The argument in one line.
A self-improving AI agent that records its own decisions and packages them into reusable skills can replace the repetitive front-end of an agency — outreach, meeting analysis, and follow-up — without a team.
Read if. Skip if.
READ IF YOU ARE…
- You run a service business or agency with 1–10 clients and want to stop manually writing cold emails and follow-ups.
- You're already using AI tools (OpenClaw, Claude, ChatGPT) but want something that persists memory and learns your specific workflows.
- You have Fathom or Fireflies call recordings and want them automatically summarized, tagged for hot leads, and actioned.
- You're comfortable in a terminal and want to build a custom dashboard rather than pay for another SaaS seat.
SKIP IF…
- You have zero clients or no existing sales calls — the meeting intelligence engine needs real call data to train on.
- You want a no-code solution — this build uses Claude Code, pnpm, Supabase edge functions, and terminal commands throughout.
- You're looking for a polished product to buy rather than a system to build and maintain yourself.
The full version, fast.
Hermes is an open-source AI agent built by NousResearch that writes its own SOPs as 'skills' after completing complex tasks, compounding its usefulness over time instead of resetting on every session. This video is a full walkthrough of building a three-component agency stack on top of it: a meeting intelligence engine that ingests Fathom calls and flags hot leads, a self-improving SDR that learns which outreach messages close, and a React/Supabase mission control dashboard. The build uses Claude Code as the programmer agent while Hermes operates the system post-handoff, creating a clean orchestrator-builder-executor separation that reduces both cost and cognitive overhead.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:33
01 · Cold Open — Stripe + Promise
Revenue screenshot hook, introduces Hermes as the agent behind the results, promises a full from-scratch build.
01:33 – 04:21
02 · What Is Hermes
Explains Hermes vs OpenClaw, the trajectory-to-skill loop, and NousResearch's self-improvement thesis.
04:21 – 06:44
03 · Cost and Model Options
Free model options (OpenRouter, local), Anthropic API costs, and why $0–20/month is realistic for starting out.
06:44 – 08:26
04 · Three Levels of Hermes
Level 0 (terminal/Telegram), Level 1 (Hermes Workspace dashboard), Level 2 (custom mission control) — most tutorials stop at Level 0.
08:26 – 15:28
05 · Install — Level Zero
One-command install, API key configuration, Telegram bot setup, and testing the base Hermes install.
15:28 – 41:08
06 · Hermes Workspace — Level One
Full walkthrough of the open-source dashboard: chat, files, terminal, cron jobs, memory files, skills browser. Includes troubleshooting the pnpm/gateway setup.
41:08 – 43:39
07 · Knowledge Fortress — Five Layers of Context
Documentation, security, technical, strategy, and historical research layers — how to build deep project context for any AI agent.
43:39 – 44:37
08 · Agent Taxonomy — Orchestrator, Builder, Executor
How to separate agent roles by capability tier to maximize efficiency and control cost at scale.
44:37 – 51:44
09 · OpenClaw vs Hermes — The Honest Comparison
No clickbait: they solve different problems. OpenClaw as Swiss Army knife, Hermes as memory/skill specialist. Migration guide overview.
51:44 – 57:00
10 · The Build Begins — Mission Control Architecture
System design for the custom dashboard: what components are needed, how Hermes and Claude Code will divide the work.
57:00 – 1:19:10
11 · Design and Theming — Screenshot-Driven UI
Using a screenshot of the existing Agents in a Box app to drive Claude Code's design decisions. Building the React/Supabase scaffold.
1:19:10 – 1:32:52
12 · Webhook Automation — Automatic Meeting Sync
Fathom webhook integration that pushes new call recordings into Supabase automatically, triggering the meeting intelligence pipeline.
1:32:52 – 1:45:12
13 · AI-First Architecture — Every Feature Must Be API-Accessible
Design principle: all dashboard features must be callable by agents, not just humans. Building the integration guide and endpoint registry.
1:45:12 – 1:50:39
14 · Hermes Takes Over — Builder Hands Off to Operator
The moment Claude Code finishes scaffolding and Hermes begins operating the system via natural language. The handoff pattern explained.
1:50:39 – 2:00:37
15 · Meeting Intelligence — Sales Call Review and Patterns
Auto-analysis of Fathom calls: key takeaways, next steps, hot lead scoring, and linking meeting data to outreach campaigns.
2:00:37 – 2:26:03
16 · Cron Jobs — Meeting Analyzer, Hot Leads, Daily Briefing
Building scheduled tasks entirely via natural language: nightly meeting processing, hot lead alerts, and a morning Telegram digest.
2:26:03 – 2:40:52
17 · Self-Improving SDR — Outreach Engine That Learns
The agent analyzes reply rates, updates its own messaging patterns, and runs the next campaign with refined copy — no manual A/B testing.
2:40:52 – 3:00:52
18 · Email System — Resend Integration
Wiring Resend for transactional and sequence emails: domain verification, API integration, and tracking opens and clicks in the dashboard.
3:00:52 – 3:14:27
19 · Nurture Sequences — Make Them Show Up to the Call
Automated pre-call sequence with no-show detection: webhook from Cal.com triggers a multi-touch Resend sequence to cut no-show rates.
3:14:27 – 3:45:00
20 · Campaign Builder — CSV Upload and AI Personalization
Apify scrapes leads, CSV uploads into the dashboard, Hermes personalizes each message using meeting intelligence context.
3:45:00 – 4:05:33
21 · Wrap-Up — System Review and Next Steps
Full system recap, what was built vs. planned, Agents in a Box community pitch, and preview of the Cloud Code Masterclass.
Lines worth screenshotting.
- Every agent resets when the session ends — except Hermes, which writes a 'trajectory' after each complex task and converts it into a reusable skill file.
- Hermes learns from 1,713 sales calls to build a personal selling model that mirrors how you handle objections — not a generic script.
- $54,613 in 30 days from six clients without writing a single cold email is the claimed outcome of running this stack for one quarter.
- The three-level Hermes stack: terminal only (level 0), open-source web dashboard (level 1), custom mission control (level 2) — most tutorials stop at level 0.
- Skill files Hermes writes for itself include pitfalls, lessons learned, and verification steps — more like an internal wiki than a script.
- The Orchestrator-Builder-Executor taxonomy separates OpenClaw (orchestrator), Claude Code / Codex (builder), and Hermes (executor) — each handles a different cost and complexity tier.
- Claude Code builds the app while Hermes runs it: the handoff moment is when the builder agent finishes scaffolding and the operator agent takes over with natural language commands.
- Hot lead detection in the meeting intelligence engine flags discovery calls where the prospect matched the solution pattern — no manual tagging required.
- Cron jobs on Hermes need no code: one sentence like 'every morning at 9am, pull my Fathom meetings and send a Telegram digest' generates the full scheduled task.
- Memory in Hermes lives in two files: memory.md (agent's notes on tools and workarounds) and user.md (your profile and preferences) — both loaded into every session automatically.
- The Knowledge Fortress is five layers stacked into a Claude Project: documentation, security, technical, strategy, and historical research — context depth determines output quality.
- Hermes' 65,000 GitHub stars grew by 35,000 during the four hours of filming — it became the fastest-growing AI agent on GitHub in that window.
- OpenClaw and Hermes solve different problems: OpenClaw is a Swiss Army knife with 50+ integrations; Hermes is a specialist with self-improving memory and skill creation.
- A nurture sequence built with Resend and Hermes can detect no-shows and automatically re-book them — triggered by webhook, no human in the loop.
- AI-first architecture means every feature must be API-accessible from the start, so agents can take actions without a human opening a browser.
How to build an agency brain that never resets
WHAT TO LEARN
The gap between an AI assistant and an AI operating system is persistent memory and self-written SOPs — and closing that gap is a system design decision, not a tool purchase.
01Cold Open — Stripe + Promise
- Revenue numbers in the hook are earned credibility, not decoration — the specific Stripe screenshot ($54,613, 12 payouts, 6 clients) sets a falsifiable standard the rest of the video has to meet.
02What Is Hermes
- Hermes creates 'skill files' from its own completed tasks — standardized SOPs it wrote for itself that load automatically next time, compounding rather than repeating.
- The trajectory-to-skill loop is the mechanism: after any task requiring five or more tool calls, Hermes logs every decision, evaluates reusability, and writes the skill file with pitfalls and verification steps.
04Three Levels of Hermes
- Three-level deployment matters: most tutorials show Level 0 (terminal only), but Level 1 (dashboard) and Level 2 (custom app) are where the meeting intelligence and outreach automation actually live.
07Knowledge Fortress — Five Layers of Context
- The Knowledge Fortress pattern (five context layers stored in a project) is what separates an agent that gives generic answers from one that knows your specific stack, clients, and preferences.
- Level 1 context (name, role, communication style, current projects) is the minimum useful threshold — Level 2 and 3 are enterprise-depth and take 10+ hours to build properly.
08Agent Taxonomy — Orchestrator, Builder, Executor
- Separating orchestrator, builder, and executor roles prevents cost blowouts — use the cheapest capable agent at each tier, not your best model for every operation.
11Design and Theming — Screenshot-Driven UI
- Giving Claude Code a screenshot of an existing app you like is faster and more reliable than writing a design spec — the model reverse-engineers the visual language and applies it.
13AI-First Architecture — Every Feature Must Be API-Accessible
- An AI-first architecture principle — every feature must be API-callable, not just human-clickable — is a design decision that must be made at the start, not retrofitted after launch.
14Hermes Takes Over — Builder Hands Off to Operator
- The Claude Code / Hermes handoff is the key workflow: Claude Code builds and scaffolds, then Hermes operates the finished system — don't try to use one tool for both phases.
15Meeting Intelligence — Sales Call Review and Patterns
- Hot lead detection in meeting intelligence works by pattern-matching discovery call transcripts against historical closed deals — the agent builds the pattern from your own call history, not generic templates.
16Cron Jobs — Meeting Analyzer, Hot Leads, Daily Briefing
- Cron jobs built with a single plain-English sentence (no code) are Hermes' most underutilized feature — meeting analysis, lead alerts, and morning briefings all run this way.
17Self-Improving SDR — Outreach Engine That Learns
- A self-improving SDR learns which message patterns correlated with booked calls, updates its own templates, and applies the improved version to the next batch — feedback loop without a human in the loop.
19Nurture Sequences — Make Them Show Up to the Call
- Nurture sequences triggered by Cal.com webhooks cut no-show rates by automatically re-engaging booked leads before the call — the same webhook that books the call starts the sequence.
Terms worth knowing.
- Trajectory
- Hermes' internal log of every API call, decision, and tool use in a completed task. After finishing, Hermes reviews the trajectory and decides whether to convert it into a reusable skill.
- Skill (Hermes)
- A self-written SOP that Hermes creates from a trajectory. Stored as a markdown file, it includes step-by-step instructions, pitfalls, and verification steps — loaded automatically when similar tasks arise.
- Knowledge Fortress
- A five-layer context architecture (documentation, security, technical, strategy, historical) stored in a Claude Project to give AI agents deep, project-specific context on every session.
- Orchestrator Agent
- The manager layer of an agent stack (e.g., OpenClaw) that coordinates tasks, delegates to specialist agents, and handles edge cases — does not write code or execute tasks directly.
- Builder Agent
- The developer layer (e.g., Claude Code, Codex) that writes code, constructs systems, and wires integrations — hands off to an executor once scaffolding is complete.
- Executor Agent
- A specialist worker (e.g., Hermes, OpenClaw sub-agents) that completes specific tasks — runs skills, processes meeting data, sends emails — under direction from the orchestrator.
- Self-Improving SDR
- An outreach agent that analyzes which messages generated replies and meetings, updates its own prompting patterns, and applies the improved approach to the next batch — no manual A/B testing.
- Meeting Intelligence Engine
- The component that ingests call recordings from Fathom or Fireflies, auto-generates summaries and action items, detects hot leads, and stores structured data in Supabase for Hermes to act on.
- Fathom
- A meeting recorder and AI notetaker that captures calls, generates transcripts and summaries, and exposes them via webhook for downstream automation.
- Mission Control Dashboard
- The custom React/Supabase web app built in this video — a single-screen view of meetings, outreach status, campaign performance, and agent activity for an agency operator.
- NousResearch
- The AI research lab that trains the Hermes model family and built the Hermes agent — their core thesis is that an AI agent should get measurably better each time it's used.
- Apify
- A web scraping and automation platform used here to build the lead sourcing layer of the outreach campaign — scrapes prospects into a CSV for the SDR to personalize.
Things they pointed at.
Lines you could clip.
00:14
“I didn't write a single cold email. I didn't review a single discovery call. I didn't even draft the proposals. An AI agent did it.”
01:33
“An AI agent should get better the more you use it. Every other agent out there resets — Hermes doesn't work that way.”
44:37
“They are not the same two. They don't even compete in the same category in my opinion.”
4:03:48
“What you have now is a single brain that sees everything. Most agencies have scattered tools — CRM here, emails there, meeting notes somewhere else, founder's brain holding it all together.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00$54,613. That's how much hit my Stripe account in the last 30 days. 12 payouts, more than six clients. And that's not even our only payment processor.
00:14And here's the part that's going to sound insane. I didn't write a single cold email. I didn't review a single discovery call.
00:23I didn't even draft the proposals. An AI agent did it. Not chat GPT, not Claude, not OpenFlaw, a different agent. the one that sat in on 1713 of my sales calls, learned exactly how I sell, and now runs the entire front end of my agency for me.
00:41It's called Hermes. It's open- source, it's free, and it does one thing no other AI tool on this planet does. It gets smarter every time you use it.
00:52In this video, I'm building the whole system from scratch. the mating engine, the outreach engine and the mission control dashboard, everything. So by the end, you can run the same system in your own agency even if you have one client and no idea how to scale to 10. Let's go.
01:12And quick intro if you haven't met. I'm Mani Kanasani and I run an AI agency out of Vancouver, BC.
01:18And over the last 3 years, we've deployed production AI systems for everyone from solo e-commerce founders to Fortune 100 enterprises. Meaning when I show you something on this channel, it's not theory. It's what actually working in real client environments right now.
01:37So section one, what is Hermes and why should you care? So what is Hermes agent?
01:42I want to be clear up front. This is not your open law. People confuse them constantly.
01:49They are completely different tools. Open Claw is like your Swiss Army knife. It connects to 20 different platforms.
01:57It does everything. It's the general purpose AI assistant. Great tool.
02:02I done a whole 10-hour masterass on it. But Hermes is different. Hermes is built by new research.
02:10These are the people who actually train AI models for a living. So they built Hermes model family and they built this agent with one core belief and I love it. An AI agent should get better the more you use it.
02:26So every other agent out there, chatbot lord, a chat GPT or even open law, they reset every time you ask it something today, it answers and you ask the same thing next week, it starts from zero or repeats the same behavior. Like you have that one school friend where who never remembers anything you tell them, right?
02:52Hermes doesn't work that way. After you give it a complete task, it records everything it did, every API call, every decision, every tool it used and in what order. And they call this a trajectory.
03:07And then it looks at the trajectory and asks itself, can I package this into something reusable? And if the answer is yes, it creates a skill. A skill is a basically a standard operating procedure that this Hermes agent wrote for itself.
03:26So next time you ask something similar, it doesn't start over. It runs the skill this time faster, cleaner, and if something has changed, it refineses the skill. It gets better.
03:40This is amazing. I'll show you the actual skill files on my machine. You can read them.
03:46They are just markdown files in a folder. Now, this has 27,000 GitHub stars and over 200 contributors, right? It's MID licensed, completely free, and it runs on basically any model.
04:00You can use your chat GBD subscription or use cloud subscription, which I was until they banned it 3 days ago. We'll discuss about that too.
04:10Or you can use free models on open router. Helm's own models or local models using Olama. You can pick anything and you can use any model.
04:22Section two cost. Let me talk about cost because there's a lot of outdated information out there. I've been using Arby's AAI with my cloud subscription and it was great.
04:35OP 4.6 is the best model there is. There were no issues. But 4 days ago, Anthropic banned all third-party tools from using cloud subscription.
04:46And that includes open flaw, that includes Hermes, that includes everything that's not cloud code basically. Now, they gave me a $200 onetime credit because that's what my subscription is. So, they call it something called extra usage.
05:03These are not API credits. So, that's what I'm using today.
05:08So, Opus 4.6 on Anthropics credit is what we are using. I want to show you what Herby's can do at its absolute best. But when the credit runs out, I'm switching to Chad GPT codeex link for 20 bucks a month.
05:25You can get you do have some limits, but it's okay. It doesn't have any per token billing. You can set that up in the onboarding wizard if you want.
05:35For most of you watching that don't have any credits, maybe that's the way you can start with chat GBT plus 20 bucks. It has codics included best value. So if you want to use the free option, you have new research offers their own model.
05:52I think it's called Hen 3 on open douer just completely free $0. Not the smartest but it will work.
05:59And if you are into local, you can use Olama. You run a model on your own GPU. It's free, it's private, it's slower, but it works.
06:09And finally, this would be the most expensive option. If you are someone who is in a position to generate tens of thousands of dollars every month, you can pay for tokens too, right? I wouldn't personally wouldn't.
06:26So if you use Anthropic API, you pay per token. Best quality with Opus. And what I'm using today is my credit.
06:35So the real cost can start with zero at max 20 bucks a month is what I would recommend for something that helped me close six deals in 3 weeks. So section three, three levels of Hermes. So before we install, let me show you the three levels of using Hermes because most tutorials I feel like they stop at level zero.
07:00Level zero is what I call using your terminal and telegram. You install Hermes, you chat in the terminal or you connect your telegram, you text it from your phone. That's what most people are doing.
07:15That's where they stop it. But you are flying blind. You can do much more and there are people who are level the Hermes workspace.
07:26Most people are here. It's an open-source web app that gives you a dashboard. You can see every conversation.
07:33You can browse the memory. You can manage the skills. You can watch the agent actually think in real time, which is really amazing.
07:43And it's like going from having a command line to your actual cockpit. But level two, this is where you level up. I mean, literally, this is what we are building today.
07:57Your own custom command center, a web app. We will use cloud code to build it that connects Hermes.
08:04It pulls in wherever your meetings are. For me, it's Fathom. It tracks your outreach.
08:11It manages your clients, everything in one screen. And this is really tailored to your own agency. And nobody else has built this yet.
08:22So we are going through all three. Let's start with the install. Section four install level zero.
08:32Level zero is just one command and it barely takes 60 seconds. And this is their Helm's website. And you can see these are some of the steps how you can start chating.
08:45It's a simple web app. Nothing too crazy what we see with OpenClaw. You can copy this command.
08:53Obviously, I did this because it's Mac OS. But if you are on other platforms, you can go to install here and maybe a good way to quick start is the documentation and Windows users install this first and then run the command above.
09:12And they have some other installation steps too if you want to do it in Nexo setup and all of that. So detailed documentation but what we need since I'm using a Mac I'm using this on a fresh machine I haven't installed it on this yet.
09:34So agent installer opensource AI agent by news or new research. It's installing. Perfect.
09:45We will do a quick setup. And now we have options. I talked about free options.
09:52You can go to no portal and you can go open router. There are some free models if you want to use. Right now it says anthropic clock models API key or clock I'm not sure they got the news or they have updated whether they know obviously they would know cloud code anthropic now banned it you can still use it through extra usage that's why maybe that's why it's there and also can use your API key if you don't have API key and you don't have the extra usage I strongly suggest to use codex because it can API is super expensive.
10:35So GitHub go by lead hugging face you can configure all of those. So now we have choices one two or three I will select one and it logged me in authorize. So for some reason first time failed.
10:54So we are trying the quick setup again. So I was researching with cloud code. So there was an error going on.
11:03The 500 error is a nonanthropic server site. But because of in the last 48 hours people are complaining that they have been running into some issues. So the transition is not smooth.
11:18So you can use OpenAI Codex router or Kimmy. But I did manage to make it work after three tries.
11:27So, I will use my opus, my credit, but you might have to keep trying or by the time you watch the video, they might have fixed it already. Now, let's start with setting up our Telegram. But before we do that, I just want to mention there was an issue with Opus 4.6.
11:51Although it accepted the token and the onboarding once I started using it, it is giving me error after error. I think that's the ban in effect. I'm not sure why they said you can use extra usage.
12:06I don't even know where to use that extra usage now. But it is what it is. They are not letting us use it.
12:14So this tutorial now we moved it to GPT 5.4. So I was using my codeex subscription and to set up our telegram these are the steps. We need a bot token from bot father.
12:30So if you've been using this for a while now it is one of the most standard steps by this point. If you are here from any 10 days, you already know and have used your fair amount of botfather.
12:51So let's begin. Have a token ready. We can either open here or create a new token.
13:03Either thing works. So to create a new token, you can just do a forward slash new port. Target chat ID is also needed.
13:19To find your target chat ID, you can open one of these. There is also a user info bot. This is my Telegram user ID.
13:31After I gave it my bot token, it configured the whole thing. It took some time and said done. I configured my configure telegram in your Hermes environment and it saved a bunch of things, right?
13:47And then I said hi and it was not working. I did say hi and nothing happened and I took a break. It's not like it's working for 30 minutes.
13:57You see a big difference. So then when I said hi and it's not working it found the issue the Hermes was not running as a single clean telegram gateway instance we added your telegram settings. So it whatever it did it did that on its own and restarted the telegram gateway and said please send another message hi to test.
14:23Obviously when I see this message it is probably configured. I said hi and it immediately got back how can I help you?
14:32So most people we basically text our AI from our phone right it's running on the machine that's level zero honestly for a lot of people this is enough because I can ask it to how do I give access to my Google workspace and then it gave me two things it gave me full workspace access we use oath 2 and the option one Gmail only I don't really need to see all of this happening in background, right?
15:02The code and all of that. And the skill view Google workspace, I think instead of running all of this, it is showing me this on a high level. So, you won't really miss too much if you're honestly speaking and it doesn't make super sense technically.
15:21For those of you who are into no code, no code or a lot of people, this is enough. But I want more visibility.
15:34And let me show you level one which is a section five Hermes workspace. And this is the level one. This is a opensource web dashboard.
15:47And we can set it up. We can clone and run this Hermes workspace. So I opened a new terminal and you need to paste this command that I'm about to paste here get clone and it is cloning this repository.
16:07It is done and let us open the workspace and we open the workspace and let's do an install pnpm install. Okay, pnpm is not found. So, we ran all the commands.
16:23I will link what commands you should be running in the free guide that you can get in the second link in the description. Perfect. We have welcome to workspace.
16:36This is on our local host, right? This is where we can share products with open air compatible back end gateway APIs connect back end. But right now, let's skip it.
16:50Now is a good time to get your cloud code involved. Anytime you run into any kind of issues, I would copy in the commands where I took a screenshot of it. You can go to your plot code agent and you need to set up the back end or if you want you can have your Hermes agent do that too.
17:14How many of these can you run? I go ahead and run it.
17:18So essentially if you don't have cloud code or you can't talk to your happy sage most of these commands they can run it for you that's how I've been doing it the lazy way you can obviously when it usually when we are following this we will run into some issues and anyway I'm not going to use this too much but I will give you step by step and usually when it run into any issues it will let us know that hey this is not working and I tried this one this approach worked dependencies installed now start the gateway and so these dependencies if we look at ran three commands and what commands it ran see we have it set the path and step two clone the Hermes agent gateway to this and flowing into Hermes agent path setup Python VNV and install and these are the two things that it did but Python version issue.
18:34So all of these one of the things that would make your life so easy is depending on other AI agents, builder agent or a coder agent. A codeex can do this for you to setting up the back end or I should have tried with Hermes itself.
18:55See how that goes. I wrongfully tried to do do this with the cloud code out of habit. But I believe Hermes can do this for realistically because if I go to terminal and look at what Hermes was able to do for us.
19:17For example, if you look at this all of these commands running this command gateway starting and all of this is an intelligent agent. This is Codex. GPT codeex 5.4 was able to run a bunch of commands for us.
19:39So I could have told GPT, hey, I am running this thing. I'm setting up my Hermes workspace and I'm running into these issues. I want you to go fix it.
19:54And the exact steps what you're seeing here obviously we rely on hope is more than codeex. I'm not sure if you are into anthropic or open AAI but these are the steps that would usually fix it.
20:13So once this back end is done I will start speaking again. Okay it is actually doing the work. gateway is already running from here setup. Now go back to workspace and it'll show setup screen.
20:32Click auto restart. Auto start Hermes gateway. Come here.
20:37Click on auto start. Maybe a good thing if I refresh connecting to back end. Auto start.
20:48Okay, let's see what did should detect the already running gateway and connect. Go to local house 3000 in your browser. A workspace should connect now.
21:02Is it working? So I still see an error in the same error that we were seeing earlier but it should be a simple issue. Set path.png PN will now restart to the current one and rerun PNBMD dev.
21:19So we or if you need me to do it. Yeah, I think it did already. So let's refresh just in case to start.
21:35No, maybe it said let's reenter PNPM dev. I'll open my terminal and go to the terminal which has Okay, that was not the plan in BM.
22:05localhost 30001. Let's copy this. P.
22:17So it finally worked. There was an issue earlier and I managed to fix it and I managed to ask advis agent that there is an issue and then it looked at it ran few commands and looked through the files and there was an issue with web API and it fixed it all by itself. All I had to do was give commands and it took a bunch of tries, right?
22:46So realistically you're looking at 10 15 minute time savings here at least. If you are not technical, this will drive you nuts.
22:56All right. And I don't want mobile access right now. I just want to show you this is like the level two.
23:06Sorry, level one from just using your inner chat to using this and you can chat with the agent here right now. You can browse its memory, everything it remembers, every skill it's created. You can run terminal commands all from this window.
23:26This is what most users are using it right now. But I personally need more. I need my Fathom meetings in here.
23:34I need my outreach data. I need my client projects. This workspace and doesn't have that.
23:41It's a generalpurpose dashboard. So I started using Opus 4.6 again. I think my extra usage is kicking in.
23:49But before that, now that we installed this M workspace, let's do a quick let me give you a quick overview of this is we can look at every panel. This is the chat. As you can see, I was I added some message and it gave me step by step on the way I work with AIS.
24:09The way I like to work with AIS is there are some steps that AIS can do, some of the steps they can't do and they want us to do. So, anytime I ask it how to configure a chron job, give me step by step how many of these steps can you do and I should do.
24:27Okay, this is basically chat. This is where you actually talk to Hermes. Same thing as terminal but in browser.
24:35You type a message Hermes will respond and it will stream text as well and tool calls will also happen in real time like little cards that show you what to using what arguments is passing back and forth. And if I want to do a 9:00 a.m. email, how to set that up?
24:57Should I give you Gmail access first? And when I send this message, you can see the little card that is thinking and once it responds, it will show the response. And the models that I can use from my subscription is also available here.
25:19So next up we have files. This is like your workspace file browser. You can navigate your project, open files, edit them here.
25:29You can use the Monaco editor that they have. This is useful when Hermes creates a skill or writes a script. You can actually see it right here without switching to your file explorer.
25:44Previously if cloud code were to do something like this I have to open anti-gravity or I have to open my folder to watch it like visual studio code but now I don't have to and we have terminal terminal has full shell access right in the browser this is a real terminal you can run any command people use this to monitor their Hermes gateway and you can check logs or run scripts along alongside their chat.
26:15You're chatting with Hermes on the left and running terminals on the right. So everything is one window. To test this, we can run a quick command like Hermes version and we'll get the version.
26:30So that is terminal and we have jobs right now which is not available but this is where it gets really interesting. Aron has a built-in chrome system similar to cloud good schedule tasks or open class chron jobs.
26:47So you can schedule tasks and run automatically if you want every hour, every morning, once a week. And the jobs panel will show you every scheduled task and when it last and what it's running next and the output. So for example, you could say every morning at 9:00 a.m. you can check your famous news sites like maybe hacked news for any AI news and send me a summary on telegram.
27:18You can just say that and Hermes will create a chron job runs it on schedule and delivers the result to your phone automatically. No code, no n, no make.com. Just from one sentence, it will build the whole thing.
27:34And the better your model, better it will perform. If you having a hard time if you're using like a free model or a local model, you might have to use a better model to set this up.
27:48You might have to spend some API or if you use your GPT subscription, that's fine, too. But people are doing wild things with this. I've seen someone set up a crown that monitors their competitor's pricing page every six hours and send them a telegram alerts when something changes.
28:07And another person has one that pulls their Stripe revenue every Monday morning and sends a Slack summary to their entire team. And remember these cron jobs can have skills attached. So meeting notetaker, we are about to build a meeting intelligence agent.
28:26You could schedule it to process your fathom meetings or your fireflies meetings every night at midnight and have a summary waiting for you in the telegram when you wake up listing all the action items that you need to do that day. And how amazing is that?
28:47So now we have memory. This is under knowledge memory. This is where Hermes stores what it knows about you and your environment.
28:57So it uses two files. One is memory.mmd and this is essentially the agent's personal notes, the things it learned about the tools, your project structure, workarounds it discovered. And user.md is your profile, your preferences and your communication style and how you like things done, right?
29:20And these are small files about 2,000 characters, maybe 2,200 I think to be exact. But they are loaded into every single session.
29:29So every time you start a new conversation, Hermes already knows who you are, how you work, and it doesn't start from zero. So if you haven't watched my open master class, I used something called introducing yourself to AI. I have a strategic context prompt.
29:49So this is a level one context bond. It is for beginners who want to avoid starting each conversation from the beginning was intent in initially made for using with chat models but it works well with open flaw. It works well with Hermes any agent even cloud code.
30:08So if you look at my context, what is your name, your title, organization, key responsibilities, your work style including communication style, your approach, your preferences, current projects, challenges, goals, areas of interest, specific ways you want AI to help you, and preferred behavior. So all of this is only level one, and it's already one page long.
30:34And if you go level two, it's really deep. And level three I can't even begin we can this is like enterprise level could take good 10 hours or more to finish the whole thing 10 15 hours so I will add all of this so you can add this in your user.md file not everything but at least the level one context it knows exactly who you are and what you do everything and we have skills this is like the brain of the learning system.
31:08Every skill Hermes creates lives here. You can browse them, read them, edit them, install new ones. So, this is a marketplace that's already available right here for GitHub issues, PowerPoint, popular web designs.
31:23The these are available already pre-built. But in the community hub, we have close to 2,000 or over 2,000 skills. everything from GitHub PR workflows to even Kubernetes management.
31:37But the real power is not the pre-built ones, it's the skills Hermes creates itself. So it will create skills based on what you do and how you work. And when you give this a complex task, something that takes five or more tool calls, it means records the entire process, every API call, every decision, every step basically.
32:02Then it asks itself, can I package this into something reusable? And if the answer is yes, it writes a skill file and a markdown document with detailed instructions, pitfalls to avoid or any lessons learned during the process and any verification steps. And the next time you ask for something similar.
32:25It doesn't start from scratch. It will run the skill faster, cleaner, and it will be more consistent. And if something changed, let's say you added a new data source, it refineses the skill.
32:39And this is what I really loved about Hermesia. This is like procedural memory. And the agent is literally writing its own playbooks and I did not see that anywhere.
32:52The sessions here is where you can see all of your past sessions, every conversation you saved. You can click any of these and you can pick up where you left off. You can continue.
33:05This is a session when I was having issues with it and I was not working previously was GPT codeex which I hit limits with. So $20 plan does have its limit. You cannot use it like you do with the regular $200 subscription.
33:22So Hermes also has a full text search across your sessions. So 6 months from now you want to ask a question right?
33:31If you want to search about a certain thing let's say Google is the word I want to search anything that I spoke about or said if it's in skill if it's in chat if it's in file everything will show up here and this is game changer for a lot of people. So this is Hermes workspace. You've seen dashboard, chat, files, terminal, jobs, memory, skills, sessions, everything in one window.
34:03But here's what it doesn't have. It doesn't have my Fathom recordings, my meeting recordings, my outreach data, my client pipeline, my CRM. So this workspace is a generalpurpose command center.
34:19I need something specific to my business. So let's go to level two and we will use plot code to build and helmies will run and learn from this.
34:30Before we build out our application, I want to show you something called clobbery. This is my mission control where my agents can come in and start working. This is one of the earliest versions or actually the version that I continuously improve and give to my community.
34:50This is where I test out new things and skills and tools. It has a ton.
34:56To give you some context, this is what we are building today. A meeting intelligence agent where we have all the meetings that I had like hundreds of meetings or 1,700 meetings. Which month has how many number of meetings?
35:16What are my busiest days? What are external organizations I had meetings with?
35:22All of that in detail and each call will have key takeaways, topics, next steps, action items, linked so it knows which team member to contact, all their emails and everything. So this is what we are trying to build and not just this taking whatever we have in this what's happening in your meetings daytoday and sending a digest learning from it and if it's discovery calls that you're looking at how to write proposals how to create prototypes and all of that will make Hermes do it and you might be wondering okay you built an application of web app how does an agent control So it is just only scratching the surface, right?
36:09We have some YouTube related apps that I built here. And I also have a research hub, autonomous brain.
36:16I have an email agent that I can send emails with. This is actually in test, but we have another agent that is actually sending the emails and everything. So this is also something we would build and we will track how many opens, how many bounces, how many delivered, how many clicked, all of that engagement.
36:40We' be able to see it and analytics. We'll build all of this too in our outreach agent and also phone agent.
36:48All of this I covered actually in my 10-hour openlow video. I also show you how to build this light version of this mission control in that video as well. And we have some agent teams.
37:03We have multi- aent orchestration agents that can assign work to each other and completely autonomous autopilot. We talk about a lot of things.
37:14So that is not the key. All the three topics that we are going to discuss here is significant improvement from what I discussed in my past video. So if you are wondering how we are going to connect the agents from whatever application that you're building to this is we have something called integration guide.
37:37Integration guide has you can agent can register their identity. It can pull for tasks. It can look at what tasks are being assigned to it or by human or another agent doesn't matter.
37:52It can pull it can run crrons. So there is an onboarding prompt and this works with cloud code agent open agent Hermes agent anything. So the sheer amount of functionality that this has, if you look at this AI tasks, it has 14 types and 14 types.
38:11All of these are API requests. By the way, if I open this 14 tasks, each task will further have more actions that you can take. So these are 14 different features.
38:24Task has eight actions. You can create a task. You can update a task, delete a task, list a task, get a task.
38:32So essentially we are giving all of these amazing features to our agents that they can take actions. They can cue the work. They can create a sub agent or manage a sub agent.
38:46They will create sub agent in their own platform. But they can also create an entry here. So we can see what's happening with that sub agent activity all log questions agent where place where agent can ask you questions status showing you in real time if they are idle if they are online every single minute if you want to and the assenee subtasks insights budget report raw report identity.
39:19So there are bunch of actions, a bunch of tasks AI can do and there are report web hooks AI can generate process reports. There is goal analyzer.
39:31This is related to API. You can create office, you can create animated workspaces where you can create a bunch of little agents that they can work um when you assign tasks. It's so funny to watch, but that's a video for another day.
39:48And if you want to get access to this, we are going to build this by the way. And if you have to skip everything and get this so you can set this up in one click, it's available in agents in a box.
40:02And we have bunch of things in here. The main mission control and we have templates of every video that I make. And I have a master class where we take an idea, raw idea to prototype to production.
40:18And we also do three-hour hackathon sessions every week, every weekend on Saturday where we build live applications and give that away at the end. So if you are interested, do check that out.
40:33And before we begin, I want to just add a quick disclaimer. This is not how I usually build systems because this is not how to build your systems kind of video. This is more of a how to use Hermes agent to improve your business or agency kind of video.
40:52How do I do this? If I'm building a full-blown application, a real application, if I look at my workflow, if I want to build an actual application, this is what I would do. I would create a five layer context.
41:12Call this a . We have documentation layer, PRDS, database, schemas, UIUX. Security layer, robust security best practices.
41:21Cloud code has some, Codeex has some, whichever application that you're using has unique security documentation that you need to follow or frameworks to follow. So technical layer, prompting best practices, text tack selection, strategy layer, personal preferences and project direction, knowledge layer, research, any historical interactions that you have about this project or any research that you did online.
41:50All of this combined goes to the clo project that we are doing in your builder agent. So it goes to a folder, it can be a markdown file, PDF file, doesn't matter. So what is the actual workflow?
42:04Workflow looks like something like this from idea to context generation whichever idea. So let's say we are building this agency mission control or Hermes mission control.
42:16So it goes from idea to generating context. We have a chat with clock opus 4.6 and generates just the documentation layer. Everything else is standard.
42:26We already have it. And then once we generate this five layers of context, we generate a prototype. And how do you generate prototype?
42:34I like to use Lovable because that's where I can create amazing designs. Now from the past month or so, we've been using Google Gemini 3.1 Pro which is amazing. And then once you build that prototype with amazing looking UIUX, you will build logic, you will build the actual back end, actual front end and you wire it with superbase or AI models and create integration guide.
43:02We are building applications that both humans and agents can use. How do humans use beautiful UIX? How does agents use it?
43:11through API building all the edge functions we do that and finally we deploy and also most people don't realize this how important this is this is agentic AI this is architecture for efficiency I distinguish agent roles into three categories this is to maximize efficiency and control cost I have been creating content using V coding tools from October 2024 so I've been doing this for a really long time And I categorize agents into three areas.
43:48Orchestrator agent, builder agent and executor agent. So orchestrator agent is your open flow. This will be like the manager that will coordinate the task and will delegate to other agents and handle edge cases.
44:03All of that will be your orchestrator and your builder agent is your cloud codeex. This is your developer that will write the code and construct the systems and executor agents. These can be sub agents or agents themselves like you can use cloud code agent SDK or you can use open claw, you can use Hermes or you can use openflow sub aents as executor agents as well.
44:32These will be your specialized workers that complete specific tasks. Okay, before we start building, I need to address something because this is a question that most people have. Should I use open claw or Hermes or did Hermes kill openclaw or should I switch?
44:49I used both extensively. I have a 10hour openclaw master class on this channel. So, let me give you a honest answer, you know, not the clickbait version.
45:00So, they are not the same two. there don't even compete in the same category in my opinion. Think of it this way. I covered it in the intro as well.
45:11Open Claw is your Swiss Army knife. It has uh 50 plus platform connections. Claw Hub with over 13,000 communitybuilt skills.
45:21It has native Telegram, you know, Docker, 170,000 people in their Discord, thousands of contributors, 346,000 GitHub stars. It does everything.
45:33It's like a general purpose workhorse. It does a ton. But Hermes is your specialist.
45:41It has 65,000 stars and it's fastest growing agent on GitHub right now. From the time I started this video till now, it gained over 35,000 stars.
45:54So, it does fewer things, but the things it does native self-improvement or three layer persistent memory skill creation that you know compounds over time, it does better than anything else on the market. Now, I want to be fair, open has been catching up. they have been uh you know they added memory viki for cross session memory in recent update and it's like the highest rated skill on all of clawhub is actually a self-improvement agent skill that mimics what Hermes does natively.
46:34So the community is building as a plug-in but in Hermes it's the core architecture not an add-on right but here's my honest take on what category uh on where each one wins right so this is in when it comes to self-improvement open flaw it's an add-on it it is available in clawhub but in Hermes it is native it is the core feature and memory memory wicki it it is across sessions um Hermes agent it has three layer persistent memory when it comes to speed open claw if we call it standard this is almost Hermes is almost 40% faster on repeated task because it has skills and the stability there is you know frequent breaking updates with open flaw you know a lot of contributors lot lot of friction, lot of you know u it they're improving but breaking all the time.
47:38Hermes is more stable. Am I saying it's perfect? Not at all.
47:42It's not even closed but comparatively more stable than open claw and skills ecosystem. Clawhub has over 13,000 skills and Hermes agent has some inbuilt skills, has community skills, but the biggest value comes from selfgenerated skills and selfimproved skills. And when it comes to integrations, it has OpenClaw has more than 50 plus platforms and Hermes has six native channels and it works with any API and Telegram native built in both of these platforms.
48:19And when it comes to GitHub stars, 346,000 stars, you know how viral OpenClaw was. Every person and their dog was starring just out of the virality. Not saying everyone is using it today but 65 plus this is slow steady and it's you know one of the fastest growing right now and by the time this video is aired it might have more than 100,000 200,000 who uh community 170,000 discord members and Hermes agent is built by no research I don't think they have a discord code quality is good for open flaw better for Hermes according reports uh best for orchestration and multi-level.
49:04Hermes agent is more for execution, you know, deep work, learning, autonomy. Now, here's what comments are saying. I I like to review all the, you know, most viewed videos across the topic.
49:18For example, Axen's video has more than 95,000 views at the time of filming this. And there are other videos talking about Hermes as well.
49:27But some of the top comments that are reading is after a day of playing with Hermes I deleted open claw and some people say I switched from open claw to Hermes and it's night and day in quote quality and then the third person said I tried Hermes for 15 minutes and gave up open immediately and I had a lot more people that talked to me talk about Hermes how significant improvement they've seen.
49:53Obviously the features that matter most to them, Herm is doing it so much better. But there is also the other side right open ecosystem is bigger.
50:04Open claw has way more integrations and 50 plus platforms and I run both but I see openclaw as an orchestrator and Hermes as an executor. So the real answer it depends on what you're building and if you need a multi-platform orchestration hub you know ton of integrations 13,000 skills on flaw hub docker deployment at scale open floor is still the play uh both the native both have native telegram by the way it's not just um one uh if you need an agent that puts uh that gets better over time and remembers your business context across all the sessions and uh it creates its own standard operating procedures and produces higher quality code.
50:54Hermes is the play but honestly power move is both in my opinion. Open law as an orchestrator delegating tasks Hermes as your executor doing all the deep work and um you know autonomy that's where uh that's what I built in my system.
51:12If you go to my agents in a box community, we have something called uh claw buddy. That's the exact system being open as an orchestrator. Everything else as uh execute a cloud code agent, Hermes agent, you can add other open claw agents as executors or sub agents as executors.
51:35So um if you're interested, see you there. 3hour weekend hackathons. It's amazing.
51:44You can find one on the about page. If you're interested on how that works. Next section is migration guide.
51:51Uh now for those of you who watched my open claw master class and you're thinking okay money I'm sold. I want to switch to Hermes. How do I actually migrate?
52:03Do I lose everything? No. Here's what the most valuable thing you build in open flaw is not the agent.
52:10It's your context, your strategic prompts, your skills, your memory. All of that will transfer. And I've put together a complete migration guide.
52:20Uh the link will be in the description below. It has everything that you would need and you can the complete stepbystep migration in five commands and I also added whatever I described right now is also included in this guide a quick table on why each of where each of these two uh shines and um you can run both like I said and the first step is to find your open cloud data and this is how uh I could have showed you but this is not my main machine.
52:59Now openflow is running in my Mac Mini. Um I don't use it on my main machine. So if you want to run it, you need to this is the first step.
53:10You need to find your openflow data and you will see all the skills memory config and you need to copy one by one. You start by copying the skills and this is the expected output and you transfer the context and then um you know system prompts and then you will start with moving your API keys and you start your headmies and introduce yourself.
53:36If you want this guide it will be available in the third or fourth link in the description. It's completely free. uh you can get it at any point if it's not working or skills some of the skills are not um going through or working fine.
53:53Uh I showed you I added uh troubleshooting section what you should be doing and if it's if you want to run both simultaneously what happens and uh yeah if you run into elite challenges you can comment in the video I will see you know comment whenever I have some time uh but if you need more support uh you can join school as well.
54:19Okay, now you know the difference and you know how to migrate. Let's build something. So what are we doing here?
54:30We are building using cloud code builder agent and our orchestrator and executor will be Hermes and we build this mission control application for both humans and agents to use and humans interact with UIUX and agents will interact with API. That is the takeaway from what I've just explained. So before we start at the bare minimum we need a claude MD file to begin with.
55:01So how do you come up with this claude MD file? I'm actually doing a master class on cloud code next week as soon as this video is up and if it's over one week it should have been up on my channel already. So I really encourage you to check that out.
55:22But what we are doing in this cloud MD file is basically whatever we are building. So if you look at this markdown file mission control what are we doing command center web app that visualizes and control heration system three integrated roles a meeting intelligence outreach autopilot custom dashboard. So the front end we are using react beat and tailwind CSS. back end agent, Hermes agent run autonomously, builder cloud, code agent, APIs, fathom resend, Hermes gateway, and now messaging, telegram bot for mobile notifications, and we have I should have blurred that, but we do have all of this to go off of.
56:13And I also created a design.md file. This is important because we want our application to look really nice like dashboard meet sci-fi command center. The reason I like to create this design.md file or give this cloud code or any of my builder agents some kind of an inspiration.
56:37It is humans the agents that will see the application. agent can use API and doesn't care how the UI looks, but I care. I like having to look at really good applications and interact with it and use it.
56:56So, we have couple of options where you can go to dribble and you can search for dashboard for really good-looking dashboards and you can say do the same thing with v 0ero.dev also. And you can look for community templates. You don't have to pay for anything.
57:16If you like something, you can either take a screenshot for the inspiration that you can use it for building this application or you can just rely on design MD file that we created. So I like this one better but I will say I need this in a blue theme.
57:40I will copy this and I will go to this builder agent and I will just say read clot.md file. We are building that application and I need to give you the API key for fetching my Fathom meetings.
58:02And also I'm hoping the UI would look something like this futuristic. This has a green theme, but I want to go with the blue theme. Feel free to improve the visual aesthetic of this application.
58:19Make it futuristic and also let me know how to add the API keys securely in the chat. Also, I don't want to share my API keys in the chat, nor you should display them in the chat. Tell me what the secure way to share them is.
58:39You can give me av.local file for me to update. So, this is the prompt I'm starting with. And all of the prompts, all of the cloud MD files and everything will be shared in a free guide.
58:53Find the second link in the description that says kit. You can get it for free.
58:58Now we have a place to add these API keys securely. But if you look at the folder itself, we can't see it. It's hidden.
59:07So you need something called an ID or those of you who don't know what an ID is, it's a code editor. Visual Studio Code, anti-gravity, all of these are IDs. So I will open the folder.
59:25Once I open this Hermes mission control, once I open the folder, I should be able to access.local. So, it already populated my superbase URL and now it's asking me for my Fathom API key. So, I will just add the Fathom API key and send API key here.
59:50If you don't know what both of them are, I will cover that in a minute. But I won't be opening or showing my API keys on here. Okay.
1:00:03So, Fathom is a noteaker that joins my meetings automatically and takes notes. And it's amazing the level of detail that you can get, the questions you can ask.
1:00:15It can even you can ask questions across all of your meetings and it can go through 1,715 meetings I have so far and get answers from there. It takes some time. It is slow and it's not even perfect.
1:00:30It's not like Opus level intelligence. It may be some GBD mini or a highQ kind of thing, but it's cool to have.
1:00:39So to get your API key, you need to go to settings. You need to scroll down and it doesn't matter what note taker that you are using. Fireflies order all of these will have an API key that you can get from wherever you have your meetings.
1:01:00Also, I realized something when I was adding the API keys. It also has where to get these API keys from. I thought this is really convenient.
1:01:12You don't even have to go to your browser and do all of this manually. Okay. And now we I replaced my Fathom API key.
1:01:22I want you to make the API calls and let me know how many meetings I have. Based on that, we will make a plan to import it over.
1:01:33So, it will take a while because Phantom doesn't have a feature to check how many meetings we have right away. It has to do some page nation stuff. And there is also a bit of rate limit.
1:01:48It has to wait a minute, check again. So, it it is a 5 to 10 minute thing, but I will be back when it's done. It almost feels like cloud code is a bit lazy today.
1:02:01It would come back after all the meetings are done. Right now, it says your Fathom account has around 1,160 plus.
1:02:12Actually, I do know that my Fathom account has over 1,712 meetings, but I don't need everything. I need all the calls that happened in the past 45 days. Let's import them to our superbase and create a new table with relevant details.
1:02:33Date, meeting name, transcript, summary, action items, company name if you can figure it out, emails, any kind of details that you can type of the call. We are bringing over 45 calls and also I'm giving you the UI UX of the application that I like. You can follow the design guidelines that are available in your folder and also make this app really futuristic and my inspiration has a green theme in it but I want a blue theme.
1:03:14So I will add the screenshot that we have earlier and I will send the cloud code. I use this plan mode where it plans first and then I prove the plan but again I'm not doing this like a builder tutorial. So I'm doing these things a bit differently now.
1:03:36So usually if we do plan mode, it will come up with proper to-do list and update the to-do list as it goes and works really well. But this is going to take a bit of time.
1:03:50We already have the API key. We already gave it the inspiration. Now it's up to Claude to build this basic structure, basic foundation for us to use for our next phase.
1:04:04It took some time but it managed to create some. This is pretty good for the first try especially when we are not using any tool like Lavable or Google Gemini Pro to do UI works and if this is coming up with the first go then it's nice.
1:04:23Right now I do like the elements. I like the glassmorphic design. And I like the glow animations and it actually fetched the 30-day meetings and meetings I have are up to date and you can see April 6th and when I open the meeting nothing is happening which is okay.
1:04:45We only asked it to build the skeleton of this application and it it was a big task. So moving forward what I try to do is maybe go to plan mode and then we also have a bunch of things we can do.
1:05:01So 111 meetings for the last 45 days and we have good lots of existing tables but no further meeting. Okay this is happening because we are reusing an existing superpase database.
1:05:16If you are using a free tier and you need to log in to superbase CLI and once you give access plot code will have will take care of the rest. You don't have to manually run SQL code. So for those of you who haven't logged into sbase CLI and cloud code can ask you from time to time this is the SQL code I want you to run it in your account.
1:05:42In the prompt itself I came in the environment variable I started with the project URL and on key it knows even if you don't have superbase CLI I think that is enough to a certain extent now you can see the number of turns fetch 45 meetings and then loaded tools and ran two commands ran seven commands user tool installing all dependencies reading all the files so it did a bunch of things and of all of the things once it's done app is running looking great I can see right side of your screen sidebar with armies branding dashboard navigation okay this is good now that's your second monitor showing different chrome window so it was using plot in chrome I can see the arm mission control app running beautifully in center of your screen here's what's visible so it actually took control of the entire screen.
1:06:43I was weird because I did give access to the computer use it can now see the entire screen which is great. Here's what we accomplished. Fetched all the meetings.
1:06:58Fathom 45 days with full transcript summaries. Perfect. Now I want some visual improvements.
1:07:05Okay, tell me prompt. Now I want some visual improvements to this application right now. If you look at the text, it is really tiny and the sidebar menu is ugly and it is not readable.
1:07:21In some areas text is just too light. We need to make it brighter, whiter.
1:07:28And also I would need a dedicated page for these meetings as well. Based on what you have in Fathom meetings table, I want you to come up with a visually pleasing page design. Let's make a plan for it.
1:07:47I want you to do two things in this patch. First is improving the text size, making it visually pleasing. And the second thing creating a page for each meeting a page loader. design a visually pleasing page based on what we have data that we have in uh random meeting but also and additional sections leave additional dedicated sections for Hermes to add notes or generate proposals or generate custom action items on top of what Fathom already generated insides is a section I'm looking for in this page and anything you feel like Hermes can add to this page should be taken into account as well.
1:08:46So when you do this in plan mode whatever you say it will take that and this is why I love plan mode. Imagine you are explaining AI a simple command right you're saying hey I want to build amazing web page that's it when you go to an AI like Opus 4.6 six maybe you can get away with it because it follows the best practices it knows and it will think okay user is asking me to create a web page what do I know about it and if I have introduced myself as an AI agency owner and I sell these services it would assume okay money is asking me to build him a web page it will assume everything else it will assume the text stack it will assume whatever it needs to create this web page.
1:09:46Now models like 4.6 they will ask you for certain context. Oh before I build go out and build this is what I would need from you.
1:09:56What is this page about? What is what do you need? Do you have any assets? all of these questions which is great but for the sake of our example what I'm telling you is if you give the same thing to the plan mold before it goes to an actual AI and actual AI works on it will make this amazing plan whatever you give it it will update it right now when we open the plan it will show you so much more than what I initially gave It would talk about the text stack things.
1:10:31It would talk about shads and UI components and it would say for this page these are the sections that are important and because I don't even know what it created for tables right what are the columns that are available in superbase I don't even know that but this AI knows and if I skip the plan mode and just give this command it will come up with something that it assumes Now a lot of things will be lost in this assumption and plan mode we will show you.
1:11:08Imagine this prompt going to an AI and whatever the plan AI creates going to AI. This is significantly better than the previous version and we will see that once the plan is done. So it created a really detailed plan.
1:11:26It's amazing. So if you look at the context it says mission control app is running the text is too small too dim and then meeting detail page is too basic just three tabs needs to showcase a page for filming and also we need to understand because meetings page only had this external internal and it is not routing it is not giving me a dedicated page.
1:11:55So assumed that there is no meeting page. There might be meeting page we will see. So sidebar make nav readable dashboard bright and everything meetings metric card step three.
1:12:11So imagine this prompt going to plot and it took a good 7 or 8 minutes for to come up with this plan. So that's why I don't go to plan mode for simple edits. For simple actions, I will just ask it directly.
1:12:30Opus is smart enough to handle that. But when I need something really detailed, big or when I'm doing this for the first time, usually I would go with the plan mode. Now, I've been doing something about this application.
1:12:49After visual improvements, the spacing looked a bit off and the text was a bit slow. The padding, everything, it almost felt like it was made with a junior there. Right?
1:13:04Even though Opus 4.6 is a good model, sometimes you cannot just ask this to improve it. So, it doesn't work. it stops at one point. So what I started doing initially is using claw chat models to improve my code models or using codeex to improve plot code models.
1:13:30How does that work? It's almost like King Solomon's advice. It doesn't work for him.
1:13:37It works for everyone else. So it's that kind of thing. So I've been making some changes.
1:13:43I said and it doesn't look like there is padding in corners. Logo has empty space before the menu starts. Maybe should we add a mission control?
1:13:53So you can talk to this like a friend and it won't go out and make a plan like we did for 8 minutes 9 minutes. It will answer to you under a minute and this is what it came up with and I've been fixing it in few rounds now. So we have this checklist this round three.
1:14:12I will copy this. We are asking for some padding.
1:14:15We are asking it to move it to middle somewhere here and improve the spacing. I don't think I added spacing. I will ask that here.
1:14:24Also, I'm not sure if you realize the dashboard. Some of the cards they look like they are overlapping. There is no proper spacing and we need to work on that as well.
1:14:36So, I haven't sent in the round three prompt yet. I want you to update.
1:14:41So used to clicking this before it generates. So I will just add this. I know it is going to give you round four.
1:14:54So we have everything we need. We'll copy the whole prompt. It's a really big prompt and paste it.
1:15:03Hit send. It will take some time obviously because it will update the tool list. So whenever we give it like a proper checklist, proper prompt like this, it will come up with this to-do.
1:15:18You can see rebuilding app layout, rebuild sidebar, fix dashboard, fix main content padding, build trust and verify. So now it has access to this site. It can see the site and it is also using the computer to take control and looking at this site as well.
1:15:38So we should have an update within few minutes. This is done and this got significantly better. I don't think it still has the padding.
1:15:51I will do one last pass. You don't have to. If you're okay with it, it's fine.
1:15:57But I want my other patient to comment on. So I observed this. any prompt that you give, you can be sure that 80% of this is being worked on. That's what this CL told me.
1:16:14I don't think code agent worked on padding at all. The corners, it still looks going to visual cards, but when I look at the sections, they're still overlapping.
1:16:33It fixed the row but not the column if that makes sense. So I will do one last visual polish and call it work on our actual features. But if you were to use this in your actual workflows, you would want something that's visually pleasing.
1:16:52You want to feel like using it. Obviously we have we will have an integration guide for other agents can come in and use it. So that will be my next prompt.
1:17:04Once this is done, this is what we are going to do. We are going to ask it to create an integration guide page that will have all the functions that this app is capable of doing it in such a way that both humans and AI agents can use it. How does AI agents use it?
1:17:23They have edge functions. We need to create superbase edge functions to pull that off. So let's see if the prompt is done.
1:17:31Yes, it is done. While this is done, I want to discuss what it said for the honest take or something. When code agent partially completes a multiart prompt like this, it usually means it stopped at the first file it could find a match in and didn't traverse up to the root layout.
1:17:51Two ways to prevent this next time. tell it explicitly which file to edit. Instead of add outer padding, say open app.tsx or whatever the root layout is and wrap top level div with padding container. So run the visual test yourself before acceptting.
1:18:12Okay. Also by the way the logo block starts all of them are nice. It automatically opened it and it's checking what the issue is.
1:18:24So the code agent says they're already fixed and it is confirming visually but I don't I don't see it. I don't think you see it too. It's choosing opus 4.6.
1:18:38The prompt that we ran last time is you can see it added padding right and also everything else is fine. But when I try and open this, it is failing.
1:18:50It is not opening. So I went to my cloud code agent and asked it great it works. But when I open one of the meetings, go to a blank page.
1:19:01For example, the first meeting and open routing working. So once whatever it built so far is working, we go to the next stage.
1:19:14What is the next stage? I want to build a web hook. I don't want to manually sync meetings every day.
1:19:23I want to happen it automatically. As soon as a meeting is done, it has to show up in my database. How does that happen?
1:19:34I have to set up few things. We need to create a web hook and you need to add that web hook into our Fathom to show that okay anytime you have a meeting send this data to the web hook and this web hook is made in our superbase in our application so it is not tied to your local machine you don't have to keep it on to make it work it runs automatically the meetings will show up in the database automatically so let's see it said it fixed it and when I open it still Not we didn't do a refresh.
1:20:20Okay. Even then it's not meetings. Okay.
1:20:27Second meeting is working. What about third meeting? Some meetings are working.
1:20:42They're opening. For example, out of three meetings that are listed, second one is opening. First and second leads to a blank page.
1:20:56So, I'll go take a screenshot of this and share it. And someone, you're not seeing the weight times in the video because we've been cutting it out, but someone said accurately the one of the worst things about wing, however magical this technology is waiting while waiting. I'm not sure if this is a popular term.
1:21:29I heard it from someone. The sheer amount of time you have to wait for this to finish is crazy. And obviously you can multitask, you can work on other things, come back, but it breaks the flow.
1:21:45And I think devs have it easy now. I've been dev for I graduated in 2015, so I've been in tech for more than 10 years at this point. It's easier now more than ever to build applications, do things, but it's also gotten to a point where you're not doing anything except speak to words.
1:22:10Everyone these days are using these speech to text models. And I think if you haven't been using it, it is one of the biggest unblocks. I am not even kidding.
1:22:23For example, I will show you my stance. I was using two applications and this is the main application and it to give you some context it saved me 422,000 words and that's 4 days of typing and I've been using it I think almost 8 months now not even 1 year and it's completely changed the way I work and makes you so much faster and with all the tools right now it's one of the important things edges that you can pull unwrap double encoded JSON react addit safe parse okay let's refresh before we try yes that's what still didn't work okay command shift R let's do a hard refresh so the second meeting which was earlier is also working.
1:23:29It did not fix the issue. I want you to research deeper why that's happening. I did the hard crash and did not work.
1:23:41I want you to make a plan and fix it and be thorough. Okay, quick lesson here. It created a plan.
1:23:51If you see three changes in two files, it identified the root cause in bunch of areas. Line 383 of meeting detail renders directly inside a p tag. Fathom stores a sane as an object not a string.
1:24:10React cannot render objects as JSX children. This throws. So this always existed.
1:24:17there was a problem but I was not able to identify when I did not make a plan. It depends on how much time you have or how you want to make this work. Usually people keep trying the first approach till they brute force it.
1:24:36Now I tried three four obvious things but I think this is the one. So you can naturally end up at this solution and also not every problem has to have one solution. You can have other solutions as well.
1:24:55And maybe this is not the only issue or only problem there is. So two things here.
1:25:04What should you do? I will share what I would do anytime I ran into error. The first time I ask it to hey this is not working can you fix it and once it does that and comes up with a response I will try if it fixes great usually it doesn't fix the issue what do I do I would say I can sit here if I'm prepared to do it six more times saying hey this did not work this did not work this did not work and usually I'm not I don't want to do that.
1:25:43Will that be choice? No. Maybe not.
1:25:45I don't know. That's why if one time fails and it did not fix the issue the first time, I will go to plan mode because plan mode eight out of 10 times will fix the issue because now it has a lot of time to review, go deeper and do a bunch of things. So yeah it is the issue is fixed.
1:26:11So the hack obviously bugged. What is the next step? Like I said we can go to Hermes right away but that's not valued right?
1:26:21I wanted to show you the complete engine. So the next part in my meeting intelligence engine is automatic web hooks. So if you look at my claw buddy application that I'm giving in my school community and if I open meeting intelligence all of this is happening automatically in my automation section I can go to functions and web hook.
1:26:47I created a fathom meeting function and this is the fathom meeting web hook. So I will take a screenshot of this and say you don't need a screenshot. You can just explain this to cloud code and it would be able to figure it out and whatever tool that you're using.
1:27:08Let's say if you are using Fathom or if you are using Fireflies, you just need to go to the settings and you will find an API key and within the API key, you can add an API key, generate an API key or public app or you can edit one of the existing keys. It already has access to this API key.
1:27:32I will copy the web hook secret. We need to build web hooks and functions. What are web hooks and functions?
1:27:43Web hook is what we set up in Superbase. And functions are like the prompts a user or an agent can create based on the web hooks we have. Let's say I created a Fathom web hook.
1:28:01I can add this web hook to fathom. So that way anytime there is a new meeting this web hook receives the data and once it receives the data it has to process it. So a human or an agent can prompt or add additional context.
1:28:22We call it functions. These functions are basically explaining in natural language what needs to happen with the data that it's receiving with this web hook.
1:28:35This can be simple instruction or a full-blown multi-page prompt. I'm adding a screenshot of an application that I'm currently using where I implemented this feature. And also you have access to the Fathom API key.
1:28:51Fathom also needs a web hook secret. This need to be included when we are building this Fathom web hook.
1:28:58And once you give me the web hook, I will update that in our actual Fathom application. Let's get to it. Or you can use flaw in Chrome if you want to add that gradually or automatically.
1:29:14So that is a big prompt. I want you to come up with a detailed plan to implement this and a checklist as well. So now I want to go with the plan mode and also if you have your favorite creator tell you something else how they would use plan mode I think it's completely fair because right now we are in a stage where everyone is figuring it out.
1:29:45So if you want to start and do build things with React or Python or some established tool or language or framework what have you have a standard operating procedures now these technologies have been around for decades or at least multiple years now. So everyone knows about everything and they can give you details step by step.
1:30:12So it would be obvious if this person knows what they're talking about or bullshitting about this feature or this tool. But right now with AI, I don't know whether to call this like an opportunity or it will be harder to guess because everyone is figuring out what the best practice is. I use plan mode for something.
1:30:33Nate Hug can use for something. Nick Sarv can use for something you know you can use for something else. At the end of the day, we are sharing what works for us individually.
1:30:46Sometimes it's work in progress. Next week, Flaw changes something, launches something that will make some of this obsolete. We don't even have to go to plan mode.
1:30:59Maybe I don't know. But here we are today, April 8th, 2026. This is what's working for me.
1:31:07It might not for you or your favorite creator, but keep that in mind whenever you're watching these tutorials. This one's going to be a big plan. So, it has 10 steps actually.
1:31:20Database migration, adding web hook secret. And then we have superbase edge functions and query helper library sidebar and routes web hook list page add web hook model web hook detail page function gate. So this is a really big build and it takes some time but we will verify whether that's working or not.
1:31:45So we hit a minor roadblock here. It said I built everything but superbase CLI is needed for deployment.
1:31:53I said can you check if you have superbase CLI. It said it is installed at homerew ready to deploy. But all of this sudden it says I need to authenticate first.
1:32:07So we are going to authenticate and this is how you would do it. I thought this would be a good thing to include in our video. So I will add open a new tab.
1:32:22Enter to login automatically. I think last time we used GitHub. Let's do that again.
1:32:29Enter the verification code and go to the terminal. You're now logged in the new version of Bundra clearly updated. Okay.
1:32:39Now we'll copy this and paste it with cloud code. If it can update it, it will update it anyway. If you have a token, you can set as environment variable.
1:32:54So token is also a good way to use it. Where do you find the token? You can come here or you can just ask it to give you the access token link.
1:33:06So it as soon as you click on the link, you will be redirected to the page name. So it actually created the web hooks page and when I click on register web hook actually created a web hook as well. If you want to resend web hook so include in web hook pist web hook now right now it's by default giving me I don't think I saved the new one.
1:33:34Okay, let's see. Also, you are the UI is Fathom web hook secret doesn't apply for new web hooks or every web hook. Right now, I'm trying to create web hook, but it comes with a default web hook secret which is not needed.
1:33:53We need to update both the functionality and UI. So if you read this right now, Fathom web hook is saying missing web hook ID. So what we need to do is come here wherever you have this you need to click on add web hook and the web hook that we created previously copy it and come here paste the web hook my recordings external recordings we don't have teams so I want to give transcript also we don't have I enabled basically everything that it let me enable and then we can come back and Say I added the web hook to Fathom.
1:34:38It should work now. Test that. And then you already have also we don't really need plan mode right now because these are straightforward instructions.
1:34:49Once I send these and then it fails and then again it's showing me the web hook secret which is only applicable for Fathom not others. I would then go in and say, "Hey, even after I asked you not to repeat it, you are repeating it. It's not fixed.
1:35:11Go to plan mode." And now it takes some time, but it fixes them. So clean build. Here's what's fixed.
1:35:20We hook test web fully functional operational. And then we have a UI fixes service selector popup. And we have per book secret is conditional.
1:35:32Fathom specific config success state only shows secret if exist per book list cards fixed existing recent web hook corrected service from fathom to recent clear the incorrect web hook secret hard refresh check web hooks. So we will go to web hooks and simple refresh is fine I guess. recent web hook.
1:35:59We have the web hook functions and events. I love this. Great job.
1:36:06We can create functions. We should say two things right now. Just make it so that the data that we receive from Fathom for this webbook is automatically saved to our meetings.
1:36:22In future when we want to add additional functions we can do but for now just save that as an automatic behave. This way it gives you it almost becomes like a make.com clone right this is what we basically doing with make it you trigger a web hook you receive data you send it to other applications via API all these make functions are API calls at the end of the day and now think of it like a vending machine right people ask me money what is the difference between an N10 automation and um agent, right?
1:37:06The difference is you have an intelligent entity reviewing everything and doing things for you while compared to previous NA10 automation or whatever. So think of it like having a vending machine.
1:37:22You click a button, you receive a chip or a co whatever you press that for, you will receive it. But think of AI agent almost like a chef, right? It knows who you are, what your preferences you like, spicy, medium, mild or it knows what dish you like and it will create that dish, make that dish and works anytime you want it to.
1:37:50So that's the real difference and why it's super valuable today now more than ever. So that's already done. and fathom poke automatically upsets every incoming meeting. So the next thing is actually big bell and significant step in making this work with enemies or any other agent.
1:38:15We are AIing our application. What do I mean by that? You need to create another page called a settings page or an integration page.
1:38:28What this has is this integrations page will have all the API calls that an agent can do to use our application. So if it wants to read the tables, create web hooks or functions, everything a human can currently do. move certain things, fetch certain meetings, add certain meetings, create web hooks or any other functionality that we are adding in the future whether that's outreach or meeting or calls I don't care whatever feature we humans are using AI should be able to use it as well and AI agent should be able to use it.
1:39:16So we do that by creating edge functions that an agent can call. So this is going to be a big build and we want to AI.
1:39:27So we are making this app usable by both humans and AI and every future moving forward. So if you want to update that in your MD file, you can. But also this is an important change and we need to build that edge functions and integration guide.
1:39:50The idea is as soon as I copy the integration guide and paste it to my AI agent, it should be able to pick up everything and start using the application. It can be Hermes agent, claude code agent or open claw agent.
1:40:14So this is a big thing and we need plan mode because I don't want to halfass anything because especially this is a big thing right so once it comes the plan comes up with a plan I will review the plan and then we will see what plot creates. So the recording failed. It created a plan.
1:40:45As you can see, this was before I ended this. This is a significant architectural change. And now the big one.
1:40:54And it came up with 19 endpoints. And I can I'll see if I can find any plan related language here. No, I don't.
1:41:04Maybe if I open this plan, will I have it? Yeah. So this is it.
1:41:09Make every feature in airbase mission control usable by AI agents. Any AI agent, not just humans. So it came up with 19 endpoints that agents can call and also if you look at the plan also update the cloud MD the files that need needs to change.
1:41:30Add a section documenting that every feature must be API accessible. Future features, outreach, calls, etc.
1:41:44So this is a significant to-do and the to-do right now step 1 2 3 is done and we have three more steps. I'm building integration page writing into sidebar wiring into sidebar and app.js jsx and also updating plot.md with the i first principle. Once that done we will review these features together because right now we have an option to once this is done we will give the integration guide whatever we generate to our hermes agent and that's when this application becomes AI agent usable and from there because it has access to the database because it has access to our meeting information.
1:42:39We can build additional tools around it and functions and skills because remember her is a self-improving agent. If you have claude, if you have open claw, if you have GPD, doesn't matter.
1:42:56These will not improve themselves unless you explicitly say, "Hey, I want you to add this to lessons learned or based on what you learned in this new proposal, I want you to update our proposal prompt." So previously when we don't have this we used to have a cloud project where we have all the proposal successful proposals.
1:43:21So if it closes in a day 2 days it will be our gold standard. If it something closes sooner it will be the new one. Right?
1:43:30But whatever we are learning we have to still manually update the proposal prompt. This is the main proposal prompt. This is what we are learning.
1:43:38This is what we are doing. But Herpes takes this to a whole new level. So whenever a discovery call happens, it looks at okay what is the current proposal prompt that we have.
1:43:50And what did we learn last time when we sent this proposal? Client said I don't want to see monthly retainer. It's too expensive for me.
1:44:00Do you have any hourly support that I can use? So now it automatically flags that. And in this new prompt, new skill that it's going to update, it will create we need to do this automatically.
1:44:14So from last time to this time, whenever you run skill, if there is any kind of change, any kind of feedback, it automatically updates a skill. And I haven't seen that with any AI agent so far. Now we have sle and finished during the piping and it created two pages.
1:44:39I haven't asked this as a skills but I asked for regulations through and it has a connection sheet details. It has the API and base API and the key that it can use for high API and it has meetings by prox functions and stats.
1:45:00So what I will do now is I will copy the API key and come back to where my her says you can go to tele. So I say something along the lines of I'm planning to make you work on an integration. We will be working off of Hermes machine control application where I have access to my all my meetings.
1:45:35So I will give you the integration guide. The key is you take that integration guide and let me know your thoughts. If you need to save it first, save the integration guide and this guide will be improved upon updated frequently because we are also doing going to create we are going to create outreach related solutions as well.
1:46:07The key is right now with the meetings we will be creating skills for proposals how we run a discovery calls objections we are creating and we will create skills off of these meetings and we will improve these skills with new meetings. So that is the task and if you have any feedback about the application and what are some of the features that you would use and like to see you can suggest more but I want you to go through this guide first.
1:46:47So I will just in case take a screenshot of this because this seems important. At the very least I will include this screenshot in the chat. Everything else I've been using my natural language.
1:47:03So I pasted 287 lines of text. I pasted all block 292 lines and told me this is solid. Let me save this integration first then give you my thoughts. preparing MCP sca manage now it is going to create skill of this and I'm not really interested in that yet because I'm more interested in the self improving herself so right now it says mission control has vision skills but only okay we are ding into something interesting Okay, meetings list get we have five ps we have four sheets what did it do okay we just command let me save this conditional so I don't waste the big picture let me test you medium scheless you are save on this yeah allow for this session I want to actually allow this conditionally we will set that up later but for this session I just want to do this.
1:48:15So API is live and working. Here's what I see and my thoughts. You have total meetings of WoW.
1:48:23That is accurate. We by I'm not sure if you saw want by the time this is recorded. I actually had another meeting.
1:48:32If you remember the first meeting that we had was this one April 8. As of recording this, I had another meeting and it's added automatically added start. I not having any days between on 7.
1:48:48I did multiple actually but those happened before I set up this web hook. Right? So what happened after the web hook is more important.
1:49:00So bottom line is it is picking it up. It is showing external, internal and unique companies mine. Total hours 75 hours of recording this week so far seven.
1:49:12Right now my thoughts on integration strong rest API easy to work with. Update insights. Action is a clear feature.
1:49:21I can write structured analysis back to meetings. Web hook plus functions pipeline means we can automate on new meetings. Proposal graph showed per meeting is not what I built in this order.
1:49:35Analysis immediate pull all 78 Excel meetings analyze transcripts build discovery power feedback skill from patterns. Objection handling skill from real objections proposal template skill. Right now I have two tips.
1:49:55I love to see an API. Okay. I would want you to improve or actually rewrite your answer.
1:50:03I will give you two points. One is our proposal engine what we are currently using and the prompt and the other one is discovery call. These two documents are directly responsible for making us $50,000 in the past 30 days.
1:50:28And the funny tart which is also done by me. So I am confident but it is not as powerful as you. So I would love for you to take a go at it and improve it against what done in these calls.
1:50:45And also note all of those 72 meetings are not sales calls. He look for discovery co 30 minutes with man. These are sales calls.
1:51:04So we need to get two things. I will do that actually away from computer away from this tweet record calling because I don't want to I will show you our current disc and then actually maybe we should use a older disco ball.
1:51:39So I actually got this problem amazing this doc told me what we should do notice this is the libing document updated after close. So I will copy the entire thing. I will go to my terminal and paste it there first.
1:52:01So it knows this is a proposal document. And also for the discovery p this is what we use and it is also equally long. You have all of these regions both of them I can link it in the resources.
1:52:19It's not secret or he but this has helped us industry. So we have this discovery call framework and takes the instant prospect prospect in URL company website and code offerings and we take all these three and create a set of questions it also has that services right. So whatever we create at the end it will get the best output.
1:52:47So we are setting this up. It is taking two things right.
1:52:52It is taking the story call and also the proposal. So now it has two things. This has serious fire power.
1:53:02Let me save both documents as skills. Pull the actual side score from mission control and then give you a proper read. So now we are getting into the act.
1:53:13This is the sauce. This is what you need. You need someone that can review not just your sales clubs.
1:53:23It should identify what you are doing with your clients and it will find the upsell opportunities for you or it can find any anything that you are missing because AI does that. Let's say I do a lot of Q&A and one of the sessions that we do is called deploying right and when you go to deploy day I will show you what happened couple of weeks ago.
1:53:56So I don't want to open I just want to show you in the browser. So when I ask my father questions, I will ask how many people asked for one-on-one coach and you don't want to know how many people actually attended.
1:54:16One person might asked about possibility of one-on-one coaching and also some I mean you know they also mention better interest in one world. So the reason I'm saying this is only four people attended the session and two of them actually asked me for oneon one.
1:54:40I work on a bunch of things. It's easy for me to use that content. I don't follow up with them.
1:54:48But imagine if I have an AI agent that is actually going through all of my calls. flags. This conversation says, "Hey, you have someone that asked for why haven't you reached out? Let me reach out to you instead." Vertigo proposal engine proposal conversing.
1:55:23This is a skill. We have a skill. And now it is saving both of these proposal.
1:55:29And I think it also saved the story called framework skill which has these are the inputs that have required optionally not so many. How do you so it omitted all the 68 lines but you see we haven't run with this. Now we have an agent that is going to take care of Discord reports and you've been asking me for bunch of permissions update your allow configuration machine so you will talk to me yes.
1:56:13So we have sales codes like Jezebel Anton discovery go there and we will give transcription to so now you have a can from every And also when you your Gmail which ability should it knows which proposals are sent and what the response is and everything that happens it can even generate because this agent has the maximum context of what's going on to that point.
1:57:04He would be and this would probably be the best agent or best position to answer that before most important calls actual parents and once you do that I won't hear you like that answer what you want to see me build to take this to next level because this is not going to stop at the client level.
1:57:44We are going to look at what calls and having Q&A calls and identify opportunities. We will look at client calls like Brian or any other client meetings and find upsell opportunities. So every meeting goes through your eyes.
1:58:05That is what we need to set up. Do you want to set up a chrome to achieve that or any additional features you want in the mission control? Let me know.
1:58:23So what we did there was we allowed all of the command consigning to CPI. Now you won't have to worry about you know give access which in the end of the day is hassle and this is going to be for superb but it should be bit when you're giving any destructive commands allowing or giving it the power to access to these destructive commands you have to So it's quicker than clock code permissions fix data and serializer.
1:59:07I've now looked at actual data and different types of calls discovery calls 30 minutes pulse planning checkins hot seat onboarding updates. So now I have what I extract qualification pain points daily insights for sales calls budget objections close probability client signings up signals risk satisfaction hard seed prospects market intelligence on boardings clarity expectation so I have seen people crap about diabetes can they sure most People are realizing how important this is.
1:59:54And now all of the meetings I was told I'm the top 0.1% user of fathom because I sit in a lot of meetings. Till now all of this intelligence is not going anywhere just sitting there. But now for the first time there is an agent that can actively learn and improve itself on every single aspect and expand.
2:00:21Trying to make my content better. Trying to make my outs better. Trying to make my deals close and deal flow and how they close the conversion and everything better.
2:00:31I cannot say how powerful this is long enough. So Hermes needs your input. I can set up these crown jobs right now.
2:00:40Which do you want to first meet? Meeting analyzer runs every two hours. PS you on telegram for all genres.
2:00:50Meeting analyzer did the briefing all three led to I want it all three. So, we have While you're building that, do you want me to prompt my builder agent for any updates on her special control or do you want me to point you towards the repository and you make the changes? I prefer the first case because I can comment, I can add my own input and improve it as we go.
2:01:36So I would still prefer when you give it when you give me command so I can relay but in the interest of time if you want to make changes also I can point you to the rep pro where the folder is so you can make changes. let me know. So this is important at this point because it's OP 46.
2:02:05If I was using subscription, I would have left it to do its own thing. But now we have to worry about the extra usage. They might be running on low tokens.
2:02:19It's already doing these heavy chrome jobs reviewing all these transcripts. we are using some of that and you can see the artics building up and why did we do that I don't want to add the stress of wooding the whole thing okay so that's what is that for now I prefer option one to urinate and you make it build perfect so if you look at this is amazing right things I should without asking daily morning breing to telegram meetings today who they are and what they know about their open action items from US release meetings account help scores deployments any leads going cold weekly Friday wrap-up deals in the pipeline where they stand total hours in meetings versus global book skills that got updated this week top three upsell opportunities across all lines real time in alerts hotly deductible story call okay and what too big.
2:03:23But what I need from vision control, these are feature updates. Okay, we have what did I miss?
2:03:40Waiting for model response while you're building that. Okay. Okay.
2:03:47All three ground jobs are live. Okay. I think this is like valation.
2:03:55Copy paste there spec one readings. Yeah, I think that's the same thing and just repeating what reading give these to your region in this order of priority.
2:04:10Search transcript unblocks pattern mailisation transing into real time tags enable pipeline tracking bulk performance optimization context log CRM foundation we are implementing everything and ideally I would go to planning work but this is if it requires plan work it will do it on its own is my expectation. at this point.
2:04:40We'll see what it does. But this is a big reset and a big list. While this is going on, I want to go to my terminal.
2:04:51Okay, let's see the to-do list. Building transcript search spectum meeting tag. It is the get configuration web hook integration.
2:05:01So this is why I don't use plan mode all the time. When AI agents are giving instructions to other AI agents, usually they are detailed enough. There is no room for too many assump assumptions and things that can go wrong.
2:05:19At least that's what I found. But can you send how these three alerts look like in Telegram? Next step is I will do a resend integration where you can you have the ability to send email.
2:05:40So eventually we will change this skill to both Telegram alert and email alert. But for now I want you to send me that telegram alert how each of this would look like. So I'm not sure if this is going to be 100% accurate.
2:06:03All those things the exact you'll get these three separate messages quick in the next few words. They're running against you one swap templates. Once you see that he wore the jeans in the balls to work to spons this is what it's great if you I want your code somewhere here so my camera is not blocking Crown job response mission control.
2:06:48Now let me compile and send six meetings in total. One discovery call line kick off one premier skills updated proposal engine four patches average channel footman pricing model and added consulting rate $100 an hour six quarts with 3k for builders guidance work.
2:07:11Added a quarter payment option for existing clients. Added multi-phase agent belt pricing structuring 18.5K phase one and 150k plus one year.
2:07:23Added three new green flags to watch for discovery framework three patches. So you see after every sales call it's already improving. This is amazing.
2:07:36And channel key patterns this week. channel partnership new grower I can build it but I don't have the time fastest flows community members who watch their master class and already use the tools are closing in 24 hours so I don't want to talk too much about clients but multi-ent orchestration as a new high ticket anchor two big deals mentioned this week 150k and 35k Brian worthy is tipping point cash I don't want to.
2:08:11Okay, client falls. Let's be away from it. But now this is amazing.
2:08:19Okay, I received one key that sum the fun. This is Starbucks. I should receive the other shortly as well.
2:08:30This is amazing. Great job. So let's see what our code doing here.
2:08:38Now API has 21 D14 total 5 you want once added. So this is a great news for WP. So we base it.
2:08:52This is just the beginning once you land. Tell me more of not the landmark is short whether the insights I wish. Oh, okay.
2:09:04I completely miss it. Help me. Enjoy that.
2:09:10So, we go to meetings and maybe I will add one of these. Open one of these and scroll down. We have her insights positive deal probably 75 followup within 48 hours.
2:09:25Decision maker pricing discuss no next sent partnership proposal. And this is humping. At least I think that is amazing.
2:09:36Comment down below what do you think. And this is just a starting point like in sand. Just in this video we will be building ton of amazing features that every business every service business in my opinion should have like a self evolving proposal discovery client upsell system at their fingertips.
2:09:58This is like intelligence supercharged. Everyone I know records their meetings the way record and patch choice order. I'm sure bunch of you who are watching this video also record this meeting.
2:10:12But what are they doing at the end of the day? Sales are really useful. But flag calls, team calls, all of that is sitting there.
2:10:22But for the first time you have a selfing issue. I tell you what, we need to add some feedback mechanism at the insights level for each page so you can learn from effectively. I want to add that to the build as phase two that's coming next.
2:10:46Tell me what to do and how are you or how long how often are you pulling this MCT and I envisioning like a feedback button and how often are you programing this mission control for new readings and changes that I make and I think this comment and feedback section will be a nice addition. to for me to give feedback or additional context about anything.
2:11:21So essentially we need a place to communicate for me to add additional notes for each meeting and for you to work off of that. So give me a prompt that I can give my builder agent Bus and the API is for you. So you whatever the prompt that you need to set for the key to set to make this work.
2:11:44Feel free to use local data and terminal or anything. We got the prompt. So this is essentially make making your meetings more real now going into almost like production level stuff.
2:12:11I initially I was not planning to go this deep because this is what we do for our actual meetings but I'm in a role I I just want to create this and see where we end up because I know now we are going into Your my agent has a knowledge. It can do certain things.
2:12:37But combine that knowledge with your human and it's unstoppable. So if you look at to-do list, create need feedback table add five feedback API actions deploy API and test.
2:12:56So it's come it's okay we have daily briefing dashboard yesterday's today's action items pipeline client health alerts okay now this is amazing it needs four dig this is I've already if you remember I was teasing it haven't decided whether I should do it or not but when we are building or doing this application. I showed you how do I give access to my Google workspace.
2:13:33Then you give your ad agent access to Google workspace. It's easy. Okay.
2:13:39Now all four applied and all three. Okay. MCF system meeting analyzer DDB fails feedback 30 minutes.
2:13:49You on mission control deal probably wrong direction. This guy was super web context and action request. Okay, give that to builder.
2:14:04Actually, I gave that prompt to builder but I am hoping to give you even more visibility to what's happening in my business. So to that end, you need to access my Google book account. So how can we set that up at least amount of friction?
2:14:29Can I give you an API key so you can set up all the Google cloud required permissions and everything or should I talk to my build agent which is still not cool to do that for me or there is no way I could avoid some manual work here. What are we looking at?
2:14:56Okay. So you got everything from the transcription. Now it is a skill that's already there.
2:15:04So it knows the music wrong set up. No, it won't be stand up. I mean not set up yet.
2:15:13There's no way alone some manual steps. Google ads you to authorize in a browser but we can minimize it without 5 minutes here that here's the split about 5 minutes in one time and go to here create a project so I will do that anyway step by step so I go you know move this around and I have my Google open so that way you know what I'm doing you can see what So create a project system.
2:15:54So maybe we will here is so it's kind of why since I used any of this my list and I know this is made for the C2 go API create project API and services and click on enable API and you search for GN API. You enable that.
2:16:48And now we have K. So next thing we search Google calendar API the calendar API and head into it. I don't know how or why we ended up with test.
2:17:23I thought we better do the other one that made a bunch of things easier and cool up the drive API. enable this Google Sheets API, Google Tops API. For a second, I thought it would open Dogs.
2:18:10pinpo API. So once we enabled all of that, we need to go to credentials and create credentials. No paper is understood and create mens and client ID and application type.
2:18:58So any broken years an okay tag broke to run credentials. And this is where we need to create manuals having application type instation and click create and dig.
2:19:59I need to block the screen. We can hang this. Okay.
2:20:15Go back to the occupation. I'm sorry. I will call the JSON.
2:20:33got the J on fire. So I'm agent actually came out to he now and when one of the discovery calls we have our feedback surgery and now we heard him fly asked her m to realize drast. So this plasmorphic design is okay for now but in inside so I think it's giving me an example on how this to but this is I like it we can go My daughter was gone.
2:21:38It gave me a phone and can keep on consent. Thank goodness. Okay.
2:22:00Do we have any locals here? It's the back. Better than that.
2:22:19You now have access to my client communication so you can see what's going on there. But make sure you're not sending me emails unless I explicitly say so. And you Usually if it's open claw I would never do this because they it's really bad against starting anywhere over and I just realized has that been I was not charging for the exchange usage for some reason.
2:23:06I don't know how long this would last because in client communication end communication mode the email that we received they said I won't be able to use headm explicitly because they didn't even name open but they explicitly named hares that I cannot use it for some for some crazy reason my other visa account is also using six string usage but it's not using it is actually using my current session do feel like it's a bit faster this time around it's nowhere near what they said so it might not last it might do this we it has created what we asked the feedback and everything.
2:24:08What I want to do now is basically go to my agent and mention that it's ready. So, we work on resend.
2:24:21But now your agent has complete knowledge. Except we are going to bow. a bridge.
2:24:40So hypo we will call this s self selfimprovement kdr and the idea is we have an lic ber and a campaign creator which can create campaigns based on you can manually upload a CSV file or it and scrape based on the scrapers that I provide whether that's apply or on these scrapers. So it can either take a bio campaign where you manually upload a CSV file and it should be able to provide an example on your own CSV that it likes the format of.
2:25:35So we have the manual camping creation out and and an automatic camping creation. So for example, one of the scrapers if you build that Google maps API we can hag and we can scrape the data just by saying accounting prompts and by core I will we will use an AI rule use an API key create a search ID that searches on Google notes to get that list and once we get that list we can create a campaign.
2:26:16So we have a manual campaign, we have an automated campaign. That's the first case. And the second case and the second part of the application or the function is ability to create campaign.
2:26:31Do you want some AI customization for AI customization? And then the third was I level you create a custom date map as well like an AI opportunity for something along those lines that is the level of clarity that we wanted and the key here is we are going to use an e send because that's the best for test developers once we use sand we can create a web hook and we can track open rate, bounce rate, delivery rate, click through rate, everything.
2:27:11So the idea is when these campaigns are going on, you will learn what is happening with these campaigns and the improve campaigns every single day, couple of times a day needed. So you will be also using that recent skill to send me emails about the daily thieves about meetings up opportunities and all of that.
2:27:40So we are building complete self-improving EIS and we already have an outreach stage in our station control application but it doesn't really have any outreach and this is keep this on hold everything that I told you is important and we are going to move away as the second step. But for the first step, I want you to give me a prompt that I can give to my builder agent, cloud code agent that can create a skills page for whatever skills you have.
2:28:31You need to be populating that page these skills and show me every day what you added, what you improved. So this is a log of how far the skills came from today. Every day the journey.
2:28:49So I want to build that next. Once that is done, once we build through the whole thing, we will work on self improving. Correct?
2:29:05It broke such at nine lines. That is a big chat indeed. I was going to take a screenshot of it and include it in a guide, but it is what it is.
2:29:17You would have to watch it. Got it. AISR is saved for step two.
2:29:22First, let me understand what I'm working with skill side, then give you a builder prompt. Now this is important for few reasons. You can see the gel 81 skill spoke three custom boot today.
2:29:37Now let me save the HDR version for later then give you builder prompt. So that actually saved the version that it has for vision version 0.1 and now it also gave me the print. So now here's the prompt for your go agent.
2:30:01I will copy all of this big man. Let me set up chron skills to machine control. Okay, they want so I just want to be able to see everything and this prompt is detailed enough so I don't want to go through it.
2:30:22But if you want you can there's creating skills table seven API skills API actions share miss API deploy API and skills endpoint build skills page UI update sidebar njs update integrations page and we have build and minified the whole thing perfect we have seven scope end points to API Once that is building, we can go to our API to see what the flow will be.
2:30:59Right now, meeting analyzer is running every 2 hours. Daily briefing 8 a.m. Big weekly skills update Friday 5:00 p.m.
2:31:10Feedback responder every 30 minutes. Skills sync 9:00 a.m. daily. Scan 81 skills sync to mission control.
2:31:19What lady Manus skills page 81 sales hang four custom mission controls proposal engine discovery framework SDR vision 77 build so we 32 weekly patches proposal engineer so whatever it change whatever it land I have a highle overview of what's going on now that my impent is working on this it will take a while but I will get back now Your job is to prepare a detailed path for the self self-intering SCR that I told you.
2:31:58Keep in mind that you will be irresponsible. This is the drawing is they are feel free to many features behind build. But what I'm saying is this is not for me but for you to work on.
2:32:25And I can only imagine what would happen if you was to take over and to do something like this. It scares me and makes me excited at the same time.
2:32:42And that's one of the reasons I've been decided to post a lot. Everyone deserves or has to know what is possible and this might give you a different idea to work on your own thing. You don't have to build the seed things that I am built.
2:33:04This proposal engine helped Brian to come up with his own deck and he calls it a mal and anytime he has a client it can prepare landing pages it can prepare and the delivery work now being done automatically and based on client feedback he's improving the system and now the system learns is this a client specific or a process specific change it will ask for it.
2:33:36This is a client specific CJ won't update my actual skill file but this project specific files I will update but if it's something that would be applicable for other clients it will improve it for M3 right so the next time it does it it doesn't make the same mistake so the only bottleneck for this is your imagination and your creativity.
2:34:07Once you know what you build, what to build to make your life easier, you can build with these agents. While that finishes writing the prompt for us, it actually went ahead and created everything that we asked for in our code. So I copied this to give it to my agent.
2:34:33I know we have played the defont with there should be a better way for these agents to communicate. But now when I go to AI skills I can see reading summarizer and the story power and patch created this is nice and bent skills.
2:34:53I don't see many but the idea is claw gold builds the structure builds the B6 and Hermes will come in and do the rest. So now that it finished I will go to my terminal and it is still okay I think let me save this as a reference. I will paste this for the time being.
2:35:20And if I go to your below on the full spectr given You hit the blue area when there's still space on how clity the war is live.
2:36:03This is that I will find my hook rand bit from here. Search for rare desens. So this is what we had doing that way detail.
2:36:48Okay. No read the whole thing. It just takes forever.
2:36:58Let's see if we changed W. Okay, we have audio brand. We have solo skills change to relieve building skills.
2:37:11This is custom skills. These are our skills. Homec himal high sky radio acceler is nice.
2:37:21So if these skills ever change, I can see change work what's actually going on and I will probably come up with a custom button. We already have it but make it custom so it won't be annoying to see because I don't think I would ever use sound scene or acceler.
2:37:55It's all these things doesn't make any sense to be feeling. Here's the space on top. You'll see the skills state call ID detect any files in this way from the hash block them automatically period.
2:38:29Also, we need to save your memory. One of the sub table stles because when the memory is contacted, there is a risk of losing context. So, I like you to set up a chronicity to yours.
2:38:54Do you update your daily memory and bit? log memory which you already have but I can I should be able to see so create a memory page for us where I can see your daily memory logs that you're updating and persistent memory boo and anytime we call that we need to see do memory lock what we learned what's working and everything before we do that how is that sound Maybe my idea is still raw.
2:39:31If you can refine it somehow, do so. I have superbase CLI. It's authenticated.
2:39:38Maybe you can use it to create the table yourself. Or I don't mind what we are doing right now since the front end is a lot. I think better approach would be giving me a problem.
2:39:55I don't get I build a patient. So this is like a bonus that you can have since already self-improvement. This is like a bonus that we are looking at.
2:40:13We now came up with a set time and before call back then losing our if anything happens that needs to work. No idea is don't lose context when back here is the redefined equation the call back session currently 730 small stable high signal daily session snapshots before any complex name of context this way city that if something gets compressed away, it's regardable here.
2:40:55He inside persons and that always matter. Dest session snapshots something through the cracks.
2:41:07Let me check what you have and me and give you builder pro. This is nice. While that is working, we can prepare brought to her.
2:41:22So we have bunch of freezes reasons. To make this work, we need to get resend. He resend is a platform where you can email for developers.
2:41:40See here is so cool. But you have bunch of free stuff, bunch of you get 100 3,000 of a month for free and also you have the ability. He's so cool to pay and he so I can know that about the head.
2:42:08I will probably do this up from my city. What I'm doing is I'm creating a recent API key and I will copy this key. This it is secured.
2:42:22They're not exposing it. I copied it. I went to my env.lo local file and updated it because if I don't it will probably ask me hey you know you haven't updated your API key.
2:42:39So for those of you who updated the API key, it won't work right away will work, but the idea is it will you'll be able to send test emails and marks. To really make it work, you go to doains and you click on add domain.
2:43:00And when you click on add domain, it will ask you to okay, enter your domain. Don't add your W domain because you already have your deal emails linked to it, right? Or I would suggest you just use a completely different domain that you that is not linked to your main domain.
2:43:23And even then I like to add a subdomain. So my main email my main domain name is growthcreators.ai but I added updates.growthcreators.ai systems.iosystems.io. IO when I do this they will give me a set of DNS records that I need to update some MX records some text records CNAME records whatever they give I just need to update in my code ID mecher c panel wherever you have traffic development wherever you have your details just update them fix everything works once you update it it will take about 20 minutes less than 20 minutes and you are done.
2:44:09Yeah. And once your domain is verified, once you give the API key to your platform env.lo, you got the run. And the other thing that you need to do is you copy the recent web and you come here and go to web hooks and I want to add the web hook and I will add that hook here and all events.
2:44:41Once you give all events add. So we added arguments. I added a reason API key and env.local.
2:45:05And this is the sign in secret for web hook. I added a recent web hook that we made for the app in recent as well. And you can use group creators updates.growcreators.pi domain to set up our email sender.
2:45:35And the sender name should be Hermes vision. So that was quicker than I expected recently. So I will come and go to my terminal and I still do.
2:46:08Okay, it's now prompt is ready. So paste it text and also once there is done we have to paste another text. But look at what our builder did.
2:46:27Our Hermes built um another thing for us broke. So we go added this time installed window. No.
2:46:43Okay. Maybe this time I'll open the folder just way again. What is that?
2:46:58I forgot what it's called. Yeah, memory page builder. page will be the spell generation just place it here and the only additional thing that I'm going to do here is probably wait till what it does and whatever summary that it gives V I will just copy that somebody and give it to the agent back to my MC agent and once we are done with that we continue building our SDI.
2:47:47So I did two things first thing is the memory page builder the prompt that it's gave me I dropped it and said once this is done let me know what you built and then we are working on this one. So it finished everything. It said all done.
2:48:08This is what it worked. And then the second thing both features are complete. It also said you know recent web hook secret.
2:48:25So it just created answer. So it just created actually recreated what it already said and then the second feature agent memory dashboard. So let's go to our app and check.
2:48:45Let's do a quick refresh. And now we have memory. And if we go to memory, we have persistent memory.
2:48:54We have daily logs. Built complete machine control with outreach memory feedback systems. Massive integration day.
2:49:02Build complete. Mission control. Decision made.
2:49:05Things learned. Context. Active projects.
2:49:09Blogged items. Skills updated. You see how detailed this is.
2:49:14Daily logs and snapshots. We have day one complete. And every day.
2:49:21So this is being updated by Hermes agent. So we'll have more things. We will build up more things as we use.
2:49:29And it says syncs every 3 hours. So I don't think it hasn't been 3 hours yet. Maybe 2 hours or something.
2:49:39Now I copied whatever cloud code agent gave me and I just copied this last part with features and I went into the terminal and paste it. So it also created whatever we gave it earlier. We pasted 78 lines right the previous input for everything is light.
2:50:04Let me update mission control skills with final endpoint count and it did that Hermes mission complete day one complete and then these 78 lines actually everything the two outputs that it gave me that the summary the conclusion. Now we are here. We are give me a status update of everything what we built so far and current workflow.
2:50:37What do I see every single day? What can I expect? And also I want you to send me those three test briefs and everything through resend.
2:50:53Okay. I want to delete this first mine. Give me status update of everything we built so far.
2:51:02I really want to wait till we finish SDR and then talk about it because right now we don't really have memory sync run hours everything is back visible. So I want you to send me those three test briefs and everything to through resent.
2:51:27That is more than enough. My email I just gave my email. It's funny it said you explicitly asked.
2:51:44I thought we are only setting limits on Gmail uh recent it's all uh set up after it but you see how it actually made it a hard rule no matter the tool you don't have or you don't want to send emails without any authorization or without explicitly user asking. So this is a good thing.
2:52:16So something happened where I was using that speech to text when I was if you look at the email that I received onboarding at present.dev and the reason that it happened is because it is trying to send from growth creators updates.growthcreators.ai. It happened because of my speech for text. So I will first give this additional context.
2:52:46I want to type whether it was supposed to be. I'll talk to builder agent to fix it. But I want to give you the API key.
2:53:08What is the best way to share API key with you? So you can configure resend on your end as well to send these emails. turn my speech to text me.
2:53:37It was supposed to be Hermes at that's the domain that I already own. Can you make the changes and also create a dedicated email function that agents can use to send a re to use recent to send me emails. So cloud code will make the changes and by the time it does that pick how you want to share CS with the local file add it to MV myself because I can paste it here but it's not the best approach.
2:54:28So for those of you the safest is echo recent API and her env. So we'll copy this from here. I will do this away from the screen.
2:54:50Okay, let me check if domain is already verified or if another verified domain we can use and also try calling recent directly since API key isn't superb secret. Now when I go to my email all systems operational 76 endpoints deployed and then resend is working three briefings incoming and daily briefing. These are the meetings that I have the office and the total meetings partnership model client have acted upon it and then we have meeting analyzer hardly detected insights return back and we also have weekly intelligence this is I don't know what you think but I personally think this is amazing You have integrations and right now we have the outreach dashboard campaign leads ICPS insights.
2:55:59We will work on this features next. We need to work on outreach. Since the prompt itself is initiated by our main agent which is her agent, I will make it invoked.
2:56:17So I'm I received all the three briefings. Great.
2:56:24So what we did in the first run is we took whatever we generated in terms of meeting in terms of how the workflow is and everything application and made it AI friendly ability for AI agents to use it and we brought that guide to Hermes agent and said hey Hermes this is what I want you to used and it started using it. It asked for improvements.
2:56:56We asked for we did the improvements. In the second round, I wanted to show you how the other way works. Let's say I this is the goal.
2:57:08I want you to prepare a brief for me on how you would build this or how you would want this build so you can use it for the best of your abilities. So these will be more AIcentric than humancentric.
2:57:24What do I mean by that? If you see campaigns right now I don't see create campaign button. When it comes to leads I have test leads but that's it.
2:57:37ICPS I have Canadian accounting firms but that's it. Insights no insights yet. So everything that has been all the 300 or 350 some lines of prompt that outreach had was entirely AIcentric.
2:57:58So I purposefully made it so because I want to be as minimally involved in outreach but this is not the end result. We are going to improve this significantly. Just like how in meetings I can take actions, I can do a bunch of things in outreach also.
2:58:20We should be able to just like how we have web hooks that I can register a web hook manually and fathom resend any custom web hook. Just like that, we would be using our own human generated ones as well. So without further ado, let's read it.
2:58:39I'm glad to hear it. That's the complete system running end to end meetings come in get in analyzed insight insights written back a briefing to go out via telegram and email skills learned from every call memory backups every 3 hours. And you can see everything on mission control.
2:58:59When you're ready for the next step, we got SDR builder prompt waiting. First live campaign is just ICP definition and a lead list away. Love that.
2:59:11Okay, I just need to do one more thing. I want you to make sure the integration guide is up to date with all the new info, the outreach SDR components that you built. So once the SDR component the outreach component and the res and our other agents can use the platform vision control to send these emails or whatever.
2:59:43We have 77 endpoints in total but the integration guide may not be reflecting all of those. So it's really important for us to make sure it is updating and we give the updated guide to Claude so to Hermes as well. Okay.
3:00:07Now the integration guide has been updated. I will do simple refresh and go to integration copy follow API friends. Now I can go to my terminal.
3:00:24I almost feel is it sometimes better to use telegram instead but I can I definitely feel like terminal is much better or we can we could have used web interface as well. For some reason, I like terminal better. Not every terminal, but Herby's terminal.
3:00:50Before we go into our AISR complete buildout, I want to add something like a bonus here. What is this bonus? I want to add nurture sequences.
3:01:08This is like a level up for people. So, let's see. Hey, I want to add emailure sequences.
3:01:19Now that you can see my calendar, you can know as soon as someone books a discovery call. And when someone books a discovery call, I want to subscribe them to a email sequence. So this is a nurture sequence and this will be a proactive nurture sequence.
3:01:46Sometimes if they have a week long, we will send alternate days, a mandatory 24hour and 1 hour before the events. So two emails go out regardless. So, cal.com also sends a notification so you don't have to worry about it.
3:02:10So, when I was doing that and I was entering my prompt into the terminal, I realized cal.com also sends a notification. I don't want it to confuse between any other discovery call or a follow-up call that we book with the clients for a sales call and send these ner sequences, right? So, you don't have to worry about it.
3:02:38I added cal API key in the Hermes environment and created a web hook in our Hermes mission control and send it to one of the web hooks on Cal. So I want you to build this nurture sequence component. The goal of this nurture sequence is to make them show up to the call.
3:03:13I will give you additional case studies, inputs and all of that as well kind of businesses that we worked in the past and you can use and search any case studies as well. do some more research on on relevant applications. So two things you know how we did the echo head recent API key equal to the whole thing I did the same thing for the cal API key and we go to our Hermes uh we going to register web hook cal.com web hook register web hook I will copy this web hook click done it may be different if you are using another application but this is what you would do for cow.
3:04:12How long after user join show up on account video meeting? That's fine. Enable web hook booking cancel created all of that.
3:04:23Good. Create web hook. Now we have this and I went to the terminal.
3:04:33This is the web hook that I have added in cow.com and I will paste our case studies here so you have some context. I also added one more thing which is client project documentation. These are the case studies some of the case studies not all case studies that we worked with where we documented what problem was solved and what improvement they did who were the actors involved timeline for this and any additional context the impact and everything.
3:05:29So we can it is a great resource for nurture sequences. And when I go to my I already pasted this in the final five lines is the pasted block this document.
3:05:42And when I looked at fathom when I looked at web hook it says fathom that's one. I thought okay it may not be working or I thought since we made fathom web hook the first thing it's showing that but right now it goes deep cal.com web hook is pointing to fathom endpoint which expects a fathom payloads cal.com sends a completely different format we need dedicated handler here is the full builder prompt and let me set up the Asian side gr that Personalizes natural emails.
3:06:22Okay. Is this the prompt or is the prompt said? Okay.
3:06:26Nure sequence builder prompt. Build a natural sequence. Okay.
3:06:34We do have the temp folder. Maybe I can just go to I need to go to and ner sequence may not be the right place them looking whatever it is builder prompt is at temp nure sequence Yeah, we will take this and just drop it here.
3:07:17So, it showed me what how the flow works sequence system trigger someone works by account web hook. So, we do have bunch of calls on boarding calls. We have consulting calls, strategy calls.
3:07:37We only need to trigger this for discovery calls. Make a note of that. And this is the output from our builder agent.
3:07:49I want you to give me a followup prompt where not going to ask for a follow prompt. I will just paste whatever output that the builder agent gives in here. So it finally finished building this and it said cal.com web hook to update change from pathom web hook and point to this.
3:08:21We'll copy this and go to web hooks and we'll edit this if yeah they are letting me edit on this web hook and see the book updated successfully. I come here I go to I agent and paste it. While that is going on, we can review this.
3:09:03We have anything that it created for natural sequences. healthy persistent primary skill and nurture personalizer pron discovery call. No, not just 30 minutes with money.
3:09:32It says discovery call from CA. If you don't see that, if you just see 30 minutes with money, don't worry. Don't trigger nature sequence and for those only for discovery calls booked on cal.com.
3:09:52So it is doing everything that we mentioned. So I want to build the we did so far. I think this will be where the CRM piece we talking about right now.
3:10:09Now we are leveling up whatever we in is allowed to do. So I don't want to go off script too much. Let's continue working on AISDR.
3:10:23Right now um we have two flows right a manual and an automated manual is user updates uploads the CSV and you create the campaign and you populate the campaign with different sequences like pain agitate solution or using certain case studies. is using what we sc from their website. So we create a nurture sequence.
3:11:01I don't think we should call it nure sequence. We create an email sequence that we will use to target these cold audience and generate leads that way. That's the first thing.
3:11:18It can be a manual upload or automatic like I said. So we need a place to upload the CSV file and you take care of these campaigns. You take over.
3:11:38So you are working on a lot of things. I want to see a log somewhere. AI log for for your last poll, last check in.
3:11:48And 3 hours makes sense for memory. 30 minutes makes sense for some of the meeting related stuff, feedback, but nurture sequences should be sooner, don't you think? And shall we do 15 minutes or do you have another plan?
3:12:07How often do you want to check? And will there be a local demon that wakes you up or can you do that automatically by setting the chron? Let me know how that works.
3:12:24So people go to our mission control once and look at where we are at. We have close to 90 end points right now. We do have bunch of things that we can do.
3:12:39We have this memory feature. I don't think it is accurately syncing. And we have daily logs.
3:12:49Why does it show its ninth already? Oh, okay. 3 hours ago. Choose single iOS API function pattern log ground act analysis memory sync every local sign read only get Google box never send without approval.
3:13:21Nice. It is working. We will check back few more minutes later to see if we received any update.
3:13:31Okay. Asian memory it says blast sync never it's a good thing to flag good question let me answer each then we'll build aiming agreed nurture Facebook discovery calls cold outreaches campaigns or email sequences two different system different in timing for nurtures nurture sender it also runs every 15 minutes super base and that handles the actual sending on time my ner your personalizer at 30 minutes is just for customizing content before it sends.
3:14:08That's fine because welcome email sends immediately handled by Cal. No, wait. Bandicles are scheduled for 9:00 a.m. next day.
3:14:18I have hours to personalize 24 hours and they're all template. So the timing is already right. Watch it.
3:14:27So they run inside gateway process. or separate demon when gateway is up. Crown's fire on schedule.
3:14:35Each crown spawns a fresh Asian session, does its work and delivers the result and exits. AI activity log. This is the missing piece.
3:14:45You should be able to see exactly when I checked in, what I did, what's coming next. Build into the system and let me create a crown that manages campaigns end to end. Let me update all CRS to update log then activity hit no feedback create this activity log should be a page on Hermes machine control that's one and also when I looked at memory it says last sync never and we set it up more than 3 hours ago why did it not run and when it runs I need to see the AI log in on your dashboard.
3:15:29So, can you give me So, I'm hoping the prompt that you already gave me does this. Please confirm. So, we have AI activity log prompt.
3:15:43MD. I will go to the temp folder. and copy right here and hit send. So I'm not sure how many of you are hitting this.
3:16:09I won't add this intentionally. So you see activity log as dedicated page. We are building that now questions also add a dedicated activity page tab.
3:16:21Also make your memories and crown pares correctly going forward. So I think there may be bunch of things that it does activity what builder covers what's already working builder prompt fixed all prompts now locked activity lock let's refresh memory page so you see entire entries now 0 minutes ago perfect in memory user dot user rules and many kani founder and we have nurture only ancient memory of dated dogs don't know why it shows ninth already this is eth snapshots perfect love it great now our next task should be I want you to review what AISDM was already built and how we should take it forward in the sense of managing the campaigns.
3:17:39We agreed on timing but the next step should be how to customize and actually sending these emails through resend. what inputs that you would need from me for if you make the manual workflow where I can upload the CSV and you can create the campaigns and send and trigger the campaigns and monitor what we receive on web hook and everything.
3:18:07Then based off of that we will build an automated ICP builder campaign builder as a second step. And this application should be both human and AI agent friendly. Both should be able to use it.
3:18:31Right now it feels like AI it's only made for AI agents. See, so this CSV file create a dummy CSV file that users can download and include that in a builder prompt. And once I download that I will add real data and I will start with my personal data to see how it works.
3:19:19Then we will build an actual ICP builder maybe something simple like Google map scraper or something or appy or LinkedIn scraper and test it with real prospects. So I have full picture.
3:19:44It's working. See file. Yeah.
3:19:48Sometimes I feel like it's we are talking over each other and I'm restating a bunch of things while it gives me this new builder prompt. I want to look at what AI log was where the AI log was created. I don't see anything.
3:20:10So, we have Yeah, I don't think we have AI log. I will ask. So manual campaign what I need from human CSV file who are these people what is the goal any tone or angle preferences that's it the flow we have the entire flow and it is saved at one of the folders I believe we have CSV campaign flow builder prompt I want to go here upload heads end.
3:21:06We need another UI focused prompt when we a table. I think you misunderstood what the AI log is. I want to see a persistent miss last check in on screen. when was the last time you pulled to check tasks or the last check in and also a dedicated AI log page where I can see the activity log of everything you have been doing both on Hermes and by yourself so I don't have to leave I don't have to wonder also create a calendar based on what on my calendar and based on what task that you're working on.
3:21:57So if you see a chron if you have a chron a brief add a simple this is where the brief will be sent. This is where you would be checking for campaign. But if you are checking campaign for every 15 minutes you don't have to show up show that in calendar for tasks that are over 3 hours long.
3:22:23you they qualify for a place on calendar and any other insights you want to add to this calendar page and AI log page feel free to do so we'll also see whether this is a big UI build okay maybe it takes time let's wait for what the terminal comes back with Hermes will say okay this is the prompt I made whatever the prompt it comes back with I will just drag and drop with the builder and I know this is a telephone game but this is the best we can do right now it almost feels like I have two cloud go terminals working together now but Hermes is the more intelligent and more you have to build all of this on Claude to make it this powerful the skills self-improving skills.
3:23:33Yes, you can build that to an extent on plot core itself but goofy. It came back with the builder prompt is AI presence builder prompt and also I think campaign flow builder campaign redesign from earlier. These two give you complete human friendly experience.
3:24:01Did we not give campaign flow builder? I think already my AI is currently finished working on campaign flow builder prompt. This is the output and the next thing will be presence builder prompt.
3:24:24So if you see here agent task grounds sync. So this must have both the present all the three presence part AI log page unified calendar three features that make AI agent feel present and visible across mission control row. This is also an important thing.
3:24:54This is done presence builder is next. While that is working on that, I will come back here to the terminal and paste it here. So let's take a look at good the campaign p human friendly for the builder build some presence there to end campaign with the existing API to make sure everything works before you put real leads in.
3:25:36So we go to outreach. We go to campaigns. Create new campaign.
3:25:43Okay. Asked for test outreach Vancouver upload CSV example. Okay.
3:26:04I think it's it happened because it is working on the app right campaigns dental real estate ready for okay I think my agent is my admiss agent is working on this right now so I'll I don't want to test myself. Okay, I missed a bunch of things.
3:26:52What's my honest assessment? What's not working yet? CSV file with real leads.
3:27:06Format only email is required. Majority email preferences. Okay.
3:27:18Zoom everything and send a test campaign to mi at growthcreators.ai. And I'll just review to give you a green and move on over to making the automated flow where you can scrape the leads yourself. Okay, just send the test campaign to the about email or actually make it send it to so my builder agent must have finished a bunch of things.
3:28:05Whatever it's updating, I can see things refreshing the side. Why does it it has the AI log now manual meeting analyzer chron status and it has integrations. I thought there will be a calendar page as well.
3:28:33Last check in 2 minutes ago. Nice. Analyze three meetings, three leads.
3:28:53Oh, this is nice. Actually, this is a cool little upgrade. There is a calendar page.
3:29:03I don't see Oh, okay. I think once our agent takes a look at this, we'll update the current calendar. I checked the campaign test.
3:29:30Put a question about Acme Dental. Hi, John. Noticed Acme Dental was serving patients in Vancouver. had a quick question.
3:29:38How much time does a friend desk spend answering same questions every day? I know this is not perfect. It's not crazy but it is something.
3:29:48It all comes down to how much what is the quality of your lead list came across one realy group had a good question when a new lead comes in your website how long does it take? So I will update the copyrightiting best practices that's working for other clients and the basically the point of using Helm's AI is Hermes agent will figure out what's working and what's not working and improve every single day.
3:30:25That's the idea. So let's see if it is refreshing anything. Now that calendar 12 event showing all pages are populated with real data.
3:30:38Is it didn't populate with real data? Okay. Shift command R.
3:30:48Let's do a hard refresh. I don't see calendar populated. It's empty right now.
3:31:08Let's look at a log campaign manager ner. So these are all the checkins that's happening every single day. So there is one line fix that needs to be done with builder.
3:31:45So now that the campaign is working, I went to the Appify platform. I have $5 free credit. So for proof of concept, we'll build something simple.
3:32:02So I copied the API key. I went to my env.lo in the Hermes mission control folder. Now I just going to say I want to build an automated campaign creator or ICP builder feature that we have in um outage section.
3:32:29So when we are creating this automated campaign, we should be able to use Appify API FY. I already added the API key and the env.local file.
3:32:50I want you to take that search for scrapers. Do some research. Use proof of concept.
3:32:58Use this as a proof of concept. It doesn't have to be perfect from day one. I just want to show that it's working.
3:33:10So for you to make this work, you can depending on scraper what they would need in the ICP builder. So you can ask me for entering all the parameters or criteria required. You can make some by default.
3:33:29You ask which vertical they belong to like law firms, geographical location you can make. It can be dropped down like US, Canada and further cities. I don't want to go broad or global.
3:33:46But now, so once we have that in the campaign, you need to go out and scrape these leads. So I haven't upgraded my ampify. So we'd have to use something from their free plan.
3:34:02Keep that in mind. Free scrapers. We do have $5 credit.
3:34:06And once we scrape those leads, we would send create a campaign and our agent will our Hermes agent will take over from there. So build this feature and update the integration guide with anything new that haven't been done so far.
3:34:28So this is a big prompt a big ask. So whatever this creates once it finishes we will take the output to the agent. Now that we also am upgrading the integration guide, it knows exactly what we did and where to look for automated campaigns.
3:35:02But if you are using this for your own agency or if you want to use this for a client, obviously you would have to figure out a bunch of things in advance like what scraper you're going to use and what they because they have a targeted ICP, right? You don't want them to come up or figure out what their ACB is. So if you are building something like this in production, it would be different.
3:35:39So run a real Google map scrape a file. So we have copy the whole thing. This good decorations page escape scraper runs.
3:35:57Maybe we will try and keep everything that we have here to our agent and outreach campaigns. Fine leads and auto campaign accounting firms in Vancouver. Next, Canada, US.
3:36:26So, Vancouver, BC. Next, scrape. Max results 20.
3:36:33Can start scraping and that is takes 30 to 60 seconds. I like the UI too, design and everything. While this is processing, we can come here and give it.
3:36:54And I think an agent will come back faster than the AI. So the agent told me there was an error and when I went back to check the outreach that we are doing also through an error we are going to give our builder agent the same error.
3:37:42So we have a fixed and deployed show leads and auto campaign. I want it to be human because why not? If agent is creating a bunch of tas law firm Vancouver law firms kept my English if we see 20 starts scraping thing.
3:38:31So, I don't want to go to my agent for a second. I feel like maybe both of us tried to do this at the same time, both human and agent and did something. I haven't read the error what it was in detail, so I don't know really.
3:38:51But while this is going on, I'll just wait, see how this goes. And then involve, hey, I got a campaign for you. And to do the same thing with open claw and the cloud code, it took me a lot of time.
3:39:08So the scrape is complete. It is literally less than a penny. I'm not sure if this is accurate or even close.
3:39:17We will look into that. So when you go to campaigns law firm Vancouver AI processing I will go back to the agent and paste it. I also manually created a campaign for you law firms Vancouver.
3:39:43So law firm campaign is ready for your review. Click your email and you got the notification. That's nice.
3:39:51And I don't want to be receiving these emails just because I'm sitting here. I want to be in a position where I receive these emails even when I'm not right. So your campaign 20 leads three email sequences review and mission control.
3:40:09Now it's in local host. I'm not sure what it's doing sending me links.
3:40:14But once it's deployed, you can expect actual things. So if we refresh and go to the campaigns, plonform ready for review. Right now it's not opening, which is a problem.
3:40:32Right now the campaigns when I click on them, they don't open. I see ready for review processing but when clicked nothing happens. We need to fix that and because plan mode is taking really long.
3:40:54Okay, it's fine. I don't want to go rely on plan mode. If we just follow what I taught you earlier, your first try, you try and fix it by itself.
3:41:07If it doesn't work, you do this, right? And also I wanted to show you something the kind of web hook events that we receive when we send out these events. So we are sending recent emails and these recent emails are sending back.
3:41:26So anytime I open, anytime I click a link, all of that is tracked. So you can see 45 events that's written here. 45 events, email, click, open, delivered, all of this in detail. So, who opened it?
3:41:44What is the message that is sent? So, great information. Very useful for building the improvement flows.
3:41:58So, I want to make this a real thing, as real as possible. So if you have any thing related to they look good and I am working 100% on mission control.
3:42:29We need an approve button and you need to be polling like every 15 minutes for approvals. So we go to outreach. We go to campaigns.
3:42:46What did our builder say? Campaign guards and outreach are inflictable. They just navigation.
3:42:58Let me add you through. We get three changes. They are now clickable.
3:43:05If I open this at Asian notes from strength deletes from Google Maps placeholder local. What is that? So one is active.
3:43:20I go back. And also some of the emails have placeholder in the address. Maybe we can skip them.
3:43:40I just realized something. Campaign updates. Even though it's a test, we don't want to send from it.
3:43:47Special control. I wanted to change that. Right now I'm updating the sender to Manik Kalisani Mni at growth at updates.growthcreators.ai.
3:44:14Uh so that way the receive email from my name for these campaigns not had me set mission control builder agent also made these changes. So essentially that is it. Now we have a working campaign system that will as soon as we approve it picks up its sense.
3:44:51I'm just waiting because it's 700 p.m. where I live, not the business hours. Right now it says window is 9 to5 and limit 50. So I've already shown this and because I see the change from there too.
3:45:17Now you already seen how it works with whether we show sent delivered bounced open clicked or not everything we have seen the events in here in the web hook section. So all of these would be tracked and maybe tomorrow I will because this is not business hours like it's suggesting anyway I'll keep it that way.
3:45:44And that is it right now. What did we set out to do? We wanted to build a meeting assistant meeting intelligence and outreach that is improved every single day.
3:46:00We need to set a chron at every single night or day, whichever you prefer, based on what campaigns are working, what kind of messaging getting more opens and more clicks, whether the lead magnet ones are performing well. So based on these insights, we improve the skill and improve our copyrightiting skill.
3:46:32What do you think of this idea? Should we set a chron job for this as it should be dedicated from not bolted onto campaign manager? Here's my thinking.
3:46:49Send emails respect send windows tactical execution need to run often daily at 9:00 p.m. makes sense and window 9 to5 p.m. is closed. All of today's opens clicks have had time to register and the insights feed into tomorrow's sends.
3:47:14So it created something. Now I want to upgrade what we have in our skills page. So we have all skills we have changed log.
3:47:30This is good. But I want to add a skill builder. we go to our agent and say, "Okay, the next feature that I'm envisioning is improving or adding another tab into our skills page.
3:47:53This is called skill builder. What this does is it takes a URL or a YouTube transcript or an article, MCP, any kind of documentation and once we paste it, we extract a skill from it.
3:48:14What we can learn from this? Uh you will be the one who is processing. We will just enter and submit for review.
3:48:25AI will process and update the output. What do you think of this flow? What do you need to make this happen?
3:48:34Once we finish our discussion, we will come up with a prompt for both human and AI friendly front end, an API endpoints for you to work on it. Here's the system. So daytime rehab 9 to5 campaign manager recent web extracts events nighttime SDR intelligent analyzes results find patterns in data patches outreach copyrightiting skill tomorrow sends use improved copy PS solution he go out initial copy now intelligences on subject day 28% skill patches log inside campaign manager uses updated pattern knight more data more pattern skills I don't know how many of you got to the this point and actually reading this deserves in the first intro or book because this is amazing we are talking about an AI that can improve itself over and over every single day based on results.
3:49:46This is amazing. Okay. And now we have this prompt.
3:49:51We can run from this. We will simply enter and submit it. Okay.
3:50:02The next feature. Yeah, we it is cutting. Great idea.
3:50:11This turns curve creation from something that only happens during our sessions into something you can feed anytime. Drop a URL based transcript proced into structured skill. Here's how to define the flow.
3:50:26Paste URL transcript. Pick the category. Add notes.
3:50:30Focus on X. Click submit for processing. Extract all of this.
3:50:34Analyze what's teachable knowledge here. and when to use generate structured skill.md save as draft file and click appro course live in my skill library this is next level what I can handle URL the transcript draw text PDF what I need the content the URL based helpful key decision should each submission become one skill or should be smart enough to decide different topics yeah multiple skills each get their own approval and I want you to write a prompt for my builder ag sheet.
3:51:16This full prompt is available in the temp folder as always. You can scroll down skill builder prompt. We'll give this to agent and then discuss what it does briefly.
3:51:35You have base URL, click, extract and analyze, fetches content, analyzes testing skills, create proposals, review each proposal card, approve, edit, reject, skip, creates actual skill files. So, the builder is doing that. I like the new 15minute polar for this, too.
3:52:00Yeah, I don't want to wake up my armies or talk to her just for that. Let's go to AI log. We have sync two memory entries 3 minutes ago, which is nice.
3:52:17Campaign law ready for review 20 leads. Let's move it. Okay, the shape 20 leads Google maps.
3:52:29This is nice. That is detailed. The text is still a bit light.
3:52:36It can be brighter, I believe, but it is manageable. We've come a long way. So, the final skill that I was thinking in the final part once we build this skill builder is we need to work on I think the third part that we created already is linked to meetings and emails.
3:53:04I believe for the third build, we wanted to look for what's going on with our clients, figure out upsell opportunities, project management and all of that. It is already doing it in the emails that we are sending. I think that we can conclude this video after we built this skill builder.
3:53:29I think it's long enough. You do you managed to build the complete meeting intelligence engine.
3:53:36And now your agent can send alerts to the team members as well. If you give it team member emails and say after the meeting you assign tasks to these devs, you send emails to these devs, it will do that. All right, just like how we are doing this outreach.
3:53:58And if you look at campaigns ready for review and if I go here, it has 20 leads. AIS they are processed and reached 20 leads info. A lot of info emails.
3:54:11It's fine. I don't mind. Let's approve and see what happens.
3:54:16And again, it's not like it we are going to send this today. It is going out tomorrow. If I haven't edited or posted this video, but tomorrow I haven't deleted the video by then.
3:54:30I will make sure to record this first thing in the morning and show you how it goes. So, let's see the status. And if we look at the to-do, I think it's a big list.
3:54:43Creating skill builder tables. Add eight skill builder API endpoints. deploy and test endpoints.
3:54:53Build skill builder tab UI update integration page build and verify. So it is done. This is a summary.
3:55:04Copy it. I want to see it first. If we go to skills and skill builder.
3:55:27So we have AI is still processing. Right now I did not do anything. I it is my agent saying submission and everything is live.
3:55:41Let me do a quick end to end test. Submission is sent. Content extracted.
3:55:47Waiting for processing. Let me process it right now. So you can see the full slope.
3:55:54And now it said click approve. Okay. Recent deliverability.
3:55:59That's near. I don't see it. So maybe a refresh will fix that.
3:56:04Skill builder recent best practices. Okay. Email deliverability best practices and domain bomb bounce management line optimization.
3:56:18This is a test anyway. So when we when I go to integration it's not integration second longer we can see the crown so all the crrons are updated in ari log because I don't think I see the one for this skill builder that are pulling every 15 minutes.
3:57:091 2 3 4 5 6 7 8 I don't think all crowns are here. Crown run. We'll see.
3:57:24So, how do we make sure you're updating the AI log? Not forgetting when we add new grown which file needs updating to fix that. So the skill builder chron just updated and also when I go to skills and skill builder ready for review approved approved ones will be worked upon in 15 minutes.
3:58:02So we already checked the log that the crowns are running. Crown runs you can see earlier runs Friday. There are some weekly skills.
3:58:15There are some nurture center runs every 15 minutes via super base crown. No nurturing due intelligence standing by runs nightly at 9:00 p.m. No campaign data to analyze yet.
3:58:32So the only thing that's needed keep your machine on and all of this works. So, I just sent this message and it should fix it. But overall, this is an amazing application.
3:58:54From what I was reading, one of the messages that they sent me, it was saying, "You have all of these things, amazing things. What do you want to Where do you want to go next? Lord, it was a good way to end the session.
3:59:13That line is really strong. Where do you want to go? That's fixed.
3:59:20Two layers of protection. Skill file when I can create a new crime. No single config file controls it.
3:59:29So instructions. If you want another guarantee, your builder could make activity stats. It's okay.
3:59:39I want you to take care of it. Perfect. We have that.
3:59:47We resolve that. Everything that we set out to build, I think gets built already. We to double check that we can ask our builder agent.
4:00:00I want you to review the plot MD file. Did we build everything we set out to build or is the third feature still missing? I think the third feature actually evolved into something even better.
4:00:22Awesome. Actually, we built everything that we set out to do. Meeting intelligence, outreach, autopilot, custom dashboard.
4:00:34So, all three are built, but uh MD is still headto references all schema want me to be right. Yes, go ahead. So, that's not really that important at this point.
4:00:52Okay. What do you think about the system? I think we can it is a good point to stop at this point read whatever my helm does and maybe towards the end what we have for campaigns I will update it.
4:01:16So what we build today is not a tool. It's an operating system for your business. That's what I've been telling you.
4:01:25Most agencies have scattered tools, CRM gear, emails out there, meeting notes somewhere else. Founders brain holding it all together. What you have now is .
4:01:39So this is just scratching the surface. 50K your proposal engine made in 30 days was a static system. This one compounds. What's your take?
4:01:48What feels right and what feels off? Yes, we will improve their system. We can make an additional videos on Hermes if you are interested.
4:02:00But the next video we are going to go is Cloud Code Masterass. If you don't want to miss it, please subscribe and also agents in a box. If you want to skip the entire thing, get all of this and also the application that I was showing you earlier which is even more powerful, even more end points like I can right now this application has what close to 99 endpoints or something.
4:02:33This one has almost 3x that and so many other features. You can build skills. You can build tools.
4:02:43You can build MCP servers. You can build apps. Op center itself has meeting intelligence, research hub, phone agents, email agents, lead enrichment.
4:02:54We have a skill factory. We have agent teams. We have workspace animated offices.
4:03:00And we have I can go on and on. If you're interested and want to get access to this and also weekly 3-hour live sessions, I'll see you in my agents in a box school community. First link in the description and everything that was discussed, every prompt I used all of this step by step.
4:03:22I will create a masterass guide. That guide will be available in the second link in the description. Thank you and I'll see you in the next one.
4:03:34For those of you who are wondering how do I access this? We cannot do this forever in a local host. Right?
4:03:44I would ask this to push the code to GitHub and deploy to Netlify. You can just ask that command and it will do it. I want you to push this to GitHub and deploy to Netlifi.
4:03:57And this is the only command that it would need to do that except we if you were to push this to GitHub and netlify. I want you to use the superbase authentication where the people that are getting in should be authenticated properly and you need to be creating signup signin pages. You have all of that.
4:04:25I can show you quickly how to do that too. But the video has been really long so far and I've did that in my previous videos and we'll do that in the future video as well. This is something that you would be using as an internal tool anyway.
4:04:46So you don't really need that in my opinion. Now you have a place, a control center, a meeting intelligence engine, outreach engine, a self- evvolving AI SDR, everything at your fingertips.
4:05:00You can see all the skills. You can build new skills. You can see all the changes to your skills.
4:05:08Whatever changing, however you know your skills changes after the meetings. And if you want to see more on agency side, how to automate your business, how to build self-evolving skills around your business, comment down below and I will do that.
4:05:29And I'll see you in the next one.
The Hook
The bait, then the rug-pull.
Fifty-four thousand dollars in thirty days, twelve payouts, six clients — and not one cold email written by hand. That's the opening claim, and the next four hours are the receipts: a complete rebuild of the AI agent system that runs the front end of his agency, from lead sourcing to meeting intelligence to self-improving outreach, all orchestrated by an open-source agent that gets smarter every time it completes a task.
Named ideas worth stealing.
43:39model
Orchestrator-Builder-Executor Taxonomy
- Orchestrator (OpenClaw) — coordinates, delegates, handles edge cases
- Builder (Claude Code / Codex) — writes code, constructs systems
- Executor (Hermes, sub-agents) — runs skills, processes data, takes actions
Three-tier agent role separation that controls cost and prevents capability mismatch — use the right agent type for each layer.
Steal forany multi-agent pipeline or AI-assisted workflow
41:08model
Knowledge Fortress — Five Layers of Context
- Documentation layer (PRDs, schemas, UI/UX)
- Security layer (best practices per platform)
- Technical layer (prompting, stack selection)
- Strategy layer (personal preferences, project direction)
- Knowledge layer (research, historical interactions)
Five-layer context architecture stored in a Claude Project that gives any AI agent deep, project-specific knowledge on every session.
Steal forany Claude Project or long-running agent deployment
06:44list
Three Levels of Hermes
- Level 0 — terminal and Telegram (most tutorials stop here)
- Level 1 — Hermes Workspace open-source dashboard
- Level 2 — custom mission control web app
Progressive capability tiers for Hermes deployment — each level adds visibility and control over what the agent is doing.
Steal forexplaining AI tool adoption levels to clients or communities
01:33concept
Trajectory-to-Skill Loop
After each complex task, Hermes logs every decision and API call into a trajectory, then evaluates whether it can be packaged into a reusable skill file — the core self-improvement mechanism.
Steal forbuilding any system where AI agents should improve over repeated use
1:32:52concept
AI-First Architecture Principle
Every feature in a human-facing dashboard must also be callable via API so AI agents can take the same actions without opening a browser — baked in from design, not retrofitted.
Steal forany app that will be operated by both humans and agents
How they asked for the click.
VERBAL ASK
4:03:48product
“If you want to skip the entire thing, get all of this and also the application that I was showing you — Agents in a Box. First link in the description.”
Delivered at end after showing a more powerful version of what was built; creates aspiration gap. Complemented by a free Gumroad playbook (second link) for those not ready to pay.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
