20:47Matt Wolfe · Review
Claude Fable 5.2 Just Dropped. It's the World's Best AI.
A hands-on first look at Anthropic's new Fable 5.2: what the benchmarks say, what it costs, and what it built for Matt Wolfe and early-access users.
November 9thThe argument in one line.
A personal AI system built from portable skills, a one-page plan you write yourself, and a habit of improving skills across many conversations will outlast any single model or tool.
Read if. Skip if.
READ IF YOU ARE…
- A solo creator or founder who keeps re-explaining the same tasks to AI and wants saved skills that work in any tool.
- Someone drowning in email, Slack and calendar who wants a twice-a-day triage list with draft replies ready to approve.
- A creator who already has a few Claude or ChatGPT skills and wants to stop them from filling up with one-off rules.
- A podcaster or YouTuber who wants AI help with guest research, interview guides and short-form edits without handing over final judgment.
SKIP IF…
- You want a full coding walkthrough. This is workflow design shown with Claude, ChatGPT and GrokBot, not a build tutorial.
- You want a single best-model verdict. The speakers argue the harness matters more than the model.
The full version, fast.
Peter Yang runs a one-person business on a personal AI system made of reusable skills that work across ChatGPT, Claude Code and GrokBot. A chief of staff triages email, calendar and Slack twice a day, an advisor checks decisions against a one-page plan he writes himself, and scheduled bots email him business stats, fitness summaries and weekend ideas. His skill editor reviews many past conversations at once and proposes fixes, so skills improve without overfitting to one chat. The advice: pick one recurring task, give it good context, and keep a human pass on anything customer-facing.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:01
01 · Cold open
Matt introduces Peter Yang and the promise: one cohesive system, with simple versions built along the way.
01:01 – 04:37
02 · The system and shared skills
Why Peter builds portable skills that work across ChatGPT, Claude Code and GrokBot, and a tour of the whole map.
04:38 – 06:43
03 · Harness beats model
Which tool Peter uses for what, and why one tool is enough for most people.
06:43 – 08:09
04 · The chief of staff
A twice-daily list of what needs attention, with draft replies Peter approves by number.
08:10 – 10:41
05 · The advisor and the one-page plan
A human-written plan doc gives the advisor something to measure against. Matt builds a version and tests it.
10:42 – 13:01
06 · Building a chief of staff in Claude
Matt sets up a scheduled task with email, calendar and Slack connectors and a numbered-list prompt.
13:01 – 16:14
07 · The skill editor
Improve skills by reviewing many conversations together, not by patching after each one.
16:14 – 18:00
08 · Visual feedback app for video edits
A custom timestamped-notes interface replaces a human editor's review loop for short-form video.
18:00 – 20:33
09 · Podcast prep
Guest research, angle testing and an interview guide that can be shared with the guest.
20:34 – 24:19
10 · Scheduled single-job bots
Business review, weekend ideas and fitness check-ins, plus Matt's bass coach project.
24:20 – 27:42
11 · X trends, human review and wrap-up
Finding topics hot on X but thin on YouTube, keeping a human pass on changes, and the one-workflow advice.
Lines worth screenshotting.
- The harness matters more than the model: ChatGPT wins at browser use, Claude at writing and video editing, GrokBot at scheduled agent work on a virtual machine.
- Skills saved as plain instructions work in any tool, so you stop rebuilding your workflow every time a new model ships.
- An advisor works better from a single one-page plan than from unlimited access to your memory and documents, which is mostly noise.
- Write your goals doc yourself and tell the AI it can suggest changes but never make them. Peter calls this his slop-free zone.
- Run the daily triage bot every day, but run the direction-checking advisor only once a month. More often is too much for long-term decisions.
- A chief of staff that drafts replies and links the threads turns email from writing into approving.
- Updating a skill after every conversation overfits it to that chat. Review three weeks of threads together and fix only the repeated patterns.
- A skill editor is a skill whose only job is maintaining your other skills and keeping them simple.
- Skills never get perfect in one shot. Long prompts should grow out of feedback loops, not be hand-written up front.
- Short-form videos of a minute or so can be nearly one-shotted once the editing skill has been iterated enough. Long-form still goes to a human editor.
- A small custom feedback app with timestamped notes replaces a human editor's review loop, and took about ten minutes to build.
- Give each bot one job and define its purpose, what it needs, what good looks like, and when and where it reports.
- Scheduled results are only useful if they land where you actually look, so Peter has his bots send emails instead of burying results in chat threads.
- AI is still weak at titling episodes, and titles decide whether anyone clicks, so human producers keep that job.
- Ask the AI for the exact before and after text of every change in a numbered list, so you can approve item by item.
- Spotting topics that are hot on X but thin on YouTube is a repeatable way to find video ideas.
Build one skill, then improve it in batches.
WHAT TO LEARN
A durable AI setup comes from portable skills, a plan you write yourself, and improving skills from patterns across many conversations.
- Skills are saved instructions that work in any tool, so your workflow survives every new model release instead of being rebuilt each time.
- The harness around a model, with its browser, computer use and scheduling, often matters more than which model sits inside it.
- Start with one recurring task that would actually save you time, such as a twice-daily list of what needs your attention.
- Have the AI draft replies and link the source threads so email becomes approving numbered items instead of writing from scratch.
- Give an advisor skill a single one-page plan you wrote yourself. It is a better measuring stick than unlimited access to memory and documents.
- Tell the AI it can suggest changes to your plan doc but never make them, so your goals stay in your own words.
- Run daily triage daily and direction-checking monthly, since long-term decisions do not need constant review.
- Do not patch a skill after every chat. Review several weeks of conversations together and fix only the feedback you keep repeating.
- Prompts and skills grow through iteration. You rarely get a good long prompt by hand-writing it up front.
- Give each bot one job, defined by its purpose, inputs, what good looks like, and when and where it reports.
- Have scheduled bots send results to a place you actually read, such as email, so the output does not get lost in chat threads.
- Keep a human pass on titles and anything customer-facing, and ask for exact before and after changes in a numbered list.
Terms worth knowing.
- Skill
- A saved set of instructions for one job, so you don't re-explain the task each time. The same skill can be used from different AI tools.
- Harness
- The app or environment wrapped around a model, such as Claude Code, ChatGPT or GrokBot, which decides what tools, browser and scheduling the model can use.
- Chief of staff
- A scheduled AI task that reads your email, calendar and Slack and returns a prioritized, numbered list of what needs attention, with suggested next steps.
- Advisor skill
- A skill that checks a decision against a one-page plan of your goals, limits and energy drains, and gives blunt advice.
- Skill editor
- A skill whose job is reviewing many past conversations and proposing improvements to your other skills while keeping them simple.
- Council
- A set of different AI perspectives that you run your biggest decisions past.
- Overfitting
- Changing a skill to fix one conversation's mistake so that it carries a pile of one-off rules that do not help in general.
Things they pointed at.
02:30productChatGPT
02:30productClaude Code
02:30productGrokBot
27:20channelPeter Yang's YouTube channel
Lines you could clip.
02:08
“I actually think the harness matters more than the model.”
08:35
“I have a Google Doc that has basically like a single page with my overall plan.”
14:40
“If you do that after every single conversation, you end up with a skill full of random one-off rules.”
27:00
“Pick one workflow that would actually save you time, give it good context, and keep improving it as you go.”
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.
So I'm a total sucker for a good workflow. You know, things like automations and command centers and chief of staff setups like i love this stuff now peter yang he spent over 10 years as a product manager in the bay area most recently over at roblox and earlier this year he actually quit his job to make content full -time the way he runs what's basically a one -person business is this whole ai system that he's built that kind of runs like everything for him twice a day it reads his email calendar and slack and then tells him what he needs to go deal with.
He's got this advisor that calls him out when he drifts from his goals. It preps his podcast. It helps edit his videos and it even finds stuff for him to do with his kids on the weekends.
Now, if you've watched my past videos, I have built a lot of these little pieces myself, but Peter's figured out how to tie it all together into like one cohesive system. So he's going to walk me through how all of this works. And along the way, I'm going to jump in and build like simple versions of this stuff myself so that you can follow along and set it up.
too this is show me your workflow peter yang edition i spent like over 10 years as a product manager in the bay area right and most recently at roblox and i left my job in june to i guess become a creator full -time I just really like shitposting. No, because I feel like at a traditional job, you have a lot of meetings and there's a lot of things going on.
And I feel like to be on the cutting edge of AI, I need to actually spend time tinkering and trying out the latest tools and doing stuff myself. And it's become kind of like a full -time job. That's kind of why I quit.
And also financially, I was making enough from sponsorships and other things to support myself. No, I totally feel that. Trying to keep up with it all.
and test the different things and know what's going on. Feels like a full -time job constantly. I think as content creators, I don't know how big your team is, but I basically am by myself.
and i have a bunch of contractors for video editing and thumbnails and a lot of the work that we do as creators or pretty much any job can be streamlined a lot with ai the problem with ai of course is like every other week just like a new model and a new tool coming out and you know you and i is part of our job to cover the stuff but for like a regular person if you're constantly bouncing between models and tools it's going to be like you're not actually building anything yourself that's durable right so that's why like when i talk to people i highly recommend that they set up their personal ai system with like a bunch of skills and uh automations that is unique to them that actually helps them save time and save money that they can bring to any tool that they want so for example right now i'm using a mix of chat gpt clock code and grokbot and all three have access to my skills and work workflows so i can kind of like just do work on any of them here's kind of the system that i built and it looks pretty complicated but it's actually
hopefully not too complicated. Okay, let me break down what we're actually looking at here because there's a lot going on. So Peter uses three tools.
He uses ChatGPT, Cloud Code, and Grokbot. And all three of these can actually use the same set of skills. A skill is basically like a saved set of instructions for one job.
So he's not having to re -explain himself every single time he needs to do a task. And because these skills are sort of universal, all of these various harnesses like ChatGPT, Cloud Code, and Grokbot can actually pull the same. skills and it doesn't really matter which model you're using.
So at the top here, he's got three things that help him run his life and his business. A chief of staff that looks at everything coming in and tells him what to deal with today. An advisor that checks his decisions against his goals for the year and a council, which is a bunch of different AI perspectives.
He runs his biggest decisions past. Now, if you saw the Nate Herc episode I did a couple of weeks back, it's kind of a similar idea with the council there. And then he's got his.
content systems here. He's got skills for researching podcast guests, building an interview guide he can actually send the guest, titles and thumbnails, video editing, intros, and turning one recording into a whole newsletter as well as social posts. Then there's his scheduled tasks, stuff that just shows up in his inbox, business stats, a fitness check -in, even ideas for what to do with his family on the weekend.
And then last, he's got his skill editor. It's a skill whose whole job is keeping his other skills clean as he keeps tweaking them.
We're going to come back to that one because it's actually one of the smartest parts of this whole thing here. Now, I know this looks like a lot, but you don't need to build all of this at once. Start with just one recurring task that would actually save you time.
How do you decide which model to use for each thing? Like, have you kind of figured out that like. the OpenAI models are best at this kind of stuff, the Cloud models are best at this kind of stuff, and the Grok models are best at this kind of stuff?
Or are you just constantly rotating them in and out based on new model releases and things like that? I am rotating it a lot, although I recommend most people don't do that. I actually think the hardness matters more than the model.
So from the model point of view, I do think Opus 5 .5 is kind of a game changer, especially for video editing and motion graphics. It's very good at that. But I think the hardness is more important.
For example, ChatGPT arguably... doesn't have the best model right now, but like the browser and computer use is so good so that whenever I have to like fill a form or like do some of the manual browser work, I always just use ChatGPT to do it. And GrokBot is, in my opinion, by far the best personal agent right now.
And because it's on like a virtual machine, pretty much all my scheduled tasks, I moved over to GrokBot because like if I have it on my computer, like clock code, like, you know, if my computer's off, then it doesn't work anymore. So like GrokBot, I have. all my scheduled tasks.
It also has like X integration to like do research, which is great. The part of the reason I'd use all three is also because like I kind of get free access to all three because it's my content career. But like if I was paying 200 bucks a month for this stuff, then I'll probably only pick one.
Quick side note on models because I do get asked this stuff. quite a bit. For me, Clod is where I do most of my writing.
I think that so far it's kind of been the best at matching my voice and helping me optimize how I word things. So when I do write scripts for videos or when I'm trying to think of title ideas for videos or things like that, Clod is usually my go -to for that stuff. Codecs and the ChatGPT models are the models that I still use mostly for my coding projects.
I'm also personally using dot a lot from open ai as like my main agent or assistant kind of thing but if i'm being honest i haven't really been using it as much as i thought i would when they first announced it pretty much everything in this video is going to work with whatever model you choose to use so let's go ahead and start with the first thing he showed me the chief of staff 7 a .m and 4 p .m it gives me like a list of things to take care of here's the example list send zoe and 4pm invite reply to mercury on like mercury's the sponsor so reply with a brief pick a date for the maven granola session and because it sends me this list and also links to the email threads instead of having to like go into your manually each one and try to like reply manually i basically just reply like this one go set up the invite with zoe two can you draft the email from the review three let me handle this manually and four can you do research right so for example can you draft the email from your review and and like i
almost never just let it send emails by itself i always have to review the email but the fact that it creates this draft means i can just change some stuff like change the date or something and then hit send and then it's like way faster for me to do this stuff than try to draft this stuff from scratch myself yeah and you don't even need to get into like gmail or anything at all like it'll just you can write the email edit the email and hit send all from within clot it looks like yeah like i i think basically like i almost feel like This clockhole interface or like chat GPT is like the new browser.
Like instead of using the browser, I can just use the sync to use the browser. So the browser almost becomes like a middle layer. that i don't really use anymore it's almost like the like you know even a higher level almost like the new operating system right because it's going and running other apps for you it's going and running other you know sending emails for you it will actually go into a browser and browse for you if you want it to exactly the other thing that i have is like i have my little advisor skill so here i'm telling you to give me some advice without reviewing any personal information because we're doing this like video one thing about the advisor skill is like the lms are really good at getting context But if you just tell it to look at your memory and look at all your documents, it's like too much crap.
It's too much crap. So what I do is I have a Google Doc that has basically like a single page with my overall plan. It's got what he wants this year, what gives him energy, what drains him.
And it's not just business. He also puts family stuff on there as well. That gives his advisor something to measure against.
So when an opportunity comes in, it can say, yeah, that fits or hey. You said you didn't want to do that this year. So let me show you a simple one page version that I created, and then I'll show you how I turn it into a reusable skill.
So it's a very simple doc. It's just a few sections with my top goals, what I'm not doing this year, what gives me energy, what drains me and the family stuff that's non -negotiable. Now, I could probably add more depth to this and I likely will come back and add more to it.
But this is my example with. actual realistic details around my own business. Now, one of the rules that Peter himself is really strict about is that I write this, or he writes this, not the AI.
He calls it his slop -free zone. In fact, I put this right at the top of the doc. I only edit this doc.
AI can suggest changes, never make them. I also added a section down here for how I want the advice, basically telling it. Don't sugarcoat things.
Tell me what I need to hear, not what I want to hear. I'll grab the URL here, copy it, jump into Claude. And assuming you've got your Google Drive connector and your Google Docs connector turned on, it should be able to read that.
So I'm just going to give it that link real quick. Feel free to pause and see my exact prompt here. It took.
maybe 30 seconds to do this, but we have a skill here. And because it also is connected to my YouTube data, in my case, it will also look at that to help it out. So let's go ahead and save this skill here.
Use my advisor skill. I've been asked to speak at a conference in March, two days of travel. Should I do it?
And here's the response. It took about a minute, but it's definitely pulling from the doc. So short answer, no, unless it replaces HumanX.
You already have HumanX in Vegas, block from March 7th through 10th. A second March trip goes over your one trip per month limit. Speaking also means building a talk and your plan list keynotes is something that trains you.
So it's very clear that it's not just pulling in generic. advice. But like I said, this is a starting point.
Build something out, see what it gets wrong, and then tighten it up. The advisor especially is like whenever I'm frustrated with something or like confused about something, I kind of check it for advice and be a sounding board. And it's kind of available 24 -7.
So I think it's like, I'm not sure how healthy it is, but it definitely gives me pretty good advice. So these two sound similar, but they do totally different jobs. The chief of staff is about like right now, what needs my attention today.
The advisor is more about direction. Like, am I actually headed where I said I want to go? So that's why they run on different schedules.
The chief of staff runs every single day. Peter only runs the advisor once a month. He says anything more than that for the long -term stuff is honestly just going to be too much.
Now using Peter's outline, I'll build a basic version that you can adapt. So come over here to scheduled inside of cloud, click plus, and I'm going to set it up manually. And then for our prompt, you're my chief of staff.
Look at my email for the last 24 hours, my calendar for today and tomorrow and my Slack. Give me a number list of what needs my attention. Most important first for each item.
one line on what it is, link to it, and your suggested next step. Reply, schedule, delegate, ignore, skip newsletters, promos, and anything that doesn't need me. When I reply with numbers, do what I say for each one, draft emails, but never send them without me.
Now we can select the model. I want it to be on auto where it will just run without asking for model. pick the model you want to be your default.
I'll set mine to Opus 5 .5. And let's say we want it to send every morning at 8 30 AM to set us up for the day. And let's just do it on weekdays.
So it doesn't send anything on weekends. And then under advanced settings here, we want require this computer off. So it will actually run in the cloud.
And then we want to go to connectors and make sure that the connectors we want it to have access to are turned on. You can also turn on notifications. So it sends you a push notification, email, Slack message, et cetera.
But I'll just leave it on push notifications for now. Now, when I create this, my chief of staff is going to run every single morning.
Another thing we can add to the prompt if we want is let's go ahead and edit this that Peter suggested. Review our chats, my connected sources and your memory and tell me something special or unique you notice about me that you think I haven't realized about myself yet. Be honest.
You don't have to be positive. You can have it give you something like that every morning along with everything else it delivers. And we can see it give us a prioritized list.
Now I'm going to have to blur quite. a bit of this because it is actually pulling from my email and Slack and everything else about different deals and things I'm working on, but it did give us a numbered list of everything I need to deal with today, as well as prioritizing most important towards the top.
And it gives me action items that I can do. Down at the bottom, it gave us our something I noticed about you response, which is also a bit too personal and I'll have to blur out. One thing that people probably over -index on is how in -depth the initial prompt needs to be, because it's something that I feel like you tend to build out over time, right?
Like I find this a lot with skills specifically. I'll have it go and build me a skill to do one thing and then you know i'll notice some flaws in its response i'll tell it to do this differently next time and then update your skill so all future responses have this new information in it and you kind of like iteratively improve that skill or over time yeah exactly you can improve it over time and you know like this prompt like um this longer prompt here i did not hand write any of the stuff I just used my skill editor skill and got the AI to make updates.
I actually have a really good tip. I have all these podcast threads here where I try to get AI to draft a newsletter post or make updates. And I'll tell you, dude, I've been improving these skills for a long time.
It never gets perfect in one shot. I always have to go back and forth with it, give it feedback, and make it better. And I used to do what you just said, which is after every loop, I'm like, hey, based on this conversation, go update my podcast skills and do better next time right right so so i i think that's definitely better than not doing that but like the problem with that is like it tends to like overfit into whatever your current conversation is so what i recommend doing instead is like you just make like a new thread and be like hey can you look at all my podcast threads and you know when i've been using all the various skills and just look through everything let's say over the past three weeks and then suggest ways to improve the skills.
And then I have to use my little skill editor skill to make sure that it keeps things simple. These things have perfect memory of everything that you talked about. So it's going to look through, hopefully if it does it well, it's going to look through all these threads and then find some common patterns so that it doesn't overfit to one thing.
So what's a skill editor? It's a skill whose job is maintaining your other skills. And here's what you need one.
Every time the AI gets something wrong, it's really tempting to go update the skill so that you don't do that again. personally do that all the time. But Peter's point is if you do that after every single conversation, you end up with a skill full of like random one -off rules.
So his tip is to look at a bunch of recent conversations together. Then the AI could spot what you keep correcting over and over instead of overreacting to one random mistake. So for example, here's my short form video creator.
Let's follow Peter's suggestion and say, look at all the scripting and editing feedback I've given across all of the recent shorts we've done. Find feedback or items that I tend to repeat often that could be added to our skills. And here's the response.
I went through every round on your computer, 13 rounds across nine edits. So the B -roll doesn't show what you're saying or shows it at the wrong moment. Framing and zoom.
White text that repeats the subtitles or covers the B -roll. Sound effects and number animations out of sync. And now we can go through and look at these various numbers and say, for the editing skill, include points one, two, four, and six submit that and it's gonna go through and make the skill better based on the feedback that i constantly give it and you can see this whole loop in action in peter's recent video where he had ai call comcast for him he went back and forth with it on the music the motion the graphics and the edits and it even blurred out his address and phone number all on its own but that's the thing it still took his creative judgment and a bunch of back and forth to get there so like video editing isn't solved like at least not yet one thing i kind of find annoying is like give me a feedback across different time stamps i'll be like hey three minute in change this five minute change that so i basically just kind of build my own little skill to do this um and like i don't actually know how to use premiere it's like too complicated for me so i've been my i built my own little thing and hopefully it runs soon it kind of just shows me a layout
Where I can give feedback to the LM in like a more visual interface. That's smart. Yeah, I like that.
Yeah, I've found that when it comes to video editing, it seems to work really good for me for like short form videos. So if it's like a minute, a minute and a half, I can actually almost one shot the videos now just by iterating on the skill over and over again. I've got a skill dial then that's pretty good at just like one shotting those shorts.
But long forms, I still edit. Well, actually hire an editor who edits them for me manually now. Yeah, long from my podcast, I have a manual editor.
And then for my talking head, I'm using AI more. But yeah, this is the interface where I can play the video and I can jump to certain neurons and I can leave feedback of like, you know, only show three things in the list. Okay, yeah.
So this is almost like a frame .io sort of interface. That's like, that's a... tool that does this you put a video in it and then you can add notes at various time stamps to tell your editor like where you need fixes and it's just like this basically just built like rebuilt that for you yeah this is yeah and it didn't take that long to build it until like you know maybe 10 minutes to rebuild this thing and instead of like a human editor i'm leaving feedback for the agent right so after after i leave a bunch of notes you know let's say fix this uh then i can send notes And then if it works, it will basically, yeah, the agent's now working.
Oh, I love that. I love that. That is such a smart way to do it.
And I'm assuming that the podcast workflow is very, very similar to the video workflow. It's very similar. The podcast prep, I'm doing an episode with Matt Wolfe to share our AI workflows.
And this is a skill that I use. And basically, it does a bunch of research on you and workflows. And then it found your top insights from somewhere.
It starts each day in a dashboard he built himself. The Wolfe control tower. Okay, this sounds interesting.
And I always like to... have it run thumbnail title, which is my other skill to kind of try to figure out the angle. Because, you know, like with YouTube, the thumbnail, like if people don't click into the podcast, no one's going to watch it, right?
So you got to find the right angle. So lately I actually got it to like, basically like actually copy the thumbnail screenshots of things that are doing well. So here's some.
different angles. And based on these research examples, it suggests some combination. So like, you know, copy art, daily AI workflows, the AI workflows to YouTubers use every day.
Yeah. I don't know. I guess it's a good start.
And then it creates the draft, the draft interview guide. Yeah. Okay.
Let's start with mornings. Can you show us the wolf control tower and walk through what you're checking in that first hour? So Chris's interview guide, and obviously it's not perfect.
And then usually I tend to make a linear ticket and then I kind of do a manual pass, but eventually we get to a point where. We have an interview guide that I can basically share with you. Like if you're the guest, so you can prep for it and then we can do the podcast.
Now to make this work, you really only need to give it four things. You need to tell it who the guest is. You need to tell it what you're talking about, who the audience is that you're trying to get in front of and the actual format of the interview.
And then down here, I'm asking it for some sources, some possible angles and a handful of questions that will actually build on each other. I like these to feel like conversations and not like questionnaires or, you know, straight up podcast interviews. back with this whole report here that we can use to prep for our conversation when we come down to the episode angles look at this the ai chief of staff that runs a one million one person business which you know is pretty much what we leaned into in this version that you're seeing and then down here we have a guest guide for peter so that we can prep him for what he's about to come into now one thing i still personally don't hand off to ai is titling these episodes.
For whatever reason, I still feel like AI is pretty bad at titling. And when it comes to YouTube content, that's like one of the most important parts. So yeah, I still work with my producers and editors to come up with titles.
Now, Peter also has a bunch of stuff running on schedule. So I went ahead and asked him about that next. So what about the scheduled stuff down at the bottom?
Is that kind of the same stuff that your chief of staff does? Is that the same scheduled stuff or is that different? It's a little bit different.
Let me show you an example. So basically, I have these scheduled things and I use too many AI apps and threads. So for me to actually read this stuff, I have to get it to send me emails because I actually read my emails.
So here's some examples of stuff that I get. This is my weekly business review thing. So basically, it pulls all these stats and it sends me a summary of how things are doing.
This course I launched is doing pretty well for the newsletter. The funnel is doing well, blah, blah, blah. And then it researches a bunch of other channels.
So that's my business email. And then this is my weekend email. Because in the Bay Area, it's kind of boring.
So I get to search the local mom blogs and stuff to help me find interesting stuff to do with my kids. And then my little training thing where I actually vibe coded my own fitness app. And then it has an MCP that takes my lifting stats.
And then it sends me like a summary of how things are going. One thing Peter does that I really like, instead of one giant AI that does everything, he gives separate bots specific jobs. And this works because each one has a really clear job.
For every bot, you want to define three things. What its purpose is, what it needs to do the job, and what good looks like. If it's recurring, add when it runs and where it sends the results.
For this example, I'm going to work with something that I do. often, and that's practice my guitar and my bass. I'm actually trying to get a lot better at bass.
I've played guitar my whole life. So let's make a learn to play bass. weekly check -in.
For this one, I'm going to create a new project. So I'll click up here on projects, bass coach. Your goal is to help me learn how to play the bass, help me keep up with a practice schedule and ensure that I'm increasing in difficulty and not just playing the same stuff over and over again.
Now under our instructions, here's what my custom instructions look like. Feel free to pause on the screen, but I'm basically saying I want to learn the bass and I want it to get progressively harder and harder each week. And you can look at my practice log to help me keep up with that.
So here's a practice log of songs I've been practicing on the bass, along with how long I've practiced, my notes, my rating for how well I think I played them. And I can grab this spreadsheet here. Now I can come over here to context, click add, and then.
come over to drive and then up here, it says search resources or paste URL. I can paste the URL to that Google drive link. And if I click on context, you can see our base practice log is going to get synced via the MCP.
So it'll be able to continually watch that. Give me this week's check -in and here's what it tells me going song by song on the songs. I'm learning things that are against my goal, things that are, you know, working towards my goals and one change for next week.
Now, if I come back to this project, I can now tell it to give me my weekly check -in and have it send weekly on Sunday nights at 6 .30 PM. Maybe you don't want to use a spreadsheet or you don't have a spreadsheet.
Well, that's where creating a quick and easy app comes in. I can use this same project that already has all of the context we've given it and just tell it, build me a simple practice logger. I pick a song, enter dates, how well I think I played it, any additional notes, and it saves my history.
And just a couple of minutes later, we have a simple app here. Let's say I've been working on Hysteria and I practiced today for 20 minutes. I can rate myself on how well I did, add some notes, and it will update that session.
I really like RockBot because you can basically... build a bot for each job that you have right so like i have a i have a producer bot that does research for him for me and i have another bot here that um does whatever errands my wife wants me to run like i'm i'm like a terrible chef but now that i'm kind of like unemployed making youtube videos i kind of had to cook.
So I made like a little app for myself to cook stuff in the air fryer and in the pan. For Peter's producer bot, one of the most useful things it sends him is a list of what's getting attention on X before it takes off on YouTube, which I think is pretty cool. For video ideas, usually this AI stuff breaks on X first and then makes it over to YouTube and LinkedIn and these other platforms, right?
And then Facebook two months later. Yeah, exactly. So I have it research X data on like breaking topics and compare with YouTube.
So like, this is not perfect, but I think this is pretty useful. A lot of times it makes this like report. And this is like how many X posts over the past seven days.
And also how many channels have made videos on this topic. Theoretically you can find gaps where it's like really trendy on X, but people haven't really made a ton of videos on it. And then, and then you can like maybe make a video on it and like, you know, finding a niche for you for yourself yeah i like that it's funny the in my very very early days of youtube back in like 2022 20 early 2023 like that was my core strategy was find what was trending on x or find a cool workflow that somebody showed off or find a really cool video that was going viral on x and make a youtube video about it because it wasn't most people weren't finding it on youtube yet so it was almost like x was like the the sort of signal and i would go and find like what the signal was alerting me of.
And then I would go and make a YouTube video about that. Obviously I would credit the X creator and I would link back to them and all that kind of stuff. And that was like a very good formula for success on YouTube.
At least a couple of years ago, it was. Probably is like, I mean, you have to use your judgment too, but like it probably still works, right? Because a lot of the AI chatter happens on X first.
Yeah, absolutely. Absolutely. My one last note is like, you know, it's very easy to just get the LMs to make updates to your skills or like make updates to your docs without looking at it.
and and like i i still think uh human judgment taste really matters i may for example i may get the ai to like check my newsletter post for spelling and grammar mistakes or like do some cleanup right but i always say like hey give me the exact before and after copy that you changed in a number list you know give it to me in the chat so then i can say like okay one two eight makes sense but nine doesn't make any sense don't do nine so i can at least do a human pass like i get my human eyes on it that's with every skill you put out you kind of do an extra audit on the skill and anything that's like user facing or customer facing i always like to have my human eyes on it first yeah gotcha that's the skill the skill too if it wants to make a ton of updates to my skills then i'm like okay give me a number list exactly before and after what copy you're gonna change in my scale.
So huge thanks to Peter for actually opening up his system and showing me like the real stuff and not some polished demo. If you take just one thing away from this one, don't try to build all of this. pick one workflow that would actually save you time, give it good context, like Peter's one page plan, then just use it and keep improving it as you go.
Now, if you want more from Peter, go check out his YouTube channel. I'll link it up in the description and head over to behind the craft at behind the craft .com. That's where his membership is.
And he's got all of the various skills he's built inside of there. And that's what I got for you today. If you like this one, give it a like and a subscribe.
So I know you're digging these, show me your workflow episodes and you know. That'll make me motivated to continue making more of these. Show me your workflow episodes.
Thanks so much for hanging out and nerding out with me and Peter today. Hopefully I'll see you in the next one.
The Hook
The bait, then the rug-pull.
Peter Yang quit a product job at Roblox to make content full-time, and a personal AI system now does much of the running of his one-person business. Matt Wolfe asks him to open it up and builds simple versions of each piece on camera.
Named ideas worth stealing.
03:18list
Three jobs of a personal AI system
- Chief of staff
- Advisor
- Council
The top layer helps run life and business: what to deal with today, whether decisions fit the yearly goals, and a set of AI perspectives for the biggest calls.
Steal forany solo operator setting up a daily AI routine
21:31list
Define a bot in three parts
- Purpose
- What it needs to do the job
- What good looks like
- When it runs and where it sends results (if recurring)
Give each bot a single clear job instead of one giant AI that does everything.
Steal forany scheduled task or project
13:55model
Review across threads, not after each one
Ask the AI to read several weeks of conversations, find feedback you repeat, and propose skill edits. This avoids overfitting a skill to a single chat.
Steal forimproving any saved skill or prompt
How they asked for the click.
VERBAL ASK
27:04newsletter
“Go check out his YouTube channel and head over to behind the craft dot com. That's where his membership is.”
Soft and specific: points to the guest's channel and membership, then a short ask to like and subscribe to the series.
MENTIONED ON CAMERA
Visual structure at a glance.
