An AI chief of staff: a v0.1 to take parts from
What mine does, the rules that keep it safe, and eight prompts to build your own.
I've built an AI chief of staff out of Claude. It works through my inbox, my diary, my pipeline and my follow-ups, while I work and while I sleep. I packaged it up for a few colleagues this week, more people asked for it, so here it is.
Version 0.1, work in progress. Rough drafts, shared because people asked. Mine changes most weeks as I learn what works and what breaks, and this page will change with it.
IWhat mine does
Inbox
Triage around the clock
A session checks my inbox every 15 minutes, sorts what needs me, and drafts replies in my voice for me to send.
Email to me
One clean report, not a trickle
Every piece of work ends with one summary email to me, files attached, saying whether I must act.
Pipeline
The board kept current
When work changes a lead or a proposal, it notes it on the board and refreshes next steps on the deals I lead.
Overnight
Scheduled work on a second laptop
About fifty scheduled tasks: briefings, my diary, audio summaries, preparing the weekly email, and checks that the other jobs worked.
Meetings
Transcripts that become memory
Meeting transcripts are pulled in, names corrected at source, and decisions filed into the right project.
Building
It makes the tools it needs
Websites, client portals, small apps and data pipelines, reviewed by a second model before I see them.
It works because I've switched on things most firms switch off: a card it can pay with, its own keys to the web's tools, permission to act without asking first, a laptop that never sleeps, an inbox it can work in, and one memory across all my work. What reduces the risk is a short list of rules. The first two are enforced by tested code that blocks the action; the rest are standing instructions:
- Drafts, never sends, to anyone else. Email to a third party is always a draft. I send.
- Nothing is deleted. Old files move to an archive; deletion goes to the bin.
- Save before you overwrite. A dated copy is kept before any document or data file changes.
- Never fill a gap. A missing fact is left out or marked in square brackets, never made up.
- Never interrupt me unasked. No messages or notifications I haven't asked for.
- A second model checks the work. Anything for a client is reviewed by a model from another company before it reaches me.
Check, Edit, Own still applies: it drafts, I own what goes out.
IIRead this first
Six caveats, and they matter.
- These are rough working drafts. I wrote them for colleagues this week. They haven't been reviewed, and some figures (in square brackets in the deck) come from my own notes rather than a fresh count.
- It is not a product. Mine grew over months from hundreds of small settings, permissions and fixes, each made the day something broke. It works on my two laptops, with my accounts. Yours will need its own work, and some days it will fight you.
- The prompts haven't been tested on anyone else's set-up. Treat each one as the start of a conversation with your Claude, not a recipe. They assume Claude Code, the version of Claude that works on your own computer and can read files, run commands and build things. Claude in the browser or the desktop chat can't do most of this.
- Build the safety rules before the freedom. The two that matter most: email to anyone but you is only ever a draft, and nothing is ever deleted. Only once those are enforced, and working, would I let it run without asking permission. Mine runs that way, and it is the riskiest thing in the set-up.
- It isn't cheap the way I run it: two £200-a-month plans, flat out. You can turn that right down by running things daily rather than every 15 minutes, using cheaper models for routine work, and starting with a few tasks on one machine and one plan.
- Check your client contracts and any data rules before you point it at client email or files. Mine sees most of my working life because I've made that call for my own work. Your clients, and your employer, may see it differently.
IIITake the parts that suit you
Don't copy mine wholesale. You probably do some of this already, and some of mine only makes sense for me. Paste the prompts below into your assistant (the Copy this page for AI button at the top copies the whole page), with this:
Below is a guide I've been given: [paste it, or attach the file. It could be a prompt pack, a colleague's custom instructions or someone's write-up of how they work with AI].
Compare it with how I work: my custom instructions, my projects, what you remember about me and how I've used you recently. If you can't see some of that, tell me what's missing and I'll paste it in.
Then give me three short lists:
1. The few pieces that would fill a real gap for me. For each, one line on what it would add and what, if anything, it would replace.
2. The pieces I already do, so I can skip them.
3. Anything that would clash with how I work.
Change nothing until I say. If nothing in it is worth adding, tell me so plainly. That's a good answer, and I'd rather have it than a list made to look useful.That last line matters more than it looks. Asked to review something, a model will nearly always find something to change; telling it that "nothing" is allowed is how you get an honest answer.
Once you've picked:
Add pieces [numbers] only. If you can't make a change yourself, write out the exact text and tell me where to paste it. For each, tell me where it now lives (which instruction, project or file) so I can find it and take it out later if it doesn't earn its place. Don't add anything that sends, deletes, pays or acts without asking until the two hard rules in prompt 3 are built and tested on my set-up. Don't add anything that sends, deletes, pays or acts without asking until the two hard rules in prompt 3 are built and tested on my set-up.IVThe prompts
Eight prompts, in order, written for Claude Code. Run one at a time and let each settle for about a week before the next. The order matters: the safety rules (prompt 3) come before email (prompt 4), and both come long before letting it run without asking (prompt 7). Paste each as it is, then fill in the square brackets.
1Tell it who you are
Gives it the standing context it reads at the start of every session.
I want you to become my chief of staff over the coming weeks. Start by interviewing me, one question at a time, about:
- my work: the businesses I work for, my role, my main clients and projects
- the people I work with most, and what each needs from me
- how I write: tone, length, words I never use, how I sign off
- what you must never do (for example: send email to anyone but me, delete files, contact anyone on my behalf)
- where my files live and which tools and accounts I use
Then write what you learned into my personal instruction file (~/.claude/CLAUDE.md), in short sections with plain headings. Keep it under 20,000 characters. Show me the file and ask what is wrong before you save it. Tell me how to edit it myself later.Check: open a new session and ask "What do you know about me and how should you write for me?" If the answer is thin or wrong, fix the file.
2Give it a memory
So lessons stick between sessions.
Set up a memory system for yourself. Create a folder for memories, with one small file per fact, lesson or correction, each starting with a one-line summary, and an index file listing every memory in one line. Make sure the index is read at the start of every session.
Rules for using it:
- when I correct you, save the correction as a memory with the reason, so you never repeat the mistake
- before saving, check whether a memory already covers it and update that instead
- archive memories that turn out to be wrong (move them to an archive folder; never delete them)
- never store passwords or keys in memory
Show me the structure and save your first memory: what you learned in our first conversation.Check: correct it on something small, close the session, open a new one and see whether it remembers.
3Build the two hard rules
These reduce the risk of everything that follows. They must be code that runs before every action, not instructions it might forget.
I want two rules enforced by code, using Claude Code hooks (scripts that run before every tool call and can block it), not just written in my instructions:
1. Email to anyone other than me can only ever be saved as a draft, never sent. If any command or tool would send email to a third party, block it and say why. Email to my own addresses ([list them]) may be sent.
2. Nothing is permanently deleted. Deleting a file moves it to the bin instead. Permanent deletion is blocked unless I have explicitly approved that exact action on that exact file in this conversation.
Build the hooks, register them in my settings, and write a short test for each that proves it blocks the bad case and allows the good one. Run the tests and show me the results. Then write both rules into my instruction file too, saying they are enforced by hooks.Check: run the hook tests again, then ask it to draft (not send) an email to a colleague and confirm it is saved only as a draft. Never test the block by asking for a real send to someone else: if the guard failed, it would go. Ask it to delete a scratch file. It should go to the bin.
Then re-run the hook tests every week, for example in a Friday config check. Connector and tool names change, and a renamed tool can silently slip past the guard while everything looks fine.
4Connect your email, drafts only
Then ask it to draft replies to today's mail.
Connect to my email account ([Gmail or Outlook]) so you can read my inbox and save drafts. Use the official API with my own login, and store the access token somewhere safe on this computer, never in a file you might share.
Then:
1. Read today's unread email and tell me, in one list, which threads need me, why, and by when.
2. For the three that most need a reply, draft a reply in my voice (use my instruction file) and save each as a draft in my email, in the right thread.
3. Tell me where to find the drafts.
Do not send, archive, label or delete anything yet.Check: read the drafts in your email. Edit them, send them yourself, and tell Claude what you changed and why, so it can save that as a memory.
5Schedule a morning brief
Once a day, on your laptop.
Create a scheduled task that runs at [7am] on weekdays, on this computer, and emails me (to my own address only) a morning brief:
- email since yesterday that needs me, most urgent first, with a suggested reply where one is needed (saved as a draft, not sent)
- today's meetings, with a two-line brief for each from my email and files
- anything I promised someone and have not yet done
Use the cheapest model that does this well; start with the mid-tier model. Keep the brief under one screen. Run it once now so I can see it. Tell me what happens if my computer is asleep at 7am.Check: read the brief for a week. Tell it what to cut and what was missing.
6A closing email for every piece of work
So you never have to watch it work.
From now on, whenever you finish a piece of work for me, or I ask "close?", do this before you stop:
1. Update the notes for that work (a status file in the project folder) with the date, where it got to, the exact next step and who owns it.
2. Email me (my own address only) one short summary: what was done, what is waiting on me, with the finished files attached. Start the subject with "Done: ".
3. Tell me in one line whether it is safe to close the session.
Save this as a standing instruction and as a reusable routine (a skill) called close.7Switch on bypass permissions
Only once prompts 3 and 6 have run cleanly for a week.
I want you to run without asking my permission for each step, using bypass permissions mode. Before you switch it on:
1. Re-run the tests for my two hard rules and show me they pass.
2. List any other action that should never happen without me (for example payments over [£50], posting publicly, changing my calendar for other people) and propose a hook for each.
3. Build the hooks I agree to, with tests.
Show me the test results and the hooks, and wait for me to say yes before you switch bypass permissions on in my settings. Then tell me how to switch it off again.Check: watch the first few sessions. Anything that surprises you becomes a new rule.
8Add keys, a card and a second opinion
One at a time, low limits first.
Help me give you access to [service: for example Cloudflare, Vercel, GitHub or Apify]:
1. Tell me exactly how to create an access key with the narrowest permissions that will do the job, and what each permission allows.
2. Store it in the computer's password store (Keychain on a Mac), with a small helper so your scripts can read it, never in a project file or this conversation.
3. Propose one small, harmless real change to prove it works, and make it only after I approve that exact change. Then tell me how to revoke the key.
4. Record in my instruction file which keys exist, what each is for, and where it lives.For a payment card: use a card with a low monthly limit, kept for this alone, and add a hook that blocks any single purchase above a figure you choose.
For a second opinion:
Before anything I send to a client, I want it reviewed by a model from another company. Set up [OpenAI's Codex command-line tool, or another] so you can hand it a document and get back a critical review. Use it for client work and for any code change over 50 lines. Treat its findings as advice: tell me what it found and what you changed.What I have that these prompts do not cover
- A second laptop that never sleeps, running about fifty scheduled tasks, so the work carries on overnight and while I travel. Do this only once one machine is running well.
- A logged-in browser it can drive, for sites with no proper access.
- Rules about which model does what: the strongest for judgement and writing, cheaper ones for routine work.
- A record of every mistake it made and the rule that now prevents it. This is most of what makes mine work. Yours will grow the same way, one correction at a time.
VIKeep me posted
I'll share all the variants and features people send with everyone who signs up, and send updates as mine changes. Sign up here, or just reply to Saturday's email to say yes.
Thank you. You're on the list.
VIIThe deck
The nine-page deck I wrote for colleagues on 2nd November: what mine does, what it relies on, the rules, the cost and where to start.