Quick answer
A personal AI system is a small set of reusable instructions (“skills”) and scheduled jobs that you own, which any capable AI tool can run for you. In a new episode from Matt Wolfe, creator Peter Yang shows the one he uses to run his one-person business: an AI chief of staff that reads his email, calendar and Slack twice a day, an advisor that checks decisions against a one-page plan he writes himself, content skills for podcasts and video, and scheduled bots for business stats, fitness and family weekends. The lesson for the rest of us is simple: start with one recurring task, give the AI good context, keep a human check on anything that goes out, and improve it over time.
Key takeaways
- Own the skills, not the tool. Peter keeps his skills in one place so ChatGPT, Claude Code and Grok Bot can all use them.
- The harness matters more than the model, in Peter’s view: pick a tool for what it does well, such as browsing or running when your computer is off.
- Chief of staff = today; advisor = direction. One runs daily, the other about once a month.
- Write the plan yourself. Peter’s one-page plan is a “slop-free zone” the AI may comment on but never edit.
- Improve skills in batches, by reviewing weeks of chats at once, rather than after every single mistake.
- Drafts, not sends. Peter reviews every email and every customer-facing change before it goes out.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Write a One-Page Plan Your AI Advisor Can Check Decisions Against — the simple doc that turns a chatbot into an advisor who knows what you actually want.
- How to Improve an AI Skill Without Filling It With One-Off Rules — the skill editor method: review weeks of feedback at once, then approve changes one by one.
In this guide
- Watch: the source video
- What is a personal AI system?
- What does Peter Yang’s personal AI system include?
- Which AI tools does he use, and does the model matter?
- How does the AI chief of staff work?
- What is the AI advisor, and why is it different?
- How do you keep AI skills improving without making a mess?
- How does he use AI for video and podcast work?
- What jobs do his scheduled bots do?
- Where does human judgement still matter?
- How to start your own personal AI system this week
- Frequently asked questions
- The bottom line
- Related posts
Most people use AI one chat at a time. You ask a question, get an answer and start again from scratch tomorrow. Peter Yang does something different, and it is worth understanding before you buy yet another tool.
Peter spent more than ten years as a product manager in the Bay Area, most recently at Roblox, before leaving in June to create content full time. He works alone with a few contractors. In this episode of Matt Wolfe’s “Show Me Your Workflow” series, he opens up the personal AI system that handles much of his admin, research and content prep. Matt then rebuilds simple versions so viewers can copy them. Below, we explain what Peter showed, separate it from the bigger claims, and turn it into steps you can take this week.
Watch: the source video
What this video covers: Peter Yang walks Matt Wolfe through his AI chief of staff, his goal-checking advisor, his skill editor, his podcast and video workflows and his scheduled bots, while Matt builds simple versions in Claude so you can follow along.
Video: Matt Wolfe (@mreflow), “He Built an AI System That Runs His Entire Business”, published 11 November 2026, with guest Peter Yang of Behind the Craft. No paid sponsor is mentioned; Matt points viewers to Peter’s membership, where Peter shares his skills. This summary, the opinions and the practical advice are our own.
What is a personal AI system?
A skill is a saved set of instructions for one job, such as “prepare a podcast guest brief” or “summarise my week’s business numbers”. Instead of re-explaining yourself every time, you call the skill and the AI follows the same steps.
A personal AI system is a collection of those skills plus a few automations that run on a schedule. Peter’s key point is that this is the part worth investing in. New models and tools arrive every few weeks, and if you keep hopping between them, you never build anything that lasts. Your skills, written down and kept in one place, travel with you to whichever tool is best this month.
💡 Tip: This is the same idea behind owning your email list rather than relying on one social platform. Your skills and context are an asset you control; the AI tool is just the engine you plug them into.
What does Peter Yang’s personal AI system include?
Peter shared a one-page map of his system. Here it is in plain English.
| Layer | What it does | How often |
|---|---|---|
| Chief of staff | Reads email, calendar and Slack and lists what needs dealing with today, with links | Twice a day (7am and 4pm) |
| Advisor | Checks decisions and opportunities against his yearly goals | About once a month, plus ad hoc |
| Council | Runs big decisions past several different AI “perspectives” | For major decisions |
| Content skills | Guest research, interview guides, titles and thumbnails, video editing, intros, turning one recording into a newsletter and social posts | Per episode or video |
| Scheduled bots | Weekly business review, fitness check-in, weekend ideas for the family | Weekly |
| Skill editor | A skill whose job is keeping his other skills tidy | Whenever skills are updated |
It looks like a lot. Both Peter and Matt are clear that you should not build it all at once. Start with one recurring task that would genuinely save you time.
Which AI tools does he use, and does the model matter?
Peter uses three tools, and all three can reach the same skills:
- ChatGPT for browser and computer-use tasks, such as filling in forms. He says its model is not necessarily the best right now, but its browsing is.
- Claude Code, with Claude Opus 5.5, which he calls a game changer for video editing and motion graphics.
- Grok Bot (spelled “GroqBot” in the video description), which he rates as the best personal agent at the moment. Because it runs on a virtual machine, his scheduled tasks keep working when his own computer is off. He also likes its X integration for research.
His headline opinion is that the harness matters more than the model. In other words, choose a tool for what it can do around the model: browse, run on a schedule, connect to your apps.
⚠️ Warning: Peter is open that he gets access to all three tools free as a creator. He says that if he were paying around $200 a month for each, he would probably pick just one. For most small businesses, one well-chosen tool is plenty.
Matt adds his own setup: Claude for writing and title ideas because it matches his voice best, and Codex and ChatGPT models mostly for coding. Everything in the episode works with whichever model you prefer. If you are weighing up Claude models, our guide to which Claude model to use helps.
Not sure where to start? Pick the one admin job you dread each week, write down how you do it in five steps, and save that as your first skill. That is the seed of your personal AI system.
How does the AI chief of staff work?
Twice a day, at 7am and 4pm, Peter’s chief of staff sends him a numbered list of what needs his attention, with a link to each email thread. Instead of opening his inbox, he replies with short instructions by number, such as: set up the invite, draft an email for me to review, I’ll handle this one myself, or do some research.
The important detail: he almost never lets it send emails on its own. It drafts, he tweaks a date or a line, then he sends. Peter describes the chat window as “the new browser”: he works through the AI rather than clicking around each app.
Matt rebuilds a basic version as a scheduled task in Claude, running on weekdays at 8:30am in the cloud with his email, calendar and Slack connected. Here is the prompt in our own wording:
You are my chief of staff. Look at my email from the last 24 hours, my calendar for today and tomorrow, and my Slack.
Give me a numbered list of what needs my attention, most important first.
For each item: one line on what it is, a link to it, and your suggested next step (reply, schedule, delegate or ignore).
Skip newsletters, promotions and anything that doesn't need me.
When I reply with numbers, do what I say for each one.
Draft emails, but never send anything without my approval.
Peter suggested one optional extra: ask it to review your chats and connected sources and tell you something about yourself you might not have noticed. Matt tried it and found it too personal to show on screen, which is a useful reminder about how much these tools can see.
⚠️ Warning: Connecting email, calendar and Slack gives an AI tool wide access to private and client information. Use the smallest set of connections that does the job, keep “draft, never send” in the prompt, and read our guide on briefing a chief of staff agent on when to interrupt you to stop it nagging about the wrong things.
What is the AI advisor, and why is it different?
The chief of staff answers “what needs me today?”. The advisor answers a different question: “am I heading where I said I wanted to go?”
Peter’s insight is that giving the AI everything (all your memory and documents) produces worse advice, not better. So he keeps a single-page Google Doc with his overall plan. Matt’s version has a few short sections: top goals, what he is not doing this year, what gives him energy, what drains him, and non-negotiable family commitments, plus a note on how he wants advice delivered (no sugar-coating).
One rule matters most. Peter writes the doc himself, and the AI may suggest changes but never make them. He calls it his “slop-free zone”.
In Matt’s test, he asked his new advisor skill whether to accept a two-day conference trip in March. It said no, and explained why using his own doc and calendar: the dates clashed with an existing trip and broke his one-trip-a-month rule. That is the difference between generic advice and advice measured against your own plan.
💡 Tip: The advisor runs about once a month for the big picture. Peter says more often than that is too much for long-term direction, although he also uses it as a sounding board when he is stuck. Our guide How to Write a One-Page Plan Your AI Advisor Can Check Decisions Against walks you through writing the doc and turning it into a skill.
How do you keep AI skills improving without making a mess?
Both men admit the first version of any prompt is rough. Matt’s habit was to tell the AI after each conversation to “update the skill so you do better next time”. Peter used to do the same, but found it overfits: the skill fills up with random one-off rules from a single chat.
His better method:
- Open a new conversation.
- Ask the AI to look across all the chats where you used a skill over the last few weeks (Peter suggests about three).
- Ask it to list the feedback you repeat most often, without changing anything yet.
- Pick which points to add, by number.
- Let a “skill editor” skill make the edit and keep the instructions short and simple.
Matt tried this on his short-form video skill. The AI went through 13 review rounds across 9 edits and reported that his most common note, by a wide margin, was that the B-roll did not match what he was saying. He then told it to add points 1, 2, 4 and 6 only.
If your skills folder is already crowded, our Claude skills clean-up guide covers what to keep, merge or delete. For the batch-review method itself, see How to Improve an AI Skill Without Filling It With One-Off Rules.
How does he use AI for video and podcast work?
Peter’s content skills cover the full chain: research a guest, draft an interview guide, suggest title and thumbnail angles based on examples that are doing well, and turn a recording into a newsletter and social posts. The draft interview guide becomes a task he reviews by hand before sharing it with the guest.
For video, he built a simple review screen, similar to the professional tool Frame.io, in about ten minutes. He plays the video, leaves notes at specific moments and sends them to the AI agent to fix, rather than to a human editor. He does not use Premiere.
| Task | Who does it, per the video |
|---|---|
| Short videos (around a minute) | Matt says his skill can almost produce these in one go after lots of iteration |
| Talking-head videos | Peter uses AI more and more |
| Long-form podcasts | Both still use a human editor |
| Video titles | Matt still writes these with his team; he finds AI weak at titles |
| Podcast guest prep | AI drafts research and questions; a human edits before sharing |
Matt’s tip for guest prep is to give the AI just four things: who the guest is, the topic, the audience and the interview format. Then ask for sources, possible angles and questions that build on each other.
🔎 Unconfirmed / rumour: Matt shows that his AI-generated prep suggested an angle about “the AI chief of staff that runs a $1 million one-person business”. That was a suggested headline from the AI, not a verified figure about Peter’s business. Treat it as a pitch, not a fact.
What jobs do his scheduled bots do?
Peter has several bots that run on a schedule and send their results by email, simply because email is what he actually reads. Examples he showed:
- A weekly business review with course sales, newsletter funnel, YouTube stats and what competitor channels are doing.
- A weekend ideas email that searches local parenting blogs for things to do with his kids.
- A fitness check-in fed by a simple lifting tracker he built himself.
- A producer bot that compares how much a topic is being discussed on X over seven days with how many YouTube channels have covered it, to spot gaps.
Matt’s rule for any bot is to define three things: its purpose, what it needs, and what good looks like. If it is recurring, add when it runs and where the result goes. He demonstrates with a “bass coach” project: a practice log in Google Sheets as context, a weekly check-in every Sunday evening, and a small practice-logging app he had the AI build.
Giving each bot one clear job is what makes this work. If you are running several, our guide on splitting work between AI bots without overlap will help, and every recurring job should have a stop condition.
Where does human judgement still matter?
Peter’s final point is the one we would underline. It is easy to let the AI update your skills and documents without looking. He doesn’t. When he asks it to tidy up a newsletter or edit a skill, he asks for a numbered list of exact before-and-after changes in the chat. Then he approves some and rejects others (“1 to 8 are fine, skip 9”).
Anything customer-facing gets human eyes first. That habit costs a minute and protects your reputation.
| What was shown | What is opinion or claim |
|---|---|
| Daily chief-of-staff list with links and draft replies | That Grok Bot is “by far the best” personal agent |
| Advisor answers measured against a one-page plan | That the harness matters more than the model |
| Batch review of 13 rounds of video feedback | That short videos can be almost one-shot |
| Scheduled emails for stats, fitness and weekends | That the X-versus-YouTube gap still works as a content strategy |
| A review screen built in about ten minutes | Any income figures; none were verified in the video |
How to start your own personal AI system this week
- List your repeat jobs. Write down five tasks you do every week. Circle the one that wastes the most time.
- Write it as a skill. Purpose, what it needs, what good looks like, when it runs and where the result goes.
- Pick one tool you already pay for. If the job must run while your laptop is off, choose one that runs in the cloud.
- Keep a human in the loop. Drafts, not sends. Ask for numbered changes before anything is edited.
- Write your one-page plan so future advice is measured against your goals.
- Review after three weeks, in one batch, and update the skill once.
Frequently asked questions
What is an AI chief of staff?
It is a scheduled AI task that reads your inbox, calendar and messages and gives you a short, prioritised list of what needs you, with suggested next steps. Peter Yang’s runs twice a day and drafts replies for him to check.
Do I need ChatGPT, Claude and Grok Bot to do this?
No. Peter uses all three partly because he gets them free as a creator. He says he would pick one if he were paying. Any capable tool that supports saved instructions and scheduled tasks will do.
Is it safe to connect my email to an AI agent?
It can be useful, but it gives the tool access to private data. Connect only what the job needs, tell it never to send without approval, and check the tool’s data and privacy settings first.
What is a skill editor?
A skill whose only job is maintaining your other skills, keeping them short and avoiding one-off rules. Peter uses it after reviewing several weeks of conversations at once.
Where can I get Peter Yang’s skills?
Matt points viewers to Peter’s YouTube channel and his Behind the Craft membership, where Peter shares his skills and courses.
The bottom line
Peter Yang’s setup is impressive, but its real value is the principle behind it: own your instructions and your plan, plug them into whichever tool is best, and keep your judgement in charge. You do not need three paid tools or a dozen bots to benefit. One well-written skill that runs every morning can give you back real time.
Your next step: write your one-page plan this week using our AI advisor plan guide, then set up a single daily briefing with “draft, never send” in the prompt.
Related posts
- How to Brief a Chief of Staff AI Agent on When to Interrupt You
- Claude Skills Clean-Up: What to Keep, Merge or Delete
- AI Agents and Productivity Theater: 4 Principles to Make Them Do Real Work
