The AI Spotlight Follow
The One-Person AI Operating System: A Chief of Staff, an Advisor That Says No, and a Skill That Fixes the Other Skills
AI Agents

The One-Person AI Operating System: A Chief of Staff, an Advisor That Says No, and a Skill That Fixes the Other Skills

A solo creator runs his whole business and a chunk of family life on a stack of reusable AI skills and scheduled bots that work in ChatGPT, Claude and Grok Bot alike. Here's every piece, how to build simple versions yourself, and the one rule that keeps it from turning into slop.

Advertisement

Twice a day, at 7 a.m. and 4 p.m., an AI reads one man's email, calendar and Slack and hands him a numbered to-do list. He replies with something like "1: set up the invite. 2: draft a reply. 3: I'll handle this one. 4: research it." The AI does the rest, except hit send.

Once a month, a different AI checks his decisions against a one-page plan he wrote himself and tells him, bluntly, when he's drifting.

Another AI preps his podcast guests. Another edits his short videos. Another finds things to do with his kids on the weekend.

He left a decade-long product management career in the Bay Area in June to work for himself. This is how one person runs what's essentially a one-person media business, and the setup is surprisingly copyable.

Advertisement

The big idea: build a system, not a tool habit

The problem with AI right now is that a new model or tool lands every other week. If you keep hopping between them, you end up with a pile of half-learned apps and nothing durable.

His answer: build a personal AI system of skills and automations that is unique to you and portable to any tool.

The core building block is a skill: a saved set of instructions for one job, so you never have to re-explain yourself. Skills have become a de facto standard. Anthropic describes them as reusable, filesystem-based packages built around a SKILL.md file that loads on demand, and ChatGPT and Codex support the same format.

That portability is the point. He uses three different assistants, and all three can pull the same skills:

  • ChatGPT
  • Claude Code
  • Grok Bot

The model underneath matters less than you'd think.

The map of the whole system

At first glance, it looks complicated. Broken down, it's four layers.

1. Running life and business

  • A chief of staff that looks at everything coming in and says what to deal with today.
  • An advisor that checks decisions against the year's goals.
  • A council: several AI perspectives he runs his biggest decisions past.

2. Content systems

Skills for:

  • Researching podcast guests.
  • Building an interview guide to send guests.
  • Titles and thumbnails.
  • Video editing.
  • Intros.
  • Turning one recording into a newsletter plus social posts.

3. Scheduled tasks

Things that simply show up in his inbox: business stats, a fitness check-in, weekend ideas for the family.

4. The skill editor

A skill whose only job is keeping the other skills clean. It's one of the smartest parts of the whole thing, and we'll get to it.

Crucially, you don't build this all at once. Start with one recurring task that would actually save you time.

Harness over model

Asked how he picks models, his view was clear: the harness matters more than the model. The harness is the app or agent environment wrapped around the model: its tools, browser, memory and scheduling.

His current split:

  • Claude Opus 5.5 is "kind of a game changer" for video editing and motion graphics, in his view. Anthropic released it on September 22, 2026, pitching it at long-running agentic and knowledge work.
  • ChatGPT may not have the best model right now, he argues, but its browser and computer use are so good that he sends it any form filling or manual browser chores.
  • Grok Bot, in his opinion, is the best personal agent right now. Because it runs on a cloud virtual machine, he moved almost all his scheduled tasks there. A laptop-based agent stops when the laptop is off. It also has X integration for research.

He's candid that he uses all three because he gets access as a creator. If he were paying $200 a month for this, he says, he'd probably pick one. Most people should.

A second perspective from the same conversation is worth noting too: Claude tends to be best at matching a writer's voice for scripts and titles; Codex and ChatGPT models still lead for coding projects; and OpenAI's always-on dots agents, launched September 29, haven't been used as much as expected. Everything here works with whichever model you choose.

Piece 1: The AI chief of staff

This is the piece to build first.

How his works

Twice daily, it reads email, calendar and Slack, then sends a numbered list. A sample:

  1. Send a 4 p.m. invite to a collaborator.
  2. Reply to a sponsor with a brief.
  3. Pick a date for an upcoming live session.

Each item links to the original thread. He replies with numbers and short instructions: set up the invite; draft the email for review; I'll handle this manually; go research this.

He almost never lets it send email on its own. It drafts; he tweaks a date or a line, then hits send. That's still dramatically faster than writing from scratch, and he doesn't need to open Gmail.

His description of the shift is striking: the chat interface has become the new browser. The browser becomes a middle layer he barely touches, because the agent uses it for him. Call it a new operating system.

Build a simple version in Claude

Here's how to set up a basic chief of staff with Claude's scheduled tasks:

  1. Open Scheduled in Claude and click + to set one up manually.
  2. Write a prompt along these lines: You're my chief of staff. Review 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, and your suggested next step (reply, schedule, delegate or ignore). Skip newsletters, promos and anything that doesn't need me. When I reply with numbers, do what I say. Draft emails, but never send them without me.
  3. Set it to run automatically and choose your default model (Opus 5.5 in this example).
  4. Schedule it for 8:30 a.m. on weekdays only, so weekends stay quiet.
  5. Under advanced settings, choose to run it without your computer, so it executes in the cloud.
  6. Turn on the connectors it needs (email, calendar, Slack).
  7. Pick notifications: push, email or Slack.

Anthropic's documentation confirms cloud scheduled tasks run on Anthropic-managed infrastructure, so they keep working when your laptop is closed.

The optional twist

Add this to the end of the prompt: Review our chats, my connected sources and your memory, and tell me something special or unique about me that I haven't realized yet. Be honest. You don't have to be positive.

In testing, it worked. The daily list came back prioritized with action items, and the "something I noticed about you" note was personal enough that it had to be blurred before showing anyone.

Piece 2: The advisor that tells you no

The chief of staff is about today. The advisor is about direction: are you heading where you said you wanted to go?

Don't feed it everything

His key insight: models are good at absorbing context, but if you tell one to "look at all my memory and documents," it drowns in noise.

Instead, he keeps one Google Doc, one page, covering:

  • What he wants this year.
  • What gives him energy.
  • What drains him.
  • Family priorities, not just business.

That gives the advisor something concrete to measure against. When an opportunity comes in, it can say "yes, that fits" or "you said you didn't want to do that this year."

The slop-free zone

His strictest rule: he writes this doc, not the AI. He calls it his "slop-free zone." The AI can suggest changes but never make them.

Build your own in 30 seconds

A simple version of the doc needs a few sections:

  • Top goals.
  • What I'm not doing this year.
  • What gives me energy.
  • What drains me.
  • Non-negotiable family commitments.
  • How I want advice: don't sugarcoat; tell me what I need to hear, not what I want to hear.

Put "I only edit this doc. AI can suggest changes, never make them" at the top. Then share the doc link with Claude (with the Google Drive and Docs connectors on) and ask it to turn the doc into an advisor skill. It takes about 30 seconds.

The test

The prompt: Use my advisor skill. I've been asked to speak at a conference in March. Two days of travel. Should I do it?

After about a minute, the answer: short version, no, unless specific dates are already blocked. A second March trip would break the one-trip-per-month limit in the plan. And speaking means building a deck, and keynote prep was listed as something that drains energy.

That's not generic advice. It's the plan talking back.

How often to run it

The chief of staff runs every day. The advisor runs about once a month. Any more often, for long-term direction, is too much.

He also uses it ad hoc, as a 24/7 sounding board when frustrated or confused. He admits he's not sure how healthy that is, but says the advice is good.

Piece 3: The skill editor

This is the most underrated idea in the setup.

The overfitting trap

The natural habit is to fix a skill after every bad response: "Update the skill so you never do that again." The problem is that after enough conversations, your skill becomes a junk drawer of one-off rules that overfit to whatever happened that day.

The better loop

Instead:

  1. Open a fresh thread.
  2. Ask the AI to review all your recent threads that used a skill, say the past three weeks.
  3. Have it find recurring patterns in your feedback.
  4. Run the result through a skill editor skill whose job is to keep instructions simple.

Because the AI can see the whole history at once, it spots what you correct over and over instead of overreacting to one mistake.

A real example

Applied to a short-form video skill, the review covered 13 rounds of feedback across nine edits and surfaced repeat issues:

  • B-roll that doesn't show what's being said, or shows it at the wrong moment.
  • Framing and zoom problems.
  • White on-screen text that repeats the subtitles or covers the B-roll.
  • Sound effects and number animations out of sync.

The fix: pick the items that matter (in this case points 1, 2, 4 and 6) and tell it to fold only those into the editing skill.

He didn't hand-write his long skill prompts. The skill editor wrote and maintained them.

Piece 4: AI video editing, with a human eye

One of his projects shows both the power and the limits. He had AI edit a video about an AI calling his cable company, Comcast, for him. It went back and forth on music, motion, graphics and edits, and even blurred his address and phone number on its own. But it still took his creative judgment and many rounds of feedback.

His verdict: video editing isn't solved yet.

A homemade review tool

Leaving timestamped feedback in a chat was annoying, and pro editing software felt too complicated. So he built his own review interface, similar to Frame.io, in about 10 minutes. He plays the video, jumps to moments, leaves notes like "only show three things in this list," then sends the notes to the agent, which goes off and fixes them.

His current split:

  • Short videos of a minute to a minute and a half: close to one-shot, after many iterations on the skill.
  • Talking-head videos: increasingly AI-edited.
  • Long-form podcast episodes: still a human editor.

Piece 5: Podcast prep in four inputs

His podcast skill researches a guest, pulls their top insights and drafts an interview guide. A companion title-and-thumbnail skill looks at screenshots of thumbnails that are currently performing well and suggests angles. Its suggestions are, by his own admission, "a good start."

The draft guide goes into a Linear ticket for a manual pass, then gets shared with the guest so they can prepare.

A simple version needs only four inputs:

  1. Who the guest is.
  2. What you're talking about.
  3. Who the audience is.
  4. The interview format.

Then ask for sources, possible angles and a handful of questions that build on each other, so it feels like a conversation, not a questionnaire. In testing, one suggested angle was exactly what the final episode leaned into: the AI chief of staff behind a one-person business.

One job stays human for now: final titles. AI is still weak at titling, and on YouTube the title is everything.

Piece 6: Scheduled bots for work and life

He uses so many AI apps that he routes results to email, because email is what he actually reads. Examples:

  • A weekly business review that pulls stats, summarizes how the newsletter, funnel and a new course are doing, and researches other channels.
  • A weekend email that searches local parenting blogs for things to do with his kids.
  • A training summary from a fitness app he vibe-coded, connected over MCP to his lifting stats.

He also runs a producer bot for research, a bot for errands his wife asks him to run, and a little app that helps him cook in the air fryer and pan.

One bot, one job

Rather than one giant AI, each bot gets a narrow job with three definitions:

  • Its purpose.
  • What it needs to do the job.
  • What good looks like.

If it's recurring, add when it runs and where results go.

Example: a bass practice coach

A simple demo of the pattern:

  1. Create a Claude project called Bass Coach, with the goal of building skills progressively, keeping a schedule and not replaying the same songs.
  2. Add a Google Sheets practice log (song, time practiced, notes, self-rating) as synced context through the Drive connector.
  3. Ask for this week's check-in. It came back song by song, flagged what was working against the goal, and gave one change for next week.
  4. Schedule the check-in for Sunday nights at 6:30 p.m.
  5. No spreadsheet? Ask the project to build a simple practice logger app to record song, date, duration, rating and notes. It took a couple of minutes.

Piece 7: Spot ideas before they trend

His producer bot compares X activity over the past seven days with how many YouTube channels have covered a topic. The gap is the opportunity: a subject buzzing on X that hasn't been covered much on YouTube yet. AI news tends to break on X first, then YouTube and LinkedIn, then Facebook later.

It's not perfect, he says, but useful, with judgment applied.

The rule that holds it all together

His closing principle: human judgment and taste still matter.

When AI cleans up a newsletter, he asks for every change as a numbered before-and-after list. Then he can say "1 through 8 are fine, skip 9." The same goes for skill updates: show the exact before-and-after first.

Anything user-facing or customer-facing gets human eyes before it ships.

Why this matters

  • Portability beats loyalty. Skills written once can follow you from tool to tool as models change.
  • Always-on agents are mainstream now. Cloud scheduled tasks in Claude, OpenAI's dots and VM-based agents like Grok Bot all run without your laptop open.
  • Context design is the new skill. A one-page plan beats a thousand documents.
  • One person can run a lot. This system handles inbox triage, strategy checks, content ops and family logistics.

Counterpoints and risks

  • Privacy. A chief of staff reads your email, calendar and Slack. Review connector permissions and keep sensitive accounts out.
  • Never auto-send. Drafts-only is the safe default. Agents can misread context, and inboxes can carry prompt-injection attempts.
  • Emotional reliance. Using an AI advisor as a constant sounding board is something even its owner isn't sure is healthy.
  • Cost. Running three premium assistants adds up. Pick one and go deep.
  • Skill rot. Without periodic consolidation, skills bloat. The skill editor exists for a reason.

Practical takeaways

  • Start with one recurring task, ideally the chief of staff.
  • Make it draft, never send.
  • Write a one-page plan yourself and let the advisor measure against it.
  • Run the advisor monthly, the chief of staff daily.
  • Improve skills from weeks of threads, not one bad answer.
  • Give every bot one job, clear inputs and a definition of good.
  • Route results to where you'll actually read them.
  • Demand numbered before-and-after diffs for anything AI edits.

FAQ

Do I need all three AI tools? No. Pick one. The system is built from portable skills precisely so the tool can change.

What's the difference between a skill and a prompt? A prompt is a one-off instruction. A skill is a saved, reusable package of instructions, and sometimes scripts and references, that the AI loads when relevant.

Can these run while my computer is off? Yes, if they run in the cloud. Claude's cloud scheduled tasks and VM-based agents keep working with your laptop closed. Desktop-only automations don't.

Should the AI send emails for me? Best practice here is no. Let it draft, then review and send yourself.

How do I keep skills from getting messy? Periodically review several weeks of conversations at once to find recurring feedback, and use a skill editor to keep instructions short.

The bottom line

The most useful AI setup isn't the newest model. It's a small set of well-scoped skills, a few scheduled bots and two strict rules: write your own plan, and review everything before it leaves the building.

Build the chief of staff this week. Write your one-page plan next week. Add the skill editor when your skills start getting messy. Within a month, you'll have the start of an operating system that's yours, whichever AI happens to be on top.

Sources

AI Just Proved Something About Prime Numbers That Had Beaten Mathematicians for Over a Century. One Expert Didn't Expect to Live to See ItUp next · 16 min readAI Just Proved Something About Prime Numbers That Had Beaten Mathematicians for Over a Century. One Expert Didn't Expect to Live to See ItBuried in OpenAI's 700-paper math release was a proof that the Riemann zeta function has no zeros beyond 7/8, the quasi-Riemann hypothesis. Here's what that means in plain English, why experts are stunned, and what it does to the future of being a mathematician.Keep reading →