AI Workflow Design

Build an AI Operating System for Your Life: Peter Yang's Workflow

Peter Yang does not run his one-person creator business from a single all-knowing chatbot. In this conversation with Matt Wolfe, he shows a collection of small, named workflows: an inbox-and-calendar chief of staff, an advisor grounded in a plan he wrote himself, repeatable production skills, and bots that deliver narrow reports on a schedule. The useful lesson is how those pieces are bounded and reviewed, not the number of agents on his screen.

The useful idea

Write down what matters to you, choose one recurring job, give an agent only the sources it needs, and define what it must leave for your approval. Run the job manually until the output is dependable; only then schedule it. Peter's larger system grew by improving specific workflows, not by asking an agent to “run my life.”

Watch Peter Yang's Workflow

Host: Matt Wolfe. Guest: Peter Yang, who shares additional workflows and skills at Behind the Craft. This article explains the demonstrated system; it does not imply that a paid course or three subscriptions are required.

The System, From the Human Out

  1. Human-owned direction. Peter keeps a one-page plan with his goals, priorities, energy drains and family commitments. His advisor may question it, but does not silently rewrite it.
  2. Reusable job descriptions. A skill records when a job runs, what information it needs, the expected result and the approval point. This reduces re-explaining a task every morning.
  3. Tools with different strengths. Peter currently uses ChatGPT, Claude Code and Grok Bot. He favors Claude for much writing and editing, browser-capable tools for web work, and Grok Bot for his scheduled cloud routines and X research. Those are his preferences in this episode, not a universal model ranking.
  4. Small specialist jobs. A daily brief, podcast research, a video cut, a weekend suggestion and a trend scan each have a separate purpose. One failure is easier to spot than a failure inside a giant general-purpose agent.
  5. Human decisions. Peter reads the brief, chooses what to do, reviews customer-facing messages and signs off on changes to the skills themselves.

Peter calls the surrounding harness more important than the model in many workflows: connectors, computer access, schedules, files and the ability to verify an output change what an agent can actually finish. His view is useful, but test it on your own task before paying for multiple products. He notes that creator access lowers his personal cost and that he would likely choose fewer tools if paying the full bill himself.

A Chief of Staff Is Not an Advisor

At 06:33, Peter's chief of staff reviews incoming email, calendar and Slack twice a day. It returns a prioritized, linked list: what needs a reply, what can wait and what action it recommends. He can answer by item number to ask for a draft or follow-up. He says he almost never lets it send an email without his review. The output is a decision queue, not an autonomous executive.

At 07:58, the advisor does a different job. It checks a decision against Peter's own written plan and pushes back when an opportunity conflicts with his longer-term goals. He uses it on demand or roughly monthly, rather than making every day a strategy review. A separate “council” offers several perspectives on bigger decisions, but the episode does not establish that a multi-agent vote is more accurate than an expert or a well-prepared human decision. Peter's Behind the Craft materials describe the chief, advisor and council as distinct skills.

The Skill Editor: Improve the Instructions, Not Just the Answer

Peter's cleverest feedback loop is at 12:48. Instead of adding a new rule after every disappointing result, he reviews a batch of recent feedback. If he repeatedly corrects B-roll choices, framing, on-screen text or sound effects, the skill editor proposes a change to the video skill. A human decides whether that pattern deserves a permanent rule. One-off preferences stay one-offs.

That is a practical way to keep a skill short and useful: save the correction only when it recurs, preserve examples of good work, and test the revised skill against a real job. Claude Code documents skills as reusable instruction files with optional supporting material and evaluation. The episode shows Peter's own cross-tool setup, but OpenAI's current skills documentation says eligibility and installation vary across ChatGPT surfaces, while Grok Bot has its own skill and routine system. A portable text brief is a good starting asset; automatic synchronization across all three is not something to assume.

Production Help and Scheduled Bots

For video, Peter uses timestamped feedback so the editing agent can find exactly which cut, caption or visual needs changing. He says short clips can now come close to a usable first pass after repeated tuning, but his longer videos still involve a human editor. The same separation appears in podcast work: an agent researches a guest, supplies source links and drafts interview angles; Peter and the host still choose the questions, title and thumbnail. AI speeds up preparation and iteration without making taste or factual review optional.

His scheduled jobs are similarly narrow: business metrics, a fitness check-in from connected data, and weekend ideas for his family. One “producer” bot looks for topics attracting discussion on X before YouTube coverage catches up. That is an idea signal, not proof a video will perform. Check the original posts, avoid copying a creator's work and compare the proposed topic with what your audience actually asks.

The scheduling detail matters. Grok Bot's docs say a routine can run on its cloud computer; Anthropic says new Cowork scheduled tasks can run remotely as well. A task that depends on a local file, browser or computer-control session can have different requirements. Confirm the exact connectors, schedule, time zone and run history rather than assuming “cloud” means every local resource remains reachable.

Build One Workflow This Week

  1. Pick a recurring annoyance. Choose something you already do at least weekly and can judge quickly, such as a morning inbox-and-calendar brief. Do not start with autonomous email sending or a sensitive finance or health task.
  2. Write a one-page brief. State the job, trusted sources, format, deadline, what counts as urgent, what to do when a source is unavailable, and what always requires approval.
  3. Connect the minimum. Start with read access to one inbox and calendar, if your platform and organization allow it. Do not grant broad write access just to make a summary.
  4. Run it manually three times. Check every source link, missing item, stale date and suggested action. Keep a short error log and revise the brief only for repeated problems.
  5. Schedule and monitor. Add a daily or weekday run only after the manual version is useful. Check its first several runs, the time zone and any approval behavior. Pause it when the output becomes noise.

Here is a JQ AI SYSTEMS starter brief adapted from the episode's method, not a quotation of Peter's private skill:

Each weekday morning, review my authorized calendar for today and tomorrow and my authorized email from the last 24 hours.
Return a numbered list with: priority, reason, source link, deadline, and one proposed next action.
Separate "needs my decision" from "information only". If a source is unavailable, say so; never invent an item.
Draft replies only when I ask for an item by number. Do not send messages, change my calendar, or update a source file without my explicit approval.
After each run, report the sources checked and any failures.

The success measure is modest: fewer missed commitments and less time sorting the same inputs, without incorrect sends or extra review burden. If it saves five minutes but demands ten minutes of corrections, the workflow is not ready to automate.

Keep the System Under Your Control

Behind the Craft is Peter's commercial membership, and the episode is a creator workflow demonstration. No independent time-savings study or audited revenue result is supplied for these agents. Peter's setup also touches email, Slack, family plans and possibly health or business data. Give each agent only necessary access, keep the plan and high-stakes decisions human-owned, and audit the output before any external action. Anthropic's computer-use guidance specifically warns that computer actions have risk and recommends careful monitoring, especially around sensitive records.

The strongest takeaway from the closing discussion is not “use more agents.” It is to choose a workflow you can inspect, give it context you trust, and keep improving the result while your judgment remains the final step.

Video Chapters

TimeTopic
00:00Peter Yang's AI workflow
01:57Building a personal AI system
04:30Choosing the right AI tools
06:33Building an AI chief of staff
07:58Building an AI advisor
10:56Automating a daily briefing
12:48Improving skills with a skill editor
16:17AI video editing and feedback
18:04Podcast prep and thumbnails
20:28Scheduled bots for work and life
24:25Finding video ideas before they trend
25:58Keeping human judgment in the loop

Sources and Method

Checked 11 November 2026. The episode and supplied transcript support Peter's demonstrated workflows and statements; we did not inspect his private accounts, skill files, agent logs or billing. Behind the Craft confirms his published course structure and skill categories. Product behavior and limits were checked against Anthropic's skills docs, Anthropic's scheduling guide, OpenAI's skills guide and Grok Bot's routine guide. The starter brief and pilot checklist are our editorial adaptation, not Peter's verbatim prompts.

Common questions

Do I need ChatGPT, Claude Code and Grok Bot to copy Peter Yang?
No. Peter uses all three partly because he tests AI tools professionally and says he receives creator access. Pick one platform and one recurring job first. Add a second tool only when a specific capability or cost difference justifies it.
Will one skill automatically work in every AI app?
No. Reusable instructions can often be adapted, but eligibility, installation, permissions, tool access and synchronization differ by product. Check each platform before assuming the same skill or connector is available.
Can a scheduled AI briefing work while my laptop is closed?
A cloud-based routine can, if it has the required connected sources. A job that needs a local app, local files or live computer control may still need the desktop computer and app available. Test the exact task and review its run history.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call