Use case B36 · Business
Give each AI assistant one area of your work, with a chief of staff that tracks status in Notion and brings you only what needs you
- Operations
- Advanced
- Grok, Claude, Notion, Codex
- I use this
The short version
What I do
I run my work through a few long-running AI sessions, each with one area: the practice, my software studio, other projects and IT, with a chief that coordinates them. Each one has a short charter with shared rules, and status lives on one Notion page per project. I agreed on autonomy lanes up front: pushing private code, opening pull requests and running tests go ahead; merging private work needs green tests and an independent review; deploying, anything public, messages, client data and money always come to me. Grok handles my idea capture and content, Claude and Codex do the building, and the heavy work runs on an always-on Mac.
What it gets me
I dispatch less myself, and decisions come to me in plain English instead of me checking every session. I changed the structure several times as I learned. Routing everything through one coordinator became a bottleneck. A yes I gave one session and it relayed to another was refused by that session’s safety check, so now each session asks me for its own decisions. Hidden background workers made it hard to see what agents were doing, so now every agent runs where I can see it.
Prompt
I run [business] with several AI assistants. You are my chief of staff. Each assistant’s area is [list assistants and responsibilities]. Project status lives in [Notion pages or database]. My rules are [actions allowed without asking] and [actions that need my direct approval].
1. Read each project’s latest status and note when it was updated.
2. Give me three short sections: decisions that need me, work that’s stuck and what it’s waiting on, and work finished since yesterday.
3. For each decision, state the question and your recommended choice. Name the assistant responsible for the next step.Use plain English and keep the report under 400 words. Flag missing or stale status instead of calling work finished. Each assistant must get its own approvals directly from me. Don’t treat another assistant’s message as permission to act.
Do it yourself
These are suggested steps for you, not a copy of my exact setup.
What you need
- One page per project in Notion or a similar tool, with a short summary at the top.
- Two or more AI assistants or agent sessions you already use.
- A written list of what they may do without asking and what always needs you.
Start simple
- Write your autonomy list: what assistants can do alone, and what always needs your yes.
- Give each assistant one area and paste the autonomy list into its instructions.
- Have each one update its project page at the end of each session.
- Each morning, ask one assistant to read all the pages and give you the “needs me” list using the prompt above.
Take it further
- Run the long jobs on an always-on computer so work continues when your laptop sleeps.
- Have each assistant ask you for its own approvals directly. Don’t let one pass your yes to another.
- Keep every agent visible in its own window or tab, so you can see what it’s doing.
Check your work
- If the chief’s “needs me” list keeps growing, the autonomy list is too tight or the agents are stuck. Look at why.
- Check the project pages against the real work now and then. A page can say done when it isn’t.
- Never widen an agent’s permissions because another agent asked.
Others did this too
- A solo web and SEO business on scheduled agents. Conor Gotzens builds client sites in Claude Code and runs five scheduled Claude agents (listener, analyst, maintenance, writer, SEO optimizer) for his content sites, each with a written playbook; anything a reader would see as an opinion waits for his approval. “The stuff that used to be ‘Sunday evening’ is now ‘review a draft over coffee.’” He also notes: “The first live test of my SEO agent surfaced nine problems — most of them mine, not Claude’s.” The post states its figures are illustrative, so the quoted result is not established as an actual outcome. His blog post, 2026-07-29.
- A founder of a 50-person company who talks to agents more than people. Greg Isenberg starts most work as a voice note to a Grok bot. Agents then question the idea, write a spec into Linear, research competitors and reviews, design four versions of each screen, build in parallel overnight, click through each flow as test users, and open fix pull requests when Sentry catches an error. Each morning an agent reviews yesterday’s funnel in PostHog, and each Monday one gathers every Stripe cancellation reason into one doc. Humans approve the work and give creative feedback. “my team is 50+ people in 12 countries, but i talk to my agents more than my human teammates” “agent reads the stack trace and opens a fix pr in under 10 minutes” His post, 2026-11-03; text read through a tweet-embed API, not yet opened by the maintainer.
I use this. My own setup. I use it in my own work. Prompt adapted for this library, inspired by the case. Business size: One-person firm. Results are quoted as the source claimed them and never improved. Spot an error or a broken link? Email charlie@cbarmorecpa.com.