Source: raw/He_Built_an_AI_System_That_Runs_His_Entire_Business.md + ai-research/behindthecraft-peter-yang-2026-11-11.md (membership page, fetched 2026-11-11)
Creator: Matt Wolfe (host) with Peter Yang, “Show Me Your Workflow” series · URL: https://www.youtube.com/watch?v=pvl3z_GyU2Y · Duration: not recorded · Platform: YouTube (auto-captions, fetched 2026-11-11). No sponsor segment in the transcript.
Peter Yang spent over ten years as a product manager in the Bay Area, most recently at Roblox, and left in June to make content full time. He runs his one-person business on a set of skills shared across ChatGPT, Claude Code and Grok Bot: a chief of staff that briefs him twice a day, an advisor that checks decisions against a one-page plan he writes himself, a council for big decisions, content skills for his podcast and videos, scheduled email reports, and a “skill editor” whose job is to keep the other skills clean. Wolfe rebuilds simple versions of the chief of staff, the advisor and a weekly check-in on screen in Claude, so most of the episode can be copied directly.
Key Takeaways
- Skills are the durable asset, not the tool. Yang’s advice to non-creators: stop bouncing between models and build “a bunch of skills and automations that is unique to them” that you can bring to any tool. All three of his harnesses read the same skills.
- “The harness matters more than the model.” His split:
- ChatGPT for browser and computer use (form filling, manual browser work), even though it “arguably doesn’t have the best model right now”.
- Grok Bot for every scheduled task, because it runs on a virtual machine and keeps working when his laptop is off. He calls it “by far the best personal agent right now” and uses its X integration for research.
- Claude Code with Opus 5.5, which he calls “kind of a game changer” for video editing and motion graphics.
- He gets all three free as a creator and says that if he paid $200 a month he would pick one.
- Chief of staff (daily, 7:00 a.m. and 4:00 p.m.). It reads email, calendar and Slack and sends a numbered list of what needs attention, each item linked to its thread. He replies by number (“1, set up the invite. 2, draft the email from the review. 3, I’ll handle this manually. 4, do the research”). It drafts emails but he almost never lets it send without review.
- Advisor (monthly). Grounded in a single-page Google Doc, not his whole memory (“too much crap”): goals for the year, what gives him energy, what drains him, and family commitments. He writes that doc himself and calls it his “slop-free zone”; AI may suggest changes but never make them. The chief of staff is about today; the advisor is about direction, so it runs once a month.
- Council for his biggest decisions: several AI perspectives on one question (Wolfe compares it to Nate Herk’s Roast Room).
- Content skills: podcast guest research, an interview guide he sends the guest, title and thumbnail angles (it now copies screenshots of thumbnails that are doing well), video editing, intros, and turning one recording into a newsletter plus social posts. Interview guides go to a Linear ticket for a manual pass.
- Scheduled reports arrive by email because email is what he actually reads: a weekly business review (course, newsletter funnel, other channels), a weekend email that searches local parent blogs for things to do with his kids, and a fitness summary from a fitness app he vibe-coded, which exposes his lifting stats over MCP.
- A producer bot finds topics on X before YouTube. It compares the number of X posts on a topic over seven days with the number of channels that have made videos on it, to find gaps. Wolfe says this was his own core strategy in 2022-2023.
- The skill editor fixes a real failure mode. Updating a skill after every conversation “tends to overfit into whatever your current conversation is” and fills the skill with one-off rules. Instead, open a new thread, have the AI read all related threads from the past few weeks, find the feedback you keep repeating, and apply only that through the skill-editor skill, which keeps skills simple.
- Human review is a rule, not a step he skips. For any edit to his newsletter or skills he asks for “the exact before and after copy that you changed in a numbered list”, then accepts items by number (“one to eight make sense, but nine doesn’t”). Anything customer-facing gets his eyes first.
- Video editing is not solved. He built a frame.io-style review page in about ten minutes where he leaves timestamped notes and sends them to the agent. Wolfe one-shots short videos with a tuned skill but still uses a human editor for long-form; Yang uses a manual editor for his podcast. Wolfe also still writes titles with his team because he finds AI “pretty bad at titling”.
Wolfe’s on-screen builds
1. Chief of staff as a Claude scheduled task. Scheduled → plus → set up manually. The prompt he used:
You’re 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, 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.
- Settings he chose: run on auto (no asking), Opus 5.5 as model, 8:30 a.m. on weekdays only, the advanced option the captions render as “require this computer off” so it runs in the cloud, the needed connectors switched on, push notifications.
- Optional add-on Yang 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.”
2. Advisor skill from a one-page doc. Sections: top goals, what I’m not doing this year, what gives me energy, what drains me, non-negotiable family commitments, and how I want advice (“don’t sugarcoat things”). Top line: “I only edit this doc. AI can suggest changes, never make them.” Share the Doc URL with Claude (Google Drive and Docs connectors on) and ask it to make a skill; he says this took about 30 seconds. Test: “Use my advisor skill. I’ve been asked to speak at a conference in March, two days of travel. Should I do it?” The answer cited a trip already blocked that month, his one-trip-per-month limit, and keynotes listed as something that drains him.
3. Weekly check-in as a Claude project. Wolfe’s example is a bass-practice coach: project instructions (“help me learn bass… progressively harder each week”), a practice-log spreadsheet added as Drive context so it stays synced, then “give me this week’s check-in” scheduled for Sunday 6:30 p.m. He then had the project build a small practice-logger app to replace the spreadsheet.
4. Batch skill improvement. “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.” It reviewed 13 rounds across nine edits and returned recurring issues (B-roll not matching the words or arriving at the wrong moment, framing and zoom, white text repeating subtitles or covering B-roll, sound effects and number animations out of sync). He then picked which numbered points to add.
Bot design rule
For every bot, define three things: its purpose, what it needs to do the job, and what good looks like. If it recurs, add when it runs and where it sends results. Yang prefers many single-job bots over “one giant AI that does everything”; besides the producer bot he has one for his wife’s errands and a cooking app for the air fryer and pan.
Try It
- Start with one recurring task. Both hosts close on this: pick one workflow that would save time, give it good context, use it, and improve it.
- Paste Wolfe’s chief-of-staff prompt into a scheduled task that runs in the cloud on weekdays, with “draft, never send” left in.
- Write your one-page plan by hand (goals, not-doing list, energy, drains, family) and build an advisor skill on it. Run it monthly, not daily.
- Replace per-conversation skill updates with a batch review every few weeks: “read my recent threads that used this skill, list the feedback I repeat, and propose edits as numbered before/after changes.” Accept by number.
- Send scheduled outputs where you already read (email for Yang) rather than to another app.
- Copy the X-versus-YouTube gap check if you make content: posts on a topic in seven days against the number of channels that have covered it.
Implementation
Tool/Service: Claude scheduled tasks and projects (Wolfe’s demo); ChatGPT, Claude Code and Grok Bot sharing one skill set (Yang’s setup). Setup: Google Drive and Google Docs connectors for the advisor doc; email, calendar and Slack connectors for the chief of staff. Cost: Not stated for the setup. Yang mentions $200 a month as the price at which he would pick a single tool. His skills are sold in his Behind the Craft membership (behindthecraft.com); the fetched page lists a “/chief” skill “to brief you twice a day on what needs a reply, ready for you to act on by number”, “/advisor”, “/council”, “Skill editor” and “Human review” skills, but no price was found on the page. Integration notes: Scheduled work that must run when the laptop is closed needs a cloud runner (Grok Bot’s VM for Yang; the cloud option in Claude’s scheduler for Wolfe). Compare the “update the skill after every video” loop in animate skill playbook, which is the per-session approach Yang warns overfits.
Open Questions
- The exact label of the Claude scheduler setting the captions render as “require this computer off” was not confirmed.
- Membership price: not shown on the fetched behindthecraft.com page.
- Yang’s council structure (which perspectives, which models) is not shown in the episode.
- The “Comcast call” video Wolfe mentions is referenced but not described beyond the editing back-and-forth.
Related
- Grok Bot — Running a Business on an Agent Team — where Yang runs his scheduled tasks
- Claude Code Scheduled Tasks — scheduled runs in Claude, the feature behind Wolfe’s chief of staff
- Claude Code Personal OS (Moritz Kremb) — a chief-of-staff build from an episode Yang hosted
- Nate Herk’s Three Daily Automations (Matt Wolfe) — the previous Show Me Your Workflow episode and its Roast Room council
- Council — Native macOS App for Multi-Model Blind Deliberation — a packaged version of the council idea
- Claude Opus 5.5 — the model both hosts default to here
- Ben AI’s Skill-Engineering Framework — another way to test and improve skills