09:41Mark Kashef · Tutorial
You're Ignoring the Two Best Features in Claude Code
A 9-minute screen demo of /power-up and /insights, two Claude Code slash commands that most users have never touched.
April 3rdThe argument in one line.
Moving from AI-as-search-engine to AI-as-personal-agent is not a technology problem but a one-day setup problem: cancel your meetings, dump your workflows into Codex or Claude Code, and build the skill files that make the system run without you copy-pasting between tabs.
Read if. Skip if.
READ IF YOU ARE…
- A solo creator or operator who still copies and pastes AI outputs into other tools and feels like nothing is truly automated.
- Someone at AI Layer 1 or 2 who wants a concrete, shown-on-screen path to Layer 5.
- A product manager or knowledge worker who wants to see a real personal-OS demo, not a theory.
- A newsletter creator who wants to collapse the newsletter-to-social-post workflow into one command.
SKIP IF…
- You want AI to generate your content from scratch — this system still requires your voice, your strategy, and your last-10% human touch.
- You are not willing to spend one focused day building workflows before you see results.
The full version, fast.
The bottleneck to AI productivity is not model capability but the absence of a personal operating system on top of it. The guest built a skill-file repo in Codex covering every recurring job: newsletter editing chained with a last30days research skill and a no-AI-slop pass, social distribution to four platforms via sniffed internal APIs, a weekly strategic brief, and a personal adviser that reads a live Google Doc strategy plan. His 5-layer framework maps the journey from ChatGPT for Q&A all the way to an agent layer that does knowledge work without manual routing. The concrete week-one action: switch from the default ChatGPT interface to Codex or Claude Code, brain-dump your existing workflows, and let the system build its own skills.
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 →Who's talking.
00:00hostMarina Mogilko
00:00guestPeter Yang
Where the time goes.
00:00 – 01:09
01 · Hook + Guest Intro
Cold open with 'it almost feels like cheating'; guest framed as solo operator running 140K newsletter with no team.
01:09 – 02:00
02 · Self-Improving Skills
After every conversation where AI does not one-shot it, ask it to update the skill file — then review the changes.
02:00 – 07:13
03 · The 3 Top Creator Workflows
Newsletter brain-dump via SuperWhisper into Codex; social posting to X/LinkedIn/Threads/Substack Notes via sniffed APIs; personal adviser skill with analytics access.
07:13 – 08:02
04 · Sponsor Block (Transkriptor)
Transkriptor meeting transcription sponsor read.
08:02 – 10:43
05 · Video Editing + Strategic Brief
Codex reads transcript to flag awkward pauses and pull spicy quotes; weekly brief auto-generated for YouTube, Substack, and competitor channels.
10:43 – 12:15
06 · The 5 Layers of AI Adoption
Pyramid framework: Layer 1 everyday answers, Layer 2 daily work with projects, Layer 3 prototyping, Layer 4 personal apps, Layer 5 personal agent.
12:15 – 15:05
07 · Live Demo: Personal-OS in Codex
Screen-share of skills repo: newsletter-edit + last30days + no-AI-slop chain; podcast-prep, thumbnail-title-copy, social-writer, sponsor management.
15:05 – 19:09
08 · Personal Adviser Skill Deep Dive
Skill file shown live: reads learnings.md for persistent memory, reads plan Google Doc with goals/principles/energy audit before every response; skill-editor keeps files to one page.
19:09 – 21:50
09 · Live Adviser Demo
Real-time demo: Codex reads the Google Doc plan and describes its structure (three tabs: Journal, Plan, Research) without revealing sensitive content.
21:50 – 23:01
10 · AI Chief of Staff Discussion
Hermes Agent (Nous Research) as Telegram-native chief of staff; inputs from Slack and Granola transcripts; weekly accountability brief.
23:01 – 24:38
11 · How to Go from Level 2 to Level 5
Three steps: switch to Codex or Claude Code, brain-dump your workflows, build skills and integrations.
24:38 – 26:57
12 · Life in One Year + AI Fears
Era of solopreneurs; fear of becoming dumber offline; kids and the loss of critical thinking fundamentals; solution: make kids start businesses and learn to fail.
26:57 – 28:48
13 · Creator Economy + AI Slop
AI makes slop easy and viral; the counter-move is finding principles and values that make you unwilling to publish slop even if it would perform.
28:48 – 29:34
14 · Action Plan: What to Do This Week
Get ChatGPT Pro, download Codex, build one workflow. Stop consuming AI news; switch to build mode.
Lines worth screenshotting.
- The gap between AI Layer 2 and Layer 5 is not intelligence — it is whether your AI can write directly to Google Docs, post to four platforms, and read your analytics without you pasting anything.
- A personal adviser skill that reads a one-page Google Doc of your goals, principles, and energy audit gives every AI response a strategic guardrail, not just context.
- Keep all skill files to one page so you can actually read them — AI-bloated skills create a system you no longer understand or trust.
- The self-improvement loop is simple: after any conversation where the AI did not one-shot it, say update the skill file based on our exchange — then review the diff.
- 80% of the time spent in Codex or Claude Code involves no coding — it is just a more powerful interface because it can act across other apps without copy-paste.
- Open-source skills on GitHub like last30days which searches Reddit and X are drop-in accelerators — you do not have to build the research layer from scratch.
- Posting to Substack Notes via Codex required the AI to sniff internal browser APIs because no official API exists — the agent did it in one session.
- A weekly automated brief covering revenue, content performance, and competitor channels replaces roughly one hour of manual analysis.
- Skills are not prompts — they are text files with instructions, examples, guardrails, and integration pointers that persist across every conversation.
- The creator economy slop problem is a values problem, not a technology problem — AI makes slop easy, but it also makes human-crafted work more distinctive.
- A learnings.md file inside a skill folder gives the AI persistent memory of your preferences without relying on opaque system-level memory you cannot read or audit.
- Building personal apps to 80% quality now takes a few hours — getting them to production-polished takes days. The gap is taste and testing, not coding.
- Getting to Layer 5 requires one blocked-off day to brain-dump all your manual workflows, not a slow ramp over months.
Build the system once, let it run.
WHAT TO LEARN
The difference between feeling like AI is cheating and feeling like it barely helps is almost always a one-day setup problem, not a model problem.
- Skill files are what separate a Layer 2 user from a Layer 5 operator — a skill file is just a text file, but it gives the AI a job description, examples, and guardrails that persist across every session.
- Keep every skill file to one page so you can read it, understand it, and catch it when the AI adds slop or duplicate instructions during self-improvement.
- The self-improvement loop is concrete: after any conversation where the AI did not one-shot the output, ask it to update the skill file based on your exchanges, then review the diff before saving.
- A personal adviser skill backed by a one-page Google Doc of your goals, principles, and energy audit gives every AI response a strategic guardrail — not just context, but a filter for decisions.
- Switching from the default ChatGPT interface to Codex or Claude Code is step one because only those environments can write directly to Google Docs, post to platforms, and read your analytics without copy-paste.
- Open-source skills on GitHub for research or style are drop-in accelerators — you do not have to build every layer from scratch.
- The one-day setup cost is real, but the compounding benefit is that the system gets better with every session without requiring you to rebuild it.
- Building a personal AI that knows your decision-making style requires explicit inputs: your meeting transcripts, your messages, your past decisions — the AI cannot infer your strategy without access to how you actually work.
Terms worth knowing.
- Skill file
- A plain text file containing instructions, examples, guardrails, and integration pointers that tell an AI agent exactly how to perform one recurring job. Kept under one page so the owner can read and audit it.
- Layer 5 (personal agent)
- The top tier of AI adoption where skills and API integrations replace manual copy-paste routing — the AI takes input, performs work across external apps, and delivers a final output without the user switching tools.
- learnings.md
- A markdown file inside a skill folder where the AI records strategic patterns and preferences from past conversations, giving it persistent memory the user can read and edit directly.
- Hermes Agent
- An open-source, self-improving AI agent by Nous Research with persistent memory and multi-platform messaging support, used here as an AI chief of staff accessible via Telegram.
- no-AI-slop skill
- A skill file that passes a draft through style rules designed to remove generic AI phrasing and tighten the writing back to the creator's voice before publishing.
- last30days
- An open-source skill found on GitHub that searches Reddit, X, and other platforms for recent content, used to add fresh research to newsletter drafts.
- SuperWhisper
- A desktop speech-to-text app that records up to 10 minutes of continuous voice input, used for voice brain-dumps before passing transcripts into Codex.
Things they pointed at.
02:23toolSuperWhisper
02:28toolCodex (OpenAI)
12:26toolClaude Code
02:54toolTypefully
22:01toolHermes Agent (Nous Research)
23:17toolGranola
11:06toolLovable
11:06toolReplit
Lines you could clip.
00:00
“It almost feels like cheating. I fear that I'm getting dumber and lazier.”
06:27
“The last 10%, you gotta add your human touch to it. You can't just AI-slopify everything.”
12:46
“80% of the time I'm in Codex, I'm not coding. It's just more powerful.”
24:56
“This is the era of solopreneurs. You can get so much done with agents.”
29:10
“Unlike an employee, it's never gonna leave you, and it's only getting exponentially better.”
Where the conversation goes.
01:09 – 02:00denseSelf-improving skill files
02:00 – 07:13denseCreator workflow automation
10:43 – 12:15dense5 Layers of AI Adoption framework
12:15 – 19:09densePersonal-OS demo in Codex
21:50 – 23:01steadyAI chief of staff concept
24:38 – 26:57steadyAI fears and societal reflection
26:57 – 28:48steadyCreator economy and slop problem
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.
00:00It almost feels like cheating. I fear that I'm getting dumber and and lazier. Like, if I'm on a flight and there's no, like, Internet connection, like, that I just don't feel like working anymore.
00:08That's Peter Yang. He spent a decade building products at Reddit, Meta, and Roblox.
00:13Now he runs a newsletter for 140,000
00:15people with no team, almost all of it through AI. Unlike like a employee is never gonna leave you. And it's only getting exponentially better.
00:23Last 10%, you gotta add your human touch to it. And you don't have to be a creator to make this. I see a lot of creators.
00:27They they like to get AI to generate, like, 10 posts per hour, and, like, the slop goes viral. This is heartening. For someone who's trying to stay ahead in this AI era, what should they be doing?
00:37So step number one is use codex or a clock code. And step two is to
00:42So I really wanted to make this podcast practical for everyone who's watching. There is something that's going on in AI right now that everyone is talking about as self improving. And you recently build cloud self improving skills, and the whole narrative is like, stop prompting your AI.
01:00Make it figure out what to do next. Can you talk about that? And is self improving a reality for people who are just using AI versus researchers?
01:09Yeah. There's, like, different ways to do it. I think, uh, first of all, like, you know, I think all your audience knows the skill is just like a text file with a bunch of instructions.
01:15Right? And, yeah, I have skills for making my podcast, for editing my newsletter posts, and so on. Most basic ways to do a self improvement is after you use a skill and you have a back and forth conversation with AI, because it never gets it right in one one shot.
01:28Then you just say, like, hey. Based on our conversation, can you please update the skill to account for you know, to try to get to it in one shot faster. Right?
01:36And and they'll make a bunch of changes, and then you should review it. And then, hopefully, you'll do it better the next next time. When I integrated AI into every part of my work, it almost feels like cheating.
01:46And a year before that, you said you were drowning in meetings and projects.
01:50What are the top three things that you deployed in this year that dramatically changed how productive you are? So I well, I can tell all my creator work since we're both creators.
02:00And, you know, being a creator is like a lot of repetitive work. Right? Like, a lot of repetitive copy and pasting back and forth and, like, changing things from, like, a newsletter post to, like, a like, YouTube description to, like, social post.
02:10There's, like, a lot of kind of, like, different formats. And I I I just decided to spend one day, like, just, like, you know, no meetings and just sat down with Codex and be like, I I basically just brain dumped all my workflows to Codex, and here's how I do it manually.
02:24And I use this thing called Whisper for just, like, brain dump Yeah. My to my voice. So there's, like, a couple major ones like, you know, prepping the podcast, post production.
02:32There's editing my newsletter posts. There's posting to the various social media platforms. And then another one is, like, just, like, setting up an adviser to kinda give me, like, business advice, like, checking ideas.
02:42Is it, like, an agent that's running? It's it's it's just like a scale that's listening to a Google Doc file that has, like, a bunch of personal information about my business, and then it kinda gives me advice.
02:52Uh, talk to me about, uh, posting on different platforms. Yeah.
02:56How does it work? Do you dictate something with your voice, or does it see what kind of videos you produce and repurpose them? Uh, I have some, like, uh, cron jobs, which is basically, like, you know, like, this game morning briefings, weekly reports that send me how, like, other channels are doing, what kind of videos are popular.
03:11Now I post to social media platforms through Codex directly. Be before I was using this tool called Typefulli, these, like, social media tools.
03:18Mhmm. But, like, some of the platforms I posted, like Substack notes, they don't have any kind of, like, API integrations.
03:24Right? So I I just told Codex, hey. I I I I wanna schedule posts to Substack notes.
03:29Can you help me figure it out? And then it went to the browser, and it sniffed out some internal APIs that they're using to post. So it sniffed out internal APIs and then But it could also do computer use.
03:38Right? It can it can also do computers. Yeah.
03:39So it can do both. But now but now I can post to, like, four plan I can post to x LinkedIn threads and subseq notes at the same time.
03:47And in in in the skill, it has, like, some nuances. Like, for example, on x, you wanna tag the person that you interviewed. Right?
03:54Mhmm. You wanna find the handle, but on LinkedIn, you cannot really tag. It's it's kinda hard to tag.
03:57So, like, I I kinda just bring dumped all the codecs, and now it can actually do the tweaks for for me, so it kinda saves me a lot of time. Do you get the final approval before it posts or it just goes ahead and posts by itself?
04:09I still draft it. I I don't just get it to generate copy to post. So I I still I I just, like, you know, use voice before and bring bring dump some random thing into codecs.
04:17So you talk to it. It Yeah. Takes your voice and just, like, creates different social media posts.
04:23It creates different social media posts. It also has my, you know, examples of posts that I've done that have gone really viral. Mhmm.
04:28So it can help me, like, clean out my brain dump into, like, a more polished post. Yeah. And then I'll take a look and I'll be like, okay.
04:33Go post it to all four platforms.
04:35Does it have access to your analytics to assess what's going on right now?
04:39Uh, it does. Yes. It has access to analytics from various platforms and also Typefulli, the the thing I use, also has analytics.
04:45So it is some analysis. Like, for example, it found that on LinkedIn I don't know if it's growing up, but, like, if you just do a post with a bunch of links to all your tutorials and stuff, it tends to do better than if you do a single one. So I I I had no idea about that, so I'm gonna do more of that now.
04:59What about newsletter? Is it the same process where you talk about what you want to have in your newsletter and it creates it? Newsletter depends on what kind of post I'm making.
05:07So, for example, like, um, I think by the time this goes out, I have published a post about why I'm leaving my job. So that one, I just kind of, like, went on a walk and and kind of dictated to Whisper Flori's voice tools and and kind of just, like, brain dumped my real thoughts and then To Codex direct into Codex.
05:23I don't know. I could do it to codex, but I I use this tool called, like, SuperWhisper
05:28that records, like, ten minutes of my brain dump, and then I copy and paste the whole thing into codex. Because codex I'm just worried, like, if I dictate for ten minutes, codex will get confused. That that's my worry too, or something stops working and it just keeps processing and nothing happens.
05:42That happened to me a couple times. And it's not just Codecs. I think it's the context field that is not designed for a ten minute prompt.
05:48How many how many edits do you typically do with, uh, in this process?
05:53It's interesting because, like, with AI, I'm kinda giving a feedback for it to edit. It's actually hard to just edit yourself and just give it feedback. Right?
06:00Mhmm. So, like, I I I brain dump it. I give it a codex.
06:02It does a bunch of editing, and it uses my best examples to edit. And then, inevitably, there's something wrong.
06:09So I have to read through him, and and I can kind of voice takes a bunch of feedback, and it doesn't let her pass. Mhmm. And then maybe the last 10%, I have to manually change change stuff.
06:17Yeah. Like, it's it's actually important. Like, I always emphasize, like, the last 10%, you gotta add your human touch to it.
06:22You can can just, like, AI slopify everything. Basically, your written content is
06:27Yeah. Created with the help of AI. Talk to me about the strategic, uh, thing that you're doing.
06:32Is that, like, in the morning, it briefs you on what's performing, what's not? Or Every week, it sends me a brief about
06:39how much money I made this week and and and also, like, how did my past thirty days content perform. Mhmm.
06:46And also, like, all the other channels that are make similar content. Is there any kind of outliers? Mhmm.
06:51You know? So it and it does does does that for YouTube and also for Sub stack. Mhmm.
06:56And and for Substack, it doesn't have APIs, it has to use browser use to look up everything. Yeah. So that would probably take me, like, an hour to do manually, but, like, it it does it.
07:04If your days are back to back meetings, you don't want to spend extra time forty minutes after each meeting reconstructing
07:11what was decided, who owns what, what the next step actually is. Transcriptor records, transcribes instantly, and generates structured summaries with clear next steps.
07:21There's also a search function across all of your past meetings. You ask a specific question. It pulls who said what, which decision was made, and when.
07:30I talked to founders who actually use it for onboarding using archives of the past six months of calls instead of rewriting everything from scratch. Enterprise grade security, so it works whether you're a five person startup or a larger org. Link is in the description.
07:44Sign up with your work email. You get free three hundred minutes. You mentioned some of your editing is done with AI.
07:51Yeah. Is that where you do you dump a transcript and it edits it out, or what's the process? Yeah.
07:57For video editing, I feel like maybe I haven't gotten deep into it yet, but I still use, like, a human like, a video editor.
08:02But the transcript is very useful because I dumped the transcript into Codex, and it tells me, here's the moments where, like, you had, like, awkward pauses or, like, stuff you should cut. Here's the timestamps.
08:12And also for, like, the intro, here's kind of, like, we should pick, like, here's, like, the the spicy quotes Mhmm. From the interview that you should Yeah.
08:19It's it's fascinating how this has become so much faster versus a year ago. Yeah. It's just, uh, yeah.
08:25It's a different reality these days. So you've described all of these things that you've built. How often do you build something new?
08:32Pretty much every single time I talk to AI is something something changes. Right? Like, I'll I'll take some skills or something changes, but are you somewhere, like, from zero to one?
08:40Mhmm.
08:41Well, my hope is that I can streamline a lot of my content production. And and, you know, for the audience, I'm not gonna slobify everything. I'm I'm still applying taste, but I wanna streamline a little bit so I can save more time because I actually wanna be a builder because, you know, I I spent a decade of my career just building products inside big companies.
08:58And I want to now that we have all these agents and tokens we can use, I wanna, like, build different tools, like open source tools or things that, you know, solve my own problems. Mhmm. Yeah.
09:08So I wanna do both. Don't wanna just be a creator. I wanna be a creator and a builder.
09:11Nice. So I'm gonna spend more time doing that. Mhmm.
09:13Yeah. How do you make decision when to build something and when it's
09:17okay the way it is?
09:19Um, it's usually I I I feel like, you know, it's best to kind of build stuff to solve your own problems. So, like, for example, I I built, like, a fitness app. I mean, everyone's building fitness apps, but I I built a fitness app that helps me track my workouts, and then it has MCP to kinda, like, send me emails on my health.
09:36Mhmm. You know, that's one thing I built. I'm doing something else right now.
09:38You know, in Gmail, like, you get a lot of scam Yeah. Scammers, and and, like, my my parents have no idea. Like, they like, you know, their bank sends them an email, but it's actually, like, xyz@hotmail.com.
09:49Right? So I'm trying to build, like, some sort of extension to just highlight that this is a scam. Mhmm.
09:54How long does it take you to build something like that? I think to get to 80% there, it probably just takes a couple hour hours. Mhmm.
09:59Because you just make a plan and you get coders do it. But to actually make it good and to test it and polish it, that that takes more time. That takes, like, a couple days or, like, you know, depending on where your quality bar is.
10:08Yeah. Like, where you wanna go. Before we start building something, you have this system with five layers of AI adoption.
10:15Can you walk me through them? And let's decide which layer do we want to land on today. I made this five layer thing actually, like, six months ago.
10:22So most people are in layer one, which is just using AI for everyday answers. They're using ChatGPT and Claude and asking questions. I think the second layer is, like, using AI for daily work.
10:32So you're still in ChatGPT and Claude, but you're using projects and other things to, like, um, to not do too much copy and pasting, to, like, set up projects for different things, like, I don't know, giving you a life advice, automating your content stuff. But, um, because it's in Chattypedia and Claude, you still have to copy and paste the output to somewhere else.
10:49Right? So it's still, like, a lot of going back and forth. Yeah.
10:52Um, and then and I think this is more from, like, a product perspective, but then a lot of people, uh, in my you know, as a product manager, you wanna use AI for prototyping. Because, like, product managers, we all used to just write a bunch of documents Yeah.
11:04That none of our engineers want to actually reread. So so now with all these tools, like Lovable and, like, Replit and, you know, Codex and everything else, you can actually build prototypes of your products.
11:15So, like, I usually take a screenshot of my product and I'm like, hey. Can you make a prototype of this and make some changes and show it to my engineer and the designer? Mhmm.
11:22Uh, so that's then. And then you can use AI level four to build different apps. So, like, there's never been building a personal app for yourself and trying to, like, build an app that, like, millions of people use.
11:32Right? So when I say building apps is is a lot of it is just, like, personal apps for yourself. Does a where do the agents, uh, go in these layers?
11:39Well, if if you're just building an app, like, you know, you can just put it on Versa or something and then have a link that people can you you use. So it's it's not it's not really looping or anything. Okay.
11:48Yeah. So and that's layer four. Layer four.
11:50Yeah. And layer five? Layer five layer five is what we've been talking about, which is, like, you actually use AI kind of for, like, as a personal agent to try to, like, streamline and automate as much as you can.
12:01Right? And and, like and it it sounds, like, very complicated, but, really, it's just about, like, sitting down, like like, canceling all your meetings for a day and sitting down and, like, thinking through your past week, all the time you spent time manually doing stuff and and just, like, sharing it with your AI friend.
12:15Can you walk me through a process that we just talked about, uh, in a very practical way so that people who've been using AI on, like, level one, maybe some level two Yeah. Can move to maybe ideally level five, uh, where it's more I I I feel like the most practical step that most people should take is just to, like, stop using Chat, Chibutu, and Claw and just start using Codex or Claw code.
12:39That's, like, step step one. And, like, people get intimidated by this stuff, like Codex and Claw code. Like, this is, a programming Sounds very nerdy.
12:46Yeah. Yeah. But in reality, it's it's it's just, like, you you still just chatting with the AI.
12:49Right? And and, like, when I use Codex, like, 80% of the time, I'm I'm not building I'm I'm not coding or or doing any of that. Mhmm.
12:56It's it's just more powerful because, uh, you know, these two apps, they have you you can build skills for them.
13:02They have access to APIs and integrations so they can actually get work done across your other apps, and they're just, like, way more powerful than just default ChatGPG and Cloud. So so, basically, like, what I recommend you do is use Codex or Cloudco and set up a new folder or call it, like, personal OS or whatever you wanna call it.
13:17Just call it a folder. Right? And then I asked it, like, what are the main workflows that it can do here?
13:22And and you can kinda see here that Oh, you asked Codex to come up with Well, this is all the stuff that I've set up over time. Mhmm. Right?
13:29So, basically, like, know, as we talked about, it uses, like, various different skills, and skills are just text files to do work for me. So, like, for example, to to edit my newsletter, it it has this skill to shape a rough note into a transcript, and then it also uses other skill called last three days to add fresh research, and then it uses another skill called no AI slop Wow.
13:49To make the writing clear. I like that. What what if you don't ask it for last thirty days?
13:53It's just gonna Yeah. Omit? I think somewhere in the skill, have some instructions like, hey.
13:57Do some online searches to, you know, make sure I'm, like, you know, saying stuff that's actually true. Another thing, like, people should recognize is that there's a lot of, like, open source skills that are really good. You know?
14:06So last thirty days is, like, some open source skills that I found on GitHub, and it does, like, searches across, like, Reddit, x, and all these other platforms. Genius.
14:13I need this. And then it uses no ass slob to and and and and then eventually the newsletter post gets to a point, and, uh, I use social writer skill to make, uh, to turn newsletter into, like, a bunch of threads. Mhmm.
14:24Right? So that's for newsletter. And then for podcast, it's a very similar format.
14:27Like, I I use this podcast prep skill to research guests that I have. And then and then, you know, as part of making stuff on YouTube, you don't know if better anyway. You're gonna have a really good thumbnail title and copy.
14:37Mhmm. So I have a whole skill to just, like, test different thumbnails and titles. And then and then based on that, I decided the podcast is even worth doing, and then I I do, like, a postproduction skill, like, turning into all kinds of stuff.
14:48Mhmm. Yeah. Sponsors.
14:49We both have sponsors. It's, a lot of work. Yep.
14:51So I I I ask you to keep track of my sponsors through different skills and also help me draft and brainstorm sponsor copy, help me, you know, schedule stuff on the inventory and all that kind of stuff. Yeah.
15:02So so I I guess the TRDR is, like, uh, these workflows are done by connecting different skills together that that you build. And, like, for all these skills, I I don't actually Can you show one of them?
15:14Like, just go into, like, strategy, for example, where it's built and, like, walk them through how you built it. Yeah. Sure.
15:19Let let's just say open personal Can you show the sound? I think one of the most useful skills that I built is just like a personal adviser skill because I guess my wife is tired of me asking for advice all all the time.
15:30It literally is just a text file. Right? So it's a text file.
15:33Just give me an honest advice, and, you know, you're my trusted life and business adviser. And then keep your tone warm.
15:40Right? And then this this stuff is, like, a little bit more fan fancy.
15:44There's, like, evals and stuff, but we can talk about that later. LearningsMD. Is that your file?
15:49Yeah. Codex itself has memory, but I asked it to save an error text file called learnings dot m d to save memory from our past conversations so it kinda learns about me.
15:57This is kind of the self improving stuff that we talk about. With something they've released yesterday because they released something when it now remembers all of your conversations, like, has this expanded memory. Do people still need learnings MD?
16:08Well, I don't actually know how their, like, system wide memory works. So maybe you don't need it, but, like, it's also useful for me to read learnings.md and see what kind of, like, conversations I've I've had with it in in the past.
16:18For learnings. Md file, do you ask it how do you prompt so that every conversation contributes to that file? Yeah.
16:26So, um, you also have to think about managing the context window because you don't want it to be, like, super long. Yeah. Right?
16:31So I I have some instructions saying, like, hey. Like, after every conversation, you should ask me like, you should recognize kind of, like, patterns and, like, learnings that you have and ask me if if if if I wanna save in learnings. M d.
16:43So so ask me. And there are more strategic learnings. Right?
16:46Because this is your personal adviser. So ever after every single conversation? Uh, it should use its best judgment whether he wants to ask me or not, and then, uh, I I also ask it, like, if you're gonna save it there, just, like, have a few lines.
16:57Just, like, have one or two sentences. Like, don't don't don't write, like, huge paragraphs.
17:01So that's the way it's so interesting with these, like, how what what AI actually remembers about you. Yeah. Because sometimes you think maybe it remembers everything and you hope that it does, but in reality, it doesn't or remembers stuff that you don't really care about.
17:16Yeah. So these learnings and decent so you're basically prompted at saying,
17:21what where does this prompt go? Does it go It's it's just in the skill file, like, the the the main skill file of, like, hey. Be before you before you answer any questions I have, like, just quickly skim through learnings.md.
17:33And Yeah. Once we finish conversation or, like, in the middle of a conversation, we see something, ask me if this should be added to this learning ZMD file, and the skill runs across everything that you do. Because skill is triggered based on, like, uh, your description, your skill description in the in the top, which is like so here here I say, like, use whenever the user is stuck on a decision or working through a hard problem or asking for a gut check.
17:53But sometimes it doesn't trigger sometimes I have to, like, actually do, like, slash give me a slash personal adviser to Yeah. And then ask a question.
17:59How do you come up with all the like, for me, when I'm building this Yeah. There is so much brain work going into this, like, when I trigger this skill, all the nuances. Do you use codex for that as well?
18:10Yeah. I I I never write any of these files man manly. But you still no.
18:14It's even though writing, it's just thinking about all the nuances. Yeah. Yeah.
18:17Yes. I can think about a lot of things for you, but then there are so many
18:22different things that you want to remember, when to trigger the skill. Yeah. What types of information you wanna go there?
18:26Oh, maybe and then you wake up in the middle of the night, oh, I forgot to add this. That's right. Yeah.
18:31I have a skill to build skills. Right? And and and Anthropic actually has a skill to build skills too.
18:35Personally, I'm, like, very worried about just, like, it becoming slobify. Like, it just, like, starts adding stuff. I I don't reread it.
18:41Mhmm. And it becomes, like, super slob. Like, becomes, like, super Yeah.
18:43Because you also want to make sure you're in control of whatever is in the skill. And if AI just built a huge skill for you Yeah. Exactly.
18:50Then you don't have time to read it. So so my so my I have, a skill editor skill. It's very basic.
18:54It just, like, looks for, like, duplicate instructions or, like, slopped content and tries to remove it. Like Okay. Like, I I I tell you, should I keep my all my skills to, like, one page max.
19:02Mhmm. So I can actually read through it and see if it makes sense or not. Let's take a deeper look at what's what's next up.
19:08Yeah. So so, yeah, so, like, update the plan and and and then then and then, like so for for this personal adviser thing, I have my you know what?
19:17Why why don't I just ask it to you? Mhmm. Can you read my plan Google Doc and share its overall structure without confidential info?
19:31So so, basically, like and this is just the way I I I do it. Like, I have my actual plan or, like, know, my business plan or whatever in a in a Google Doc. Mhmm.
19:39And and, like, I I find this very useful to, as creator, to, like, have some principles to help him make decisions. So, basically, it's it's like a one pager with my goal, my principles, like, background of my business positioning and, uh, some other stuff, like energy and other stuff.
19:54Right? Like, I don't know about your goal, but my goal is just to, like, make x amount of money and just, like, actually be in control of my time and who I wanna meet with. And then my principles are, like like, I have a problem where I tend to take on too much stuff.
20:05So I asked you to, like, just keep the main thing the main main thing. Like, that's one of my principles, like which is, like, you know, people invite you to conferences and, like, know, do all kinds of like, try to write write write a book or do all kind of stuff and just and just you have to say no because you wanna you should focus on making a newsletter and a podcast great.
20:20Yeah. And it's so hard to Yeah. Focused.
20:23Right? Yeah. But but because I have that principle, so when it gives me advice, it has that principle in mind so it kinda kinda kinda check me.
20:28Mhmm. And then the business stuff is more just like, you know, like, all the financial information.
20:33And I I I think it's a very one important section is, like, what gives me energy and what takes energy away. And just, like, couple bullet points. Mhmm.
20:40But, like, you know, what gives me energy is, like, you know, vibe coding with codecs. Mhmm. What takes energy away is, like, I don't know, like, doing too many Zoom calls or or something.
20:49Totally. Yeah. Yeah.
20:50So so, like, so I I think I think it's very important. Like, you you don't have to be a creator to make this document. Right?
20:55For any kind of career you have, just make this document and putting a Google Doc somewhere and then set up a skill to reference this document when it gives you advice. And I think that that, like, is is literally one of my most useful skills.
21:07It's, like, not super complicated, but, like,
21:09it helps me make a lot of decisions for better or for worse. Most people use Claude like a search engine. They type in a question.
21:16They get an answer. Most times, they're not really satisfied with it, and they close the tab. I did the same thing for months, and I was looking at people who were saying AI is changing their life, and I'm like, mhmm.
21:26Then I spent one afternoon setting it up properly, uploaded a few files about how I think and how I work, and it completely changed. I wrote the whole process up step by step. You get it when you subscribe to my newsletter, Future Proof.
21:38It's free. The link is in the description. Yeah.
21:41Yeah. Well, this is very powerful because it then it you can use it for anything that you do, so it adheres to your principles. Yeah.
21:48Do do you have a team?
21:50Uh, I have, like, a video editor and a few contractors, but, no, I don't have a full time team. Yeah. Do you have you ever tried building, like, a a AI chief of staff?
21:58Uh, Yeah. I have one in this right now, I'm using this thing called Hermes. Uh-huh.
22:02Yeah. What can you talk about this? It's like OpenClaw Uh-huh.
22:05But it's, like, a little bit more reliable, and there's, like, a team that built it. Uh, but you can also just set up in Codex too. But, like Mhmm.
22:10I I I'm using Hermes right now because
22:12I I like having it in my messaging apps, like, in Telegram and, like, wherever I wanna use it. How do you use it? Talk to me about that.
22:19Because I'm trying to build something. You are. Yeah.
22:21I'm I'm trying to build, uh, I'm trying to build an AI chief of staff Yeah. Where first of all, like, my strategic decision making happens on Zoom calls and in Telegram where I give feedback to my team, and I say yes or no.
22:36Yeah. Uh, so I want AI to learn my strategic decision making Got style. Got it.
22:42And then I just talked to, you know, her interview, um, the founder of Kalshi. And she has a team of developers, but they build this thing where it goes across all the Slack chats they have every Sunday Yeah. And tells her where the team is stuck, what was promised to be delivered, what hasn't been delivered.
22:58So somebody who just watches the whole operation and
23:01keeps everyone accountable. I see. Well Is that where you're trying to build with Hermes?
23:05So you just have to think about where are the inputs, what are the what are you trying to do, and what are the outputs. Right? So the inputs for, like, any kind of business is probably like like Slack.
23:13Yeah. And if you use Granolah to record in meetings, there's transcripts there. Mhmm.
23:17Just set up all the just ask, like, Codex or Hermes or whatever. Just set up all the integrations. Be like, hey.
23:22Can you hook up to my Granolah? Can you hook up to my Slack? So for anyone who's at level two, let's say, and wants to get to level five, what are the steps they should be taking?
23:30I think this is the step number one. Really build something like a strategic mindset Yeah. Behind all of your AIs and projects and agents Yeah.
23:38What are the next steps? So step number one is, as I said, like, to just actually use Codex or Clockwork code. Just just start using that instead of, like, the the regular stuff.
23:46Right? That's Even projects? Like, you you Yeah.
23:49I I I switch away from projects because, like, I I set up all my projects here instead because, like, um, like, with a project, if I wanna edit a newsletter post, it spits out an output, and I have to copy and paste the output to A Google Doc. Yeah.
24:01Or, like, manually up update stuff. Yeah. But this stuff can just update Google Doc for me, so it saves the step.
24:07Mhmm. Right? Makes sense.
24:09So, now, sometimes you use the right apps, and step two is to just ask the right questions, like bring down your workflows and ask the right questions because I I feel like a lot of people don't realize what this stuff is capable of doing. You know?
24:21Like, this stuff is basically capable of doing, in my opinion, any kind of knowledge work that you have as long as you're patient with it and give it the right context. Yeah. So just ask the right questions and then set up the workflows like we just talked about and then build the skills and the integrations, and and then you're off to the races.
24:36What do you think your life is gonna look like in a year? A year? I've always worked in big companies where there's always been a lot of cross functional alignment.
24:44Right? You gotta align different teams and different orgs. I really actually do really miss my engineering team and people, like, that I enjoy working with at work.
24:51But I feel like these days, like, this is, like, the era of solopreneurs, and I feel like you can get so much done with agents, stuff that you don't wanna do, that you can it it really frees up your time to do stuff that gives you more energy.
25:05Mhmm. You know? And even the stuff that's, like, strategic or gives me energy, I I feel like now I have this partner that knows me Yeah.
25:12And can, like, just, like, kinda be my pair pair programmer or, like, partner to kinda figure stuff out. Do you think you're gonna be working less or more in a year? Uh, well, I I think I'm pretty type A, so I don't know if I'm a complete checkout.
25:23But I hope I'll be doing more work that I enjoy. And I don't think I I feel like if you and I retire, will be a bore of our mind.
25:30We'll play go play golf or something, but it'll be very boring. Of course. Yeah.
25:33I think we're gonna keep working. Yeah. We'll keep working, but we wanna do work that we actually enjoy, not like the montage copy and paste or, like, you know, that kind of work.
25:40Do you have any fear around AI? Uh, I feel like I I I fear that I'm getting dumber and and lazier.
25:47Like, if I'm on a flight and there's no, like, Internet connection, like, that I I just don't feel like working anymore. I don't I don't have a partner. You know, that's, yeah, evolution.
25:56If you take our grandparents Yeah.
25:59Even my grandparents, like, they would never understand my lifestyle because they're they were constantly working. Yeah.
26:05Yeah. Uh, it's either at home, like, my mom raised me while working, while cooking, while cleaning, and I'm like, oh, I can't clean. I have I need a cleaner.
26:12You need a cleaner. You know? Oh, I just bought a robot
26:15vacuum cleaner. I don't know how I lived without it. Yeah.
26:18We're becoming laser and laser. That's true. Yeah.
26:20Uh, but I think it's it's actually then we have more time for kids and meaningful time with, uh, with our loved ones. I I I do worry about our our kids because, like, at least you and I grew up without AI, so we have, like, twenty years of, like, doing the basic stuff, like basic critical thinking skills and, like, you know, learning how to do math and stuff.
26:38Mhmm. But I feel like our kids these days, they can they can start using codex, right, right now. They can talk to their, you know, computer.
26:44Yeah. So, like, uh, you kinda kinda have to make sure they actually still learn the fundamentals. What's what's your solution to that?
26:49What are you trying to do with your kids that's different from how yours? I have kind of like a radical thing. I I I feel like they need to and maybe this is my bias, but I feel like they need to just, like, maybe start a business or something and learn the real things.
27:02Because I I I feel like reflecting on my journey at yeah. My education, I didn't learn a ton from the classroom.
27:07I I learned by, like, failing and, like, making mistakes. Yeah. So I feel like they need to have the right mentality to do that kind of stuff and, like, learn how to fail and just go out there and, like, try to do something.
27:17Do do you ever have fear for your career? Yeah. My career?
27:20Yeah. I I'm I'm talking to a lot of creators these days Yeah. And I see them at events, and a lot of people think that creator economy is getting oversaturated
27:29because everyone is suddenly a creator because it's so much easier to create content. Yeah. Um, do you think that might be a problem?
27:36I I do think it's a problem because I I see a lot of creators, like, just making a slop. Like, they did, like they get AI to generate, like, 10 posts per hour, and, like, the slop goes viral, and it's kind of, like, disheartening to see that.
27:46Mhmm. But I feel like that I think that's why it's so important to find your principles and values because, like, I I never wanna be someone who just makes tons of slop even if it goes viral and makes me a lot of money.
27:56Yeah. You know? But but I do think but may maybe people will get tired of a slop, and people will be like, hey.
28:01I actually want human craft and human touch
28:03on this stuff. Let's finish with an action plan Yeah. For someone who's trying to stay ahead in this AI era.
28:12For the next week, what should they be doing? Okay. I think
28:16to stay ahead, get the $20 chat GPT plan, download Codecs, and just start saving time and you building workflows and building skills to save time and asking Codecs to do it.
28:27Because I feel like that's more productive for you because you will immediately recognize that you can start saving time versus what a lot of people do, which is, like, they go on Twitter and try to, like, read all the news and, like, try to, you know, take some courses or or or something.
28:40Like, switch from consumption mode to actual build mode. Output. Yeah.
28:44What if they come back and say, oh, it tried, but it didn't work. I don't know what you're talking about. It's so bad.
28:50It's so much easier to just do everything by yourself. By yourself. Well, it does take longer.
28:55So it's interesting because, like, with AI, you have to settle the system first before you can see the fruits of the output, and it does take time to settle the system. But it probably only takes, like, one day to do it. Right?
29:04And you have to have some patience because even now with all this stuff I built, it still cannot one shot exactly what I want. Mhmm. So there's always kind of a back and forth, and as long as you're patient and you give it feedback and tell how to save it in your skills and memory, it'll get better over over time.
29:18Yeah. You know? And unlike like an employee, it's never gonna leave you.
29:22It's gonna be there. It's never gonna leave you, and it's only getting exponentially better. Yeah.
29:27The mom will get better. Single month. Yeah.
29:29Peter, thank you so much. It was a very, very productive conversation. Thanks, Marina.
29:32Thank you for coming.
The Hook
The bait, then the rug-pull.
Peter Yang spent a decade building products inside Reddit, Meta, and Roblox. Then he left, went solo, and rebuilt every workflow around AI. His opening line: 'It almost feels like cheating.' What follows is thirty minutes of the actual system — not theory, not hype, but screen-shared skill files, sniffed internal APIs, and a personal adviser that reads his Google Doc strategy before answering any question.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
