Alex Finn walks through every surface of the new Hermes Desktop app and shares the session management insight that turns a $1,000/month bill into almost nothing.
Posted
3 months ago
Duration
Format
Interview
educational
Views
5.4K
324 likes
57 · 43
Big Idea
The argument in one line.
Managing your AI agent's context through sessions, profiles, and targeted skills is the difference between a $1,000/month bill and one that costs almost nothing, and Hermes Desktop finally makes those controls accessible without touching a terminal.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You are already paying for an AI agent subscription and want to bring monthly costs down without losing capability.
You use Hermes via Telegram or a CLI and want a faster, more organized workflow.
You are a solopreneur interested in using AI agents to surface business opportunities automatically.
You want to run local models alongside cloud models and route tasks to the cheapest fit.
You want a practical orientation to the Hermes Desktop UI before diving in yourself.
SKIP IF…
You are completely new to AI agents and need the fundamentals before optimization tactics.
You are locked into a different AI provider and have no plans to switch.
You are looking for deep technical architecture -- this is accessible and intentionally surface-level.
TL;DR
The full version, fast.
Hermes Desktop consolidates what previously required CLI commands and Telegram thread gymnastics into one polished interface. The core insight: every message you send includes all prior context in that thread, so a bloated single-thread setup multiplies costs by 3-4x. The fix is multiple slim sessions, profiles mapped to specific models, and pruning unused skills. The episode closes with a live demo of a cron job that runs a local Qwen model every 20 minutes to scan Reddit and X for problems the host is positioned to solve, then auto-generates prototypes for the best ones.
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.
150+ skills ship by default; each adds context overhead. Toggle off unused ones. Tool sets group skills into bundles.
19:19 – 21:44
08 · Messaging and Cron Setup
Messaging configures via UI now. Cron section gives one-click confirmation scheduled tasks were actually created.
21:44 – 28:09
09 · Reverse Prompting and the Brain Dump
Brain dump context and interests, then ask the agent for the best prompt. Live demo builds a morning brief pulling real headlines.
28:09 – 32:12
10 · Sub-Agents vs Profiles
Sub-agents are copies of the main agent for parallel tasks with the same skill. Profiles are for sequential work needing different skill sets.
32:12 – 37:05
11 · The Daily Business Opportunity Scan
Live cron demo: local Qwen scans Reddit and X every 20 minutes for founder problems, filters by skills, auto-builds prototypes.
37:05 – 39:03
12 · Local Models: Mac Studio vs DGX Spark
DGX Spark ($4,800, 128GB) is the current plug-and-play recommendation. Mac Studio preferred but mostly sold out.
39:03 – 41:59
13 · Reframing Cost as Investment
Stop comparing AI tool costs to Netflix. These are investments with ROI potential.
41:59 – 43:48
14 · The Real Way to Make Money + Closing
Aim agents at other people challenges. Find problems, stay lean as a solopreneur, build solutions.
Atomic Insights
Lines worth screenshotting.
Every message you send includes all prior context in that thread -- a single bloated thread can 3-4x your monthly AI costs.
Profiles mapped to model strengths (Opus for strategy, GPT-5.5 for coding, local Qwen for free research) beat profiles mapped to corporate roles in both cost and clarity.
The reverse prompt move: brain dump your context and goals first, then ask the agent what the best prompt would be -- it will outperform anything you would write yourself.
150+ skills ship in Hermes by default; each unused skill adds context overhead to every message; audit and disable what you do not use.
Cron jobs set via CLI or Telegram have no confirmation they were actually created; the Desktop Cron section gives one-click verification.
Sub-agents are copies of your main agent for parallel execution of one skill; profiles are for sequential work where each step needs a distinct skill set.
The highest-leverage use case: a scheduled agent that scans Reddit and X for founder complaints, filtered against your own skills, with auto-generated prototypes for the best problems.
Hermes allows hot-swapping models without agent reconfiguration; competing tools hard-code model bindings and require updates to access new releases.
The Artifacts section auto-organizes every link, image, and file sent to any agent session into a searchable second brain -- no manual filing needed.
Local models (Qwen 27b on a DGX Spark) handle research tasks for free, letting you reserve cloud spend for tasks that genuinely need the smartest model available.
Takeaway
Your AI bill is a context problem, not a model problem.
WHAT TO LEARN
Agent costs are high not because models are expensive but because one giant thread makes every message carry the full weight of everything said before it.
02Sessions and Context Management
Create a new session for each distinct topic or task; a single catch-all thread with months of history can multiply your costs by three to four times.
03Profiles and Multi-Model Workflow
Map profiles to model strengths, not job titles: use the most capable model for strategy, a high-limit cheaper model for coding, and a free local model for routine research.
06Skills, Tools, and Cron Setup
Audit the skills active on your agent and disable anything unused in the past week; each skill adds to the context size of every message you send.
Scheduled tasks set through a CLI or messaging app have no confirmation they were saved; always verify through a dedicated cron interface before relying on them.
07Reverse Prompting and the Brain Dump
Before writing any cron job or complex prompt, brain dump your context and ask the agent to generate the prompt for you -- the result will outperform what you would write yourself.
08Sub-Agents vs Profiles
Sub-agents run one skill in parallel across many tasks; profiles are for sequential workflows where each stage requires a different skill set -- choosing wrong between them wastes both money and time.
09The Daily Business Opportunity Scan
The clearest path to extracting real value from an AI agent: run a scheduled scan of social platforms for problems people are actively complaining about, filter for ones that match your skills, and build solutions for the best ones.
10Local Models and Cost Reframing
Before spending $5,000 on local inference hardware, prove to yourself that you can generate consistent value with cloud models; the ROI becomes obvious only after the workflow is proven.
Switching models in a well-architected agent takes seconds; if a task feels expensive or slow, the right response is switching the model, not rewriting the agent.
Glossary
Terms worth knowing.
Session
A single conversation thread with an AI agent. Hermes Desktop creates a new session for each chat, keeping context slim and costs low -- equivalent to opening a fresh chat window.
Profile
A fully separate AI agent within Hermes, with its own skills, personality file (soul.md), memories, and session history. Switching profiles is like handing the task to a different employee.
Sub-agent
A copy of the main agent spun up to run the same skill in parallel. Five sub-agents build five features simultaneously; they do not have independent skills or memories.
Cron job
A scheduled task that runs automatically at a set time or interval -- for example, a morning briefing at 9 AM daily or an opportunity scan every 20 minutes.
Artifacts
The Hermes Desktop section that auto-collects every link, image, and file exchanged in any session into a single searchable archive, functioning as an automated second brain.
Tool set
A named bundle of multiple Hermes skills and tools that can be activated together for complex tasks, rather than enabling them one by one.
Reverse prompting
The practice of brain-dumping your context and goals to the agent, then asking it to generate the optimal prompt or instructions for the task, rather than writing the prompt yourself.
soul.md
A plain-text personality and instruction file attached to a Hermes profile that defines how that agent thinks, responds, and prioritizes tasks.
Context window
The total amount of text an AI model can hold in memory during a conversation. A larger context window means higher costs per message; keeping it slim is the primary cost-control lever.
DGX Spark
NVIDIA's consumer-oriented inference device with 128GB of unified memory, priced around $4,800, capable of running open-source models like Qwen 27b locally.
04:04 – 06:10denseSession and context cost management
06:10 – 12:58denseProfiles and multi-model workflow
12:58 – 14:32steadyArtifacts second brain
14:32 – 17:32steadyHermes vs OpenClaw comparison
17:32 – 21:44denseSkills, cron jobs, and UI features
21:44 – 28:09denseReverse prompting technique
28:09 – 32:12denseSub-agents vs profiles architecture
32:12 – 37:05denseBusiness opportunity scan demo
37:05 – 39:03steadyLocal hardware: DGX Spark and Mac Studio
39:03 – 41:59steadyCost reframing and ROI mindset
41:59 – 43:48sparseMaking money and closing
The Script
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.
17px
Hermes desktop app just launched, but how do you actually go and make money, be more productive using Hermes desktop and Hermes agents in general? So today, I have Alex Finn on, and he gives you an entire walk through of how to use Hermes desktop app, all the nooks and crannies, explains everything so so clearly, and then we go through some use cases.
So stick stick around to the end where we show you exactly how you can set up Hermes agents that go and and build businesses for you. This is the most comprehensive tutorial on Hermes desktop app and Hermes agents on the Internet, and I can't wait for you to actually go learn, build, and have some fun.
Enjoy the episode.
Alex Finn, I need you by the end of this episode to explain to me why I need to get Hermes Desktop. I need you to walk me through Hermes Desktop.
I need you to show me use cases of Hermes that show me how I can make money, be more productive, And I need to you to explain me why I need to get off OpenClaw, why I need to shut down my telegram, and everything in between.
Alex Finn, can you do that by the end of the episode?
Greg Eisenberg, mission accepted. We will do that.
We will go through why Hermes desktop is the best AI agent experience on planet Earth. We'll go through every feature, every nuance, every secret, and then we'll show you the sauce.
We'll go through all the use cases you can do, every person watching this video so they can get tremendous amount of value out of Hermes agent. Alright.
Let's go. Alright. We'll do it.
Here we go. I'm gonna show share my screen. Great to be speaking with you, Greg.
I have been deep into Hermes agent desktop the last few days. So they released this a few days ago now. I wasn't expecting it.
I I believe this is, like, the moment that Hermes overtook Open Claw, to be quite honest. And as most people probably know, I was, like, the Open Claw guy.
I was at a poker tournament last night, and a guy was like, hey. You're the guy getting paid by Open Claw. Right?
And I'm like, no. I don't even talk about OpenClaw anymore. What are you talking about?
This is Hermes agent desktop. So before everyone was using Hermes and OpenClaw inside Telegram, inside Signal, iMessage, all of that.
I think this is the single best way to use Hermes agent now. I I I think I I don't use Telegram on desktop anymore.
I don't use any other experience. This is clearly the best way to be using Hermes agent. There is a lot here.
So, uh, few things. One, sessions all here. Right?
So before in Telegram, if you wanted some idea of sessions or multithreading, you needed to set up threads in Telegram, which is kind of, like, an advanced confusing thing to do.
I know a lot of nerds watching would be like, what are you talking about? That's easy. For the average person, setting up threads in Telegram isn't the easiest thing to do.
You gotta, like, create separate group chats and then add your bots and then, like, edit subjects. And here, you every time you talk to your Hermes agent, it just goes into a new session, and you can talk to it.
So it makes it really nice for, like, just separating out your context. I think the way most people use Hermes agent by default is they have, like, kinda like a mono thread. Right?
Like, just one big thread where they talk about everything in it. The issue is is you kinda, like, pollute your context with the 100 different things you talk about, and so your costs skyrocket.
Like, that's, like, the number one complaint with Hermes. Oh my god. I'm paying a thousand dollars a month.
You if you manage your context and your sessions well, you are not paying thousand dollars a month. You're not paying anything close to it. Alex, by the end of this episode, can you show maybe a few ways for us to get our Hermes under control, like some tips and tricks?
Absolutely. We'll do that throughout, then we'll recap at the end. I'll give you a whole bunch of sauce on saving money with Hermes as well.
Okay. So saving money with Hermes.
Number one, appropriate sessions. Context, especially if you're using, like, Opus, which has a million parameters, context can really inflate your bills.
Like, three, four x your bills. So if you keep things individual right?
Here's my content session. Right? Here's my talking about local, uh, model session.
Mode research is my stock research session. With the context nice and slim, I'm saving tons of money because the every time I send a message, it'll only be that thread itself. For those who are a little less technical, when you send a message to your agent, right, uh, if I were to say, generate me an image of a squirrel with Grok Imagine.
With this message, it sends everything said before that. So if you have one big thread like most people interact with Hermes agent, you're sending a massive messages every single time, which will cost you tons of money.
But if you take advantage of Hermes desktop with its multiple sessions and pinning an organization, you can even put sessions in folders. It keeps your context clean, means your messages are much more slim, which means you save a lot more money on your messages.
So that's one big thing there. The the session management on desktop, It's the best. It's better than Telegram.
It's better than the way 99% of people use, uh, Hermes agent. There's a lot more to it.
That's one of my favorites. One of my other favorites is now the organization of profiles.
If you can see down here, it's a little tiny. Maybe I can zoom in a little bit. Yeah.
A little bit. You can see a bunch of names down here.
We got default, Billy, coder, GPT.
Mez, librarian, Oracle, and Quen. These are all my different Hermes agents.
So for those not as familiar, in Hermes, there is this concept of a profile. You have multiple profiles.
Profiles are literally just separate Hermes agents. Each profile has a has a set of skills, has a soul dot md, which is their personality, has sets of memories, uh, has your history of sessions and chats.
They're literally just separate agents. And now with desktop, they're organized really, really nicely.
So you can just click, and it takes you to your other profile. Right?
I can just click, and it takes me to my gpt. Mez, which is my Hermes powered by GPT five five. I click on Quen.
This is my Hermes powered by Quen three seven sitting on my DGX spark on my desk. I click home. That's my default.
That takes me to Hermes. I named my default Hermes Hermes. I just like that name.
And that takes me there powered by Opus four eight. Before, like, if you wanted to switch profiles, you needed to, like, go into, like, your terminal and, like, do Hermes profile, Quinn entering, like, pops open a new it's, very hard to organize.
Now it's very neatly organized here so I can quickly okay. I want Quinn, which is my local model to do go do this for me. I want GPT Means, which is my Chad GPT model to go and code this for me.
Okay. Now I'm gonna go back and go back in my Opus one and have them do that. And now it's, like, really easy to multitask.
To give some examples of of when you'd wanna multi multitask, for instance, you know, Opus, very expensive model.
I think Opus four eight is the smartest AI model out there. It's also very pricey. So if I'm doing high level thinking and strategy, I'm going Opus.
Then I'll switch to GPT. Me's, which is my, uh, Chad GPT five five model, uh, on Hermes. Huge limits on Chad GPT.
Also, very good coding. So I'll come in here. I'll be like, hey.
Build me this app. Quinn, that's my local model. Free unlimited intelligence because it's running on my desk.
This is for, like, quick research. Hey. What's, uh, the top news in AI?
I could have a go and scour the web for the top news, and it's not gonna cost me a thing. Completely free. So you gotta know which profile is best for which tasks.
So is the way to think about profiles more around how fast or slow do you want, you know, a task to be completed or how expensive or cheap? Or is the way to think about it similar to how a paper clip works where you're kinda like designing, you know, your product manager, your marketer, your engineer, and you have different profiles for different team teammates.
Yeah. It's it's I the way I like to think about it is more on the model basis.
Each model has its own strengths and weaknesses. So we talked a little bit earlier, Greg, about saving money.
Models, if you use it in the right way, you'll save tons of money. Coding, I actually think Chad GPT five five is better than Opus four eight for coding, and you get way more limits. You get way more usage on your plan.
So if I'm gonna do coding, I'm gonna go to GPTVs, which is my Chad GPT five five profile. If I'm gonna do, uh, like, orchestration planning, doing in-depth plans, I wanna go to my smartest model, which is, uh, Hermes powered by Opus four eight.
So I'm more along the lines of what are the strengths and weaknesses of each model.
I'm gonna be honest. I'm not 100% bought into the concept of, like, having 50 different, like, product manager, graphic designer.
I I think that's a lot of I just think that takes up a lot of context in your meat brain on, like, okay.
Which one of these 50 profiles do I go to? Do I have a product manager? Do I have a designer?
Do I have a writer? I I I think that these models are smart enough and generally intelligent enough where you can basically get away with like, I can use Opus for designing a graphic image, then writing a YouTube script, and then planning my product.
So I more think along the lines of what are the strengths of each of the models, and and how can I use them to their strengths to get better performance and save money?
I like that take. It's basically, like, all that matters is the output and the outcomes. Therefore, just know which model makes the most sense for what you're trying to get accomplished and then use the right profile for that.
Yeah. I I used paper clip, and I think it's a very, very cool, well designed product. But it's bought into this the concept that you need to have an AI agent in every role.
You need to have a senior product manager who manages a design intern who manages the CTO, who manages like, and the issue is is that when you think about it in that lens, which you're totally fine to do because that's it's a more understandable lens for a lot of people. So if it makes more sense to you and you understand your organization better that way, then go ahead and do it.
The challenge is I I think there's inefficiencies there when it comes to cost. Like, if I have to go to my CEO and say, hey. Build me a a, you know, a thumbnail for my latest YouTube video.
And then they go to, like, the product manager who then goes to the intern. You have call after call after call looking at the context and the soul and the skills of each one of these.
And it's like, I could have just went to Opus four and said, hey. Build me a a, you know, an image. Like, it's it doesn't have to be that deep.
So listen. Everyone, you have your own workflows. Do it whatever way you want.
I just for me, I found the most cost savings and efficiency just having it for each model. So those are the profiles. I love going to the profiles.
I love being able to go in quickly talk to, okay. I have a quick research task. I'll go to Quinn who's living on my desk.
Give him a research task to research AI. Go to gpt.me's. Have them go build out an app for me.
Profiles, people should be leveraging profiles, setting up profiles for different models, start saving money that way, get better productivity that way. Lot of other really cool things here. Artifacts.
So artifacts basically puts all your images, files, and links in one place.
And so if you're anything like me, you've been trying to kinda MacGyver your AI agent into, like, a second brain, which it does really well and it works.
It's just kind of a system to maintain. And one of those second brain tasks I have is, like, when I'm coming across things, inspiration.
I see websites that are cool, tools that are cool. I'll take the links and give it to my agent and say, hey. Save this link for me.
File it away here. Do this for me. Now kinda done automatically with artifacts.
So you can come in here and see any links you've sent to your agent. Uh, you can see any, uh, images, media you've created or sent to your agent, any files you've sent back or forth.
They're all organized here, And now you can quickly organize them, search through them. So now I don't need to go and be like, hey.
File this in my second brain under this category and do this. I just send, like, a link. I got my librarian.
My librarian is just kinda my second brain organization agent. I'll just drop links into my librarian. I don't even tell them to do anything with it.
I'll just hit send. And now I can go to my artifacts, and all my links are just, like, tightly organized here. I can search.
I can say Reddit. It shows me all the Reddit ones. So artifacts is sneaky, like a really, really cool place to centralize all your links and images and files and everything.
That's really cool. They basically they basically productize the second brain for for an, you know, an agent platform.
Like, they probably saw how people were using it, and they're like, we're just gonna make this really easy, really visual. And that's what I'm liking, by the way.
I'm just I'm seeing this for the first time. That's what I'm liking so far with the whole Hermes desktop experience. It just feels like a more Apple esque product for the OpenClaw type of experience.
That's a really good way to put it because, you know, I I've switched over from OpenCLaw to Hermes.
And and and for the record, neither of these people are paying me. Neither of these people barely even talk to me.
I I I I this is purely for the love of the game.
And, you know, one of the Which is a big deal, by the way. You switching from OpenCloud to Hermes, like, where's the breaking tweet? You know what I mean?
Like, you you you like you said, you're a OpenCloud guy. You know?
This is a big deal. This is
I was the I was at the top you know, not to toot my own horn or pump my own tires. I was at the top of the game of open claw. This is like Michael Jordan going from basketball to baseball, you know, in 1994.
Like, that is the this is basically the equivalent of that, and I did it anyway with basically zero fanfare. But I I I have to I can't keep pushing things that I don't think are as strong anymore just because I'm getting a bunch of clicks and views. You know?
Yes. You know, it it it something I think a subtle shift happened the last couple months where I think the Hermes team, very much as you just kinda described, went more the Apple route.
Well, I think the Openclaw team went more of the Android route in a way where the Hermes updates have been very focused, polished, tested, and tied to really nice narratives that are focused on making life easier for the user.
While I feel like the Open Claw updates have been more kinda shotgun, put everything in it, don't test it quite as well, rough around the edges, Doesn't work quite as well.
We're trying to throw everything in it that will satisfy, you know, the power user, satisfy, you know, the the the people who are doing a thousand things at once.
And, like, the updates became very unfocused and worst of all, unreliable where I was updating, and I'm like, this is honors. I'm gonna break my open claw. And then it did break my open claw.
While Hermes is like their branding's incredible, but, like, their announcements, like, each update will be, like, once a week, and it'll be, like, the efficiency update. And everything will be very focused on efficiency and, like, around that narrative.
And as you kinda just said, it's, like, kinda the Apple route where it's just you know, it doesn't need to be a CLI. It could be a nice desktop experience that's pleasant for the average person to use.
Can you walk me through the skills and tools section?
Yeah. A lot of cool things here. This is a nice interface for all the skills that are installed on your Hermes agent.
One thing people don't really realize is out of the box, your Hermes agent has, like, over a 150 skills installed, and each one of those actually adds context to your agent. And so that's increasing your cost.
Now with this nice user interface, you can come in and you can turn off any skills that aren't necessary.
And you can also see the new skills being generated based on the actions you take. Not many people realize this.
They know Hermes is very self improving. They don't realize it's actually creating tons of skills in the background while you're not looking. Right?
Like, I have a three JS one file game because I build a bunch of three JS games just for the fun of it. YouTube news scripts, real time stock dashboard. Like, there's a lot of skills that get built here kinda in the background as you're doing things, and you can just turn off, turn on any ones that aren't relevant to you to to save money and increase efficiency.
Tool sets are groups of tools. So this is a new kind of concept being rolled out right now where, uh, you can group together many skills and tools all in one.
So if you're doing something complex, it groups them all together. Uh, I gotta play around more with this to be quite honest with you. Uh, but tool sets, something definitely more to be used because it's it's kind of a new concept inside Hermes where you can group together multiple tools and skills.
That is, uh, those. You also have messaging. So, again, a lot of really nice things here around just being able to quickly see and visualize the features of Hermes.
So if you wanna set up more messaging services, you just click it, set it up. Right? Before it was you had to go into your CLI, Hermes set up, navigate to whatever messaging service.
The average person does not want to use a CLI. Right? Like, I guess I I think there was this huge trend the last few months.
CLIs we're going back to CLI. CLI is the future. Everyone will be working in the CLI.
That doesn't change the fact average people do not wanna go in the terminal and do anything. Right? They don't wanna do anything.
I I get so many complaints on my YouTube videos when I do setup is where it's like, this is so easy. Paste this into your terminal. And I still get comments like, I don't wanna do that.
I have no interest in that. That's not that's not for me. So this makes it just a lot easier.
You never have to go into your terminal again. You never have to go in the CLI. You go in.
It gives you, like, exactly what you need to set up your telegram, and then it just works. So really easy to set things up here.
And then one other thing I I like is the cron section. So if you click down here on cron, you will see a lot of your cron jobs you have set up.
One of the biggest complaints I get when I do these videos is, Alex, you you said to do this to set up a a routine. I did it, but it didn't work. The next day, the the routine didn't happen.
The issue is is when you're using through things through Telegram or the CLI and you set up these cron jobs, for those who may not be as technical cron jobs, just schedule tasks your computer does, you don't really have any confirmation that they were created when you ask your agent, hey.
Every morning 9AM, do this, or every night at 10PM, do this. There's no real way to confirm and actually schedule those things. Now with the cron section, which is at the bottom, one click, you can see all your cron jobs.
You can see everything that's scheduled, and you can even create it manually here. So if you wanna come in and say, hey. Make me a morning brief.
Tell me about tell me all the unread emails from the night before.
Right? You can come in daily, deliver to the desktop, create the cron.
You can have 100% confidence this cron job is being created and will happen every day at that time.
Because cron jobs are and routines are so important to unlocking the value of Hermes, do you have any advice for how do you actually create cron jobs that are worded in a way that actually have a high likelihood of success?
Yeah.
Whenever it comes to doing things more effectively with AI or solving for, like, kinda unknown unknowns, you know, I know I can be doing this better, but I don't know how to put a better cron job or better prompt in.
The answer is always reverse prompting. Always. The answer is, I want to let's do it right now, actually.
Let's demo it live. I'm gonna do a new session. So say you're a new user, Hermes.
You wanna come and you wanna set up a cron job for, uh, every morning for a morning brief with relevant information in it. So you can go, I want to set up a morning brief with you for every morning at 9AM.
That's usually where people stop. But here's where you take it to the next level with reverse prompting. You go, my interests are generally AI, stock investing, tech, and the Boston Celtics.
Right?
What Controversial. Controversial.
The Boston Celtics?
I I would say so. You know? I think a lot of a lot of listeners of the podcast are big Knicks fans right now.
So you know? That's what I'm seeing in the from
yeah. That's what I'm seeing. Enjoy it while it lasts, Knicks fans.
It's not gonna last because Wemby's coming for you, and everyone in the country is rooting for Wemby on this one. Sorry, Knicks fans.
You get you got your one year in thirty years, so you have a mild amount of hope. Enjoy it. What so here's where the reverse prompt comes in.
Right? I just got you just lost, like, a 100,000 subscribers, by the way. I know.
Sorry, Grant. I owe you one. What would be the best prompt I can use to set this up with you?
So I did a couple things here. I did a couple very important things here. Number one, this is my favorite combo in the world, the brain dump to the reverse prompt.
You brain dump in everything about yourself, interests, goals, ambitions, skill sets, things like that.
So your agent has context around who you are, and then you reverse prompt. Based on what you know about me, what's the best way to do this?
The the the issue many people have is they just kinda wing it and tell the agent to do things. What they forget is you're talking to, like, superintelligence. Right?
You're talking to, like, the smartest thing on planet Earth. So why would you do things your way when you can get help from superintelligence and do it, like, the best way? So I'm gonna hit answer on that.
And instead of what most people do is, hey. Set set up morning brief that tells me the news on AI and the Boston Celtics and how Jayson Tatum is the best player in the NBA.
Now it's going to actually tell me the exact prompt to use.
So here's the prompt I'd use. And now it gives me this really detailed in-depth prompt.
Right? You are compiling Alex's daily morning brief. It's currently early morning eastern time.
Pull fresh information from the last twenty four hours. Do not rely on memory or stale data. Use web search and exact real headlines, prices, and scores.
Cover these four areas. And then it goes in-depth. AI, the three four most important developments in AI from the last twenty four hours.
You know, what this fix is, usually when you say things like, tell me the latest AI news, it'll give you headlines from, like, two months ago. Right? Or it'll even give you headlines from, like, a year ago because of the model's memory cutoff.
But now because of this better prompting, you'll be sure that it's from the last twenty four hours. Market and investing. Uh, yesterday's closed in premarket direction for the S and P, Nasdaq, and notable movers.
Flag one or two stock sectors with unusual news. Include any major economic events scheduled for today. Like, most people are not gonna remember to talk about the Fed CPI earnings or big names.
Tech. Two to three significant tech industry stories beyond pure AI. Right now, it's like going into extra detail.
And what happened now, and here's a big one, formatting. Lead with a one line vibe summary, use bold letters and type bullets. Like, if we we use the original prompt most people would use, hey.
Give me a morning brief to everything about AI. It just would have been a, like, a big dump of text. But now it's nicely formatted, so it's easy to read.
And so this is the format I go. Brain dump to reverse prompt. You're gonna get a lot better prompts, cron jobs, and output from, uh, your agent.
Great tip. Thank you.
Of course.
What else we got here? So that is the cron jobs. Really, really cool stuff.
Really amazing to use. You can see your context very easily, your context window very nicely at the bottom now. So you can make sure, oh, you know, our context window is filling up.
Let's do slash new to start a new session, and that'll clear that so you can save money there. You can quickly change your models. This is another really big strength of Hermes is they architected their agent away where you can ease you can switch models much easier than OpenClaw.
OpenClaw, like, every model is, like, hard coded in to OpenClaw. So you have to, like, ask them to switch models. And when new models come out, you have to wait for OpenClaw to update with that newest model.
Hermes, they updated it. They have the agent, uh, work in a dynamic way where it can just swap in models and thinking settings really, really easily. And the moment a new model drops, you can just swap it in.
And so now in here, we're taking the customization step further. You know?
Oh, I'm gonna do a simple task. Doesn't require much thinking. Let's use, like, Opus four eight with minimal effort or Haiku with minimal effort, things like that.
So another really nice way to save money with your agent. Uh, lot of other really cool things here. There's this agent window.
When you're go and you're you're multitasking, you got a bunch of agents doing things, you can see all your sub agents spin up. Usually, a question I get at this point is sub agents, what what's there between a sub agent in a profile?
So a real Hermes agent. So, again, the profiles have their own skills, personality, everything.
Sub agents are copies of the main agent. So when I go to, you know, Hermes and say, hey.
Go do research and then build an app based on that research. It'll spin up subagents. Those subagents are just copies of the main agent.
So it has the same skills, the same personality, the same everything, But profiles have their own skill sets and context and memories and everything.
And so the way you wanna think about it is, okay. For what you're doing, does it require a separate set of skills? If it does, then spin up a new profile.
Is what I'm doing can I just use that same set of skills? Just do it multiple times at once? Alright.
Spin up sub agents for that. So agents down here is where you can see those sub agents work.
What's a what's a concrete what's a concrete example of of that? Like, when when would I be in the case where I'd wanna use multiple sub agents?
Like, what sort of task?
So say I am wanting to build out a micro SaaS, and I want to it has many different features, maybe five or six different features.
Maybe it connects to the x API to pull tweets, and then it formats the tweets in a nice way. And then it allows you to share those tweets on Twitter and then send your own version of that tweet to us.
So there's, five or six different features as micro SaaS. I would wanna use sub agents for this. I would go to my Hermes.
I'd say, build me this micro SaaS. Here's the 10 different features. Spin up sub agents to build out each of those features simultaneously.
And then I'll go and spin up, and they they all have the same skill set, right, which is just coding. I just need a coding agent, but I just need multiple versions of that coding agent.
And so it'll spin up five or six sub agents to build out that micro SaaS because I just it's doing many things at once. But maybe I wanted to do something like, okay.
I'm working on a YouTube video, and I need a script. I need a researcher that will get the information that will feed into that script.
I need, you know, a thumbnail, so I need a graphic designer or an agent that is connected to the Grok Imagine API.
For that, I would want multiple profiles. Right? I'd want, you know, for the researcher, I'm gonna use Quen, which is my local model.
Say, hey. Research this script, then hand it off to hand it off to Hermes, which is my Opus four eight model, and have them actually write the script based on your research.
And then hand that off to gpt.me, which is my Chad GPT model, and have them use a Chad GPT image generator to build the thumbnail. That would be a good use case for multiple profiles.
So the way to think about it to kinda recap, different skill sets needed, use different profiles. One skill set, but just on multiple strands at once, use sub agents.
With that being said, we talked about sessions, so managing what you're talking about, managing context. We talked about artifacts, organizing all of your files, images, media, all of that.
We talked about skills, managing your skills, making sure the right skills are turned off and on. We talked about cron jobs, so the scheduled tasks.
We talked about context models. Should we kinda get into the actions, what you can do with all this, putting it all together?
Yes, please.
Alright. Let's do it. I will show you one of my use cases I do.
This is real life. I have my agents doing this all the time.
And, uh, let me just pull this open over here. I'm going to refresh this.
So one thing as a solopreneur I like to do is I like to solve challenges.
Right? You know, that's pretty much the best most ethical way to make money in this world is find other people's challenges and then solve those challenges. And so with my agents available to me, you've never had more power or leverage to find challenges other people have.
And so I have a cron job inside my Hermes agent. Let's see if I can show it to you.
Is it in this agent here? Uh, daily AI business opportunity scan. Every twenty minutes, I have my agent going and using the Quen three seven model locally.
You don't need a local model for us. You don't need a DGX spiral. I'll give you an alternative in a second.
But because I have this unlimited free super intelligence on my desk, I have to do it every twenty minutes. I have a go this is kinda like IdeaBrow.
These two would fit very well together. IdeaBrowser and this AG.
You put them together, you got magic.
But what it does is by the way a lot of people are using Idea Browser with Hermes, but that's a whole other separate, uh, episode I think I I could do if people are interested.
That's an hour episode there. You connect Hermes to IdeaBrowse, and you have it going just building out each Yeah. Now that's sauce.
Yeah. Uh-huh. But what I have it doing is I have Quinn every twenty minutes going and reading Reddit, reading x, searching for other people's challenges.
And I'm gonna show you what it builds for me here, and I'm gonna show you how I take it a step further. Now if you wanna emulate this, you 100% should. You don't need you don't need to do this every twenty minutes if you're using, like, a cloud.
The the reason why this works so well with Quen locally is it's free, so I can have it do it as much as I want.
If you're using, like, ChaiGBT or Opus, you'd be paying out the wazoo for this because it costs money every time you do something. So maybe you do this just once a morning to save a little cash. But here's the output of that, this cron job.
I get this custom dashboard it built, which does it look like AI slop?
Yes. It does, but I don't care. The the the the styling of it doesn't matter to me.
But what I get here is Henry going in, which is actually Quinn three six on my DGX spark, does it every twenty minutes, and it finds challenges from people on Reddit and X, gives me the source thread so I can see what comments the the founder said the whole agency was basically four APIs in a trench coat.
And then because my agent knows so much about me and my skill set, it tells me why I'm in position to fix this problem and what should be my first move to solve that challenge.
And so now I kinda have this automated business researcher that's going, knows me to a t, knows all my skills, everything, and finds me challenges to solve.
Right? It knows about my community, so it knows about my assets, what businesses I have, Vibe Coding Academy. And taking it even a step further, when it finds it appropriate, it even adds this button here.
Henry built a pro so it automatically built me a prototype for this business idea. And so if I click that, micro SaaS prototype built for that idea it had.
And so now I can be like, okay. Is this interesting? Let me test this.
Oh, actually, I think there's something here, and then I can productize it and make it shippable. This is kinda how being a solopreneur builder works.
You find challenges. Because you're running efficiently and you're only one person, you don't have much overhead, you can go over these tiny you can go over these tiny slim markets and make a few bucks off of it, and that's enough to pay the bills.
Right? So this is a a really nice way I have my agent acting as an employee and finding me opportunities around the clock.
Side note. So this is really cool. Um, at the very least, your creative juices flowing.
Right? So it might not even be this idea, but you might be like, you know, I'd like this. I you know?
But if it was in another market or if it was an app instead of this but my question to you is, in terms of local models, is your recommendation that people go out and buy, Like, what should they be buying basically, and why should they be buying it?
My recommendation up to this or up until recently was the Mac Studio. The Mac Studio from a speed and bandwidth perspective isn't as strong as NVIDIA hardware like the Spark and other chips, but you got a lot more unified memory, which allowed you to run much cooler models and much better models just a little bit slower.
The challenge is is the Mac Studio is pretty much sold out on any basically, every level. You can get it right now, but it's small amounts of memory.
It's not like the big the big boys that allow you to run really good models. So now I think my recommendation for most people, unless you can find, you know, a Mac Studio secondhand or something like that, the DGX Spark's a really great plug and play device.
You you just you don't even need a monitor to set it up. Like, you just plug it into the wall, and it works. It has a 128 gigabytes of unified memory, which basically means you can run any model that goes up to a 128 gigs.
So Quen, I think 27 b works really nicely. Any of the new Nemotron models.
So NVIDIA got into the model game recently. Now they're putting out their own open source models. They're building it with the the spark in mind.
So I think if you want plug and play something to run inference for you to sort of run your models, the the DGX Spark is a great option. I prefer the Mac Studio because it's running on the device I have right here, and you can use it.
It's like a really nice user interface, the Mac OS. So that's really nice. But, uh, the Spark is a is a good plug and play option if you don't have access to, like, a really good Mac studio.
Cool. And and just quickly, like, how much is the price on on that baby?
They just raised it, I think, to $4,800, which seems like a lot, and it and it is.
Everyone's in different financial situations. Of course. It for sure is.
My kind of thesis, which I think is playing out in front of us right now, is we are going to have hardware bottlenecks for a very long time now because the AI build out's happening.
And so memory is sky I just bought a memory card for my Nintendo Switch two. It cost me, like, $100. It's, like, the craziest thing ever.
Like, the the cost of memory is skyrocketing, all hardware, everything like that.
And so hardware is getting more expensive. I think, you know, getting into local models in your own hardware is a very good idea in my opinion. Playing around with LLMs locally is a very good idea in my opinion, not only for functionality wise like what I just showed you here, but also for education and learning wise.
So, you know, I think if you can find the room to afford a DGX Spark, I think it's a really good investment.
My take is play with Hermes, download the app, show that you can that you can make this as a part of your daily solopreneur or, you know, just lifestyle.
And once you've gotten your hands a little dirty, then you've proven to yourself that if you have the $5,000 that you, you know, you deserve it.
Like, don't be one of those people that goes and buys the latest and then it sits on a shelf. That's my take.
You deserve it. That's the takeaway. You deserve it.
You owe it to yourself. Get yourself a spark, man. You you owe yourself a spark if you actually
go and and and use this thing. Right? Like, understand how to use it and and play with it and actually show, like, show that you can create value for yourself.
Because then at that point, you're gonna be like, $5,000? Well, if I'm gonna make 10 x that, then it's a no brainer.
Right?
Yeah. I think the it's always like the number one question.
Yeah. Oh, how much does that cost? What's that?
How much is that? I think people have this weird framing around cost when it comes to AI.
Like, they see the cost of Claude, and they see it's, oh, $200 a month. That's really expensive. It's very they're comparing it to, like, the cost of, like, Netflix and, like, Xbox Live and Amazon Prime.
Right? Those things, they just suck your money and life forces away.
They're not adding to your life. $200 for Claude, $5,000 for DGX Spark.
These are investments in yourself to create more value in the world. And you should, if used the right way, see an ROI. So I would just encourage you to think through different lenses when it comes to costs for these different things.
Alex Finn, before we go, is there one question I should have asked you about Hermes desktop app, Hermes agents in general that I haven't, um, that you think that the audience would benefit from me asking,
or did you cover it all? I covered it all. I think if you were to ask me a question I'd get more specific on or spend more time on is how the hell do you make money with Hermes?
And the answer would be around a lot of what I showed you to the end there, finding challenges, finding opportunities to solve.
And I think if your watches at home, you know, instead of using these agents to kinda play around and experiment with I think you should be using them to solve other people's challenges because that's how you're gonna be able to create the most value for yourself and the entire world.
Alex Finn, ladies and gentlemen, thank you so much for hopping on on short notice, sharing the sauce with us as always. I'll include links for where to find Alex in the show notes, the description, his acts, his YouTube, and everything else in between.
Thank you so much. You always hype me up, and I love that about you.
Gotta be fun. You gotta have fun doing this. A 100%.
It is fun. Right? Like, how lucky are we that we can play with these tools, that they're coming out so often?
I like that you're not afraid to share your opinion in real time as new information comes in. And so I like that about you.
And, you know, from my perspective, it's like I like playing with everything and then and seeing what what what's best for me. So thanks again.
Of course. It's a pleasure as always. See you soon, man.
See you.
The Hook
The bait, then the rug-pull.
Alex Finn declared it the moment Hermes overtook OpenClaw. In a 44-minute screen-shared walkthrough, the agent power-user who built his reputation on OpenClaw explains why he switched -- and hands over the exact playbook he uses to run automated business research, a multi-model workflow, and a cron job that builds micro-SaaS prototypes while he sleeps.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A solo walkthrough of five open-source GitHub repos getting traction right now, an AI writing editor, an agent-run CRM, a video-editing agent, a skill security scanner, and a phone-controlling harness, each with the exact install command and the first small workflow to try.
Greg Isenberg names the role he thinks AI agents are about to make the most valuable job in tech, and hands over the folder structure, tool stack, and 30-day plan to become one.
Greg Isenberg and developer Vinny break down WebMCP, the experimental browser feature that lets any AI agent search, compare, and buy on a website without scraping the page, then price two startup ideas built on it.
Cloudflare quietly built the payment rail for AI agents to pay tiny fees for web resources, and that rail is about to spawn a whole generation of businesses selling clean data, agent-readiness, and packaged expertise.
00:00
01:23
01:50
03:11
04:19
04:51
05:22
05:54
06:25
07:29
08:01
09:04
09:35
10:06
10:37
11:09
11:40
12:11
12:42
13:45
14:16
14:50
15:26
16:02
16:38
17:14
18:25
19:37
20:13
21:25
22:08
22:56
23:44
24:32
26:08
26:56
27:44
28:26
29:01
29:35
30:10
30:45
31:19
31:54
32:16
32:25
32:33
32:54
33:28
34:01
35:08
35:41
36:14
36:48
37:05
38:18
38:48
39:03
40:11
40:38
41:11
41:41
41:59
42:42
42:51
43:18
43:39
24:43
35:18
44:47
28:58
44:21
34:10