How to use Claude as a Chief of Staff

Most Chiefs of Staff meet Claude the same way: you paste a messy document into a chat and ask for a summary. The summary is fine, and a week later you're pasting the same background into a new chat for the fifth time, as if briefing a clever intern who forgets you every morning.

In the Chief of Staff Network's AI Readiness Diagnostic, 86% of the 1,200+ senior leaders surveyed use AI every day, but only 54% write detailed prompts and iterate, and 12.5% connect AI to their other tools. The skill and project below fix the forgetting. I use Claude as the example because its projects and skills map neatly onto the job, but you can pick any tool your company allows; the connected system is in our guide to building an AI chief of staff.

What is a chief of staff skill for Claude, and how do you install one?

A skill is a folder with a SKILL.md file of instructions that Claude loads whenever a task matches the skill's description. A chief of staff skill holds how you do your recurring jobs, such as the weekly update, the meeting brief and the decision memo, so you never explain them again.

The project holds what Claude should know about one principal; the skill holds how you do a job, and it works in every project and chat. Claude reads only the short description at the top of each skill until a task matches, then loads the full instructions, which is why the description matters more than anything else in the file [CHECK SOURCE: support.claude.com/en/articles/12512198-how-to-create-custom-skills].

Most chief of staff skills in public libraries are written by developers for founders, and they route a question to an imaginary CFO, CTO and head of sales. That's a fun demo with little to do with making one real leader's week work. The one below encodes an actual operating rhythm: your weekly update, meeting brief and decision memo, plus rules about facts and people.

---
name: chief-of-staff
description: Chief of Staff work for [principal]: weekly updates, meeting briefs, decision memos, meeting summaries and leadership agendas. Use when asked to prep, draft or summarize for [principal].
---

# Chief of Staff skill

You help [name], Chief of Staff to [principal, title] at [company].

## Before any task
1. Read the context file in the current project, if there is one. If there isn't, ask me for this quarter's priorities before you start.
2. Check the request against those priorities. If it conflicts, say so in one line at the top.
3. Use only facts from my message, attached files and sources I name. Mark every gap as [CHECK]. Never invent a number, name, date or quote.

## Weekly update to [principal]
Four parts, in this order, under 400 words:
1. Decisions I need from you: each with my recommendation and the date I need it.
2. What moved: outcomes since last week, tied to the priorities.
3. What's coming: the next 2 to 4 weeks, with risks.
4. One topic for our 1:1.

## Meeting brief
For each meeting: why it matters to this quarter's priorities, who's in the room and what each wants, what was decided last time, the decision or outcome we need, and one question I should ask. Half a page per meeting at most.

## Decision memo
Recommendation in the first two lines. Then why now, the options (including doing nothing), trade-offs, who has been consulted, and what happens after the decision. One page.

## Meeting summary from notes or a transcript
Decisions (with who decided), actions (owner and date for each), open questions, and "what we tell the company". Leave out who said what unless I ask. Flag anything that sounds like a decision but has no owner.

## Leadership meeting agenda
Pull from: open decisions, items that missed their date, requests leaders sent in, and next week's priorities. Each item gets an owner, a time and the outcome needed (decide, discuss or inform). Decisions first.

## Rules
- Write in [principal]'s voice only when I ask for a draft in their name, and follow the writing samples in the project.
- Anything about a named person is a draft for me to verify, never a fact.
- Nothing is sent, scheduled or shared without my explicit approval.
- End every output with: what you assumed, and what I should check.

# Weekly update to the CEO · Friday, Nov 7

## Decisions I need from you
1. Approve the VP Engineering offer at the top of the band. I recommend yes, because she has a competing offer. Needed by Tuesday. The ask: reply "approve" or call me.
2. Pick the board's discussion topic: pricing or the enterprise launch. I recommend pricing. Needed by Nov 14.

## What moved
- Series B (priority 1): four partner meetings; two firms asked for data room access
- Enterprise tier (priority 2): second design partner signed
- VP Engineering search (priority 3) and platform migration: on track

## What's coming
- Nov 14: board pre-read goes to directors; finance numbers due Nov 12 [CHECK: close date]
- Risk: the new pricing page needs legal review, and nobody has booked it

## For our 1:1
- The head of sales wants to reopen the Europe decision. How do you want to handle it?

What I assumed: the board meeting is still Nov 21. What to check: the partner meeting count, against your calendar.

Two lines carry most of the weight. The description decides when Claude reaches for the skill, so it names the jobs in the words you'd actually type: prep, draft, summarize. The last rule makes every draft end with its own assumptions, which turns checking from a reread into a short list.

Edit the five formats to match how you work today. If your principal wants the weekly update as three bullets on Slack, write that; a skill written around an ideal rhythm gets dropped by the second week.

Install it

Replace the bracketed names first and save the text as a file called SKILL.md. ChatGPT reads the same format, in a beta that workspace owners switch on [CHECK SOURCE: academy.openai.com/public/resources/skills].

Menus move as both products change, so if a step doesn't match what you see, search the vendor's help center for "skills" [CHECK SOURCE: support.claude.com/en/articles/12512180-use-skills-in-claude].

Make it yours in three runs

Tara Nouveau, head of AI learning at Clarinet, gave our members a four-part formula for writing a skill in a July workshop: define the job, give it context, capture the judgment, then train it with feedback. People skip the third part because it's made of small decisions they don't notice making. Her fix was "make the implicit explicit", which in her inbox triage skill meant lines like "if it comes from Diane, it's most likely an escalation".

For the weekly update it looks like this labeled illustration. Run it three Fridays in a row and note every edit you make.

1. Week one. You move the hiring decision from the bottom to the top, because your principal reads the first two lines and stops. The skill gets: "Decisions I need come first, each with the date I need them."
2. Week two. Your principal replies, "What do you actually need from me here?" The skill gets: "Each decision ends with the ask in one sentence."
3. Week three. You cut a paragraph on a project that's on track. The skill gets: "Workstreams on track get one line between them. Only what's at risk gets detail."

Those three lines are your principal's reading habits written down, and they're worth more than the rest of the format. Split a job into its own skill when someone else, such as your principal's EA, needs it, or when its section grows past a page. Anything about a person's performance, pay or employment never becomes a skill, because a saved procedure makes it too easy to run without thinking.

Jobs that run once a day train slowly, since you get one round of feedback a day, so put two or three worked examples in the file from the start. Once a month, fold the edits you still make into the file, and if you share it, keep one owner and a date at the top. Change the format only after agreeing it with your principal, because by then they rely on its structure.

This week, run your weekly update with the skill, note your edits, and fold them into the file on Friday.

How do you set up a Claude project for your principal?

Create one project per principal, give it a short set of instructions, and add your context file and a few samples of how your principal writes. From then on, do all of that principal's work inside the project, on a plan your company has approved for work data, so every chat starts from the same facts.

Tara Nouveau called the cost of skipping this "the stranger tax": every new chat, you reintroduce yourself and re-explain the company, like walking into a coffee shop where nobody knows your order. A Chief of Staff pays it more than anyone, because the job is mostly context.

A project has two parts. Instructions tell Claude how to behave: who you are, who you serve, what good output looks like and what to flag. Knowledge tells it what's true: the files it reads before it answers. Projects exist on every plan, including the free one, which allows up to five, and paid plans hold far more because Claude searches the files [CHECK SOURCE: support.claude.com/en/articles/9517075-what-are-projects].

In ChatGPT, projects work the same way, and you can choose project-only memory when you create one so nothing leaks in from other chats [CHECK SOURCE: help.openai.com/en/articles/10169521-projects-in-chatgpt].

What goes in the knowledge

Three files cover most of what Claude needs for a principal.

1. Your context file. One page: this quarter's priorities in order, what's off the list, who owns what, the decisions that shape everything else and how your principal wants information. The template, and the Friday routine that keeps it current, are in our guide to context management.
2. Three to five samples of your principal's writing. A company update, a hard email and a board note work well. Claude learns tone faster from examples than from adjectives like "direct" or "warm".
3. The current operating rhythm. The leadership meeting agenda format, your weekly update template, the board calendar. If you don't have these written down, our guide to the operating cadence has them.

Leave out anything your company's rules keep out of AI tools, and anything about a person that you wouldn't read aloud to them.

Tara's point on samples was that people describe a leader's voice in adjectives, "concise, with a bit of whimsy", and the model can't do much with that. Real writing carries the habits nobody thinks to list: how they open, how they deliver bad news, which words they never use. If your principal rarely writes, she suggested meeting transcripts, used only where recording is approved and people were told.

As an illustration, a project for the CEO of a 120-person company might hold six files: the context file, four samples (an all-hands email, a hard note to a VP, a board letter and a Slack announcement) and one operating rhythm document. That's enough to draft in the right voice and order without searching a pile of old decks.

The instructions

These work as project instructions in Claude and in ChatGPT.

You are working with me, [name], Chief of Staff to [principal, title] at [company].

How to work
- Read the context file in this project before every task. If my request conflicts with this quarter's priorities, say so first.
- Write for [principal]: [e.g. decision first, one page, numbers in a table, plain words].
- Match the tone of the writing samples in this project when you draft anything in [principal]'s name.
- Use only facts from my message, the files in this project and any source I name. Mark every gap as [CHECK]. Never invent a number, name, date or quote.
- Treat anything about a named person as a draft for me to verify.

What I usually need
- Decision memos: recommendation first, then options, trade-offs and who has been consulted.
- Meeting prep: why the meeting matters, what was decided last time, one question I should ask.
- Weekly update to [principal]: decisions I need, what moved, what's coming, one topic for our 1:1.
- Summaries: decisions, owners and dates first; discussion only if I ask.

Before you finish
- List anything you assumed.
- Tell me which parts I should check against the source.

One project per principal, kept current

I'd resist making a project for every initiative. The pricing decision affects the board pre-read, which affects the all-hands, and splitting the work means each project sees only part of the picture. If you support two leaders, make two projects. If you also run a leadership-team program, give it a separate project with nothing personal in it.

Tara's test for what belongs where: if you keep attaching the same documents, they belong in a project; if you keep typing the same instructions, they belong in a skill. For a functional leader such as a CTO, put the function's vocabulary in the context file, because "the migration" means something specific there that Claude won't guess.

The cost is upkeep. A context file three weeks stale gives you confident drafts built on old priorities, so spend ten minutes on it right after your Friday update. When you replace a document, delete the old one the same day, and once a quarter empty the project except for the context file and samples.

Before you share the project with an EA or a deputy, read it as if you were them, because it may hold notes on people or priorities your principal hasn't announced. If anything gives you pause, build them a separate project from the shareable parts.

Do every piece of work for a principal inside that principal's project, and update its context file the same day a priority or an owner changes.

This week, create the project, paste the instructions, and add your context file plus three pieces of your principal's writing. Then redo one task you did last week and compare the drafts.

What should you do in your first week with Claude?

Spend the first week doing real work, one job a day, inside your principal's project. By Friday you should know which jobs Claude does well for you, how much editing each draft needs, and which edits belong in the skill.

Pick tasks you'd do anyway this week, so the comparison with your usual way is honest.

1. Monday, set up and test. Create the project, then ask Claude the test prompt below. Its answer shows you what your context file is missing faster than rereading it would.
2. Tuesday, meeting prep. For tomorrow's three most important meetings, ask for why each matters to this quarter's priorities, what was decided last time and one question you should ask. Check every fact against your notes.
3. Wednesday, a hard piece of writing. A decision memo, a tricky email to a peer, a message about a change people won't like. Ask for two versions with different tones and pick from them.
4. Thursday, numbers. Upload a spreadsheet you already use, such as a hiring plan or a budget tracker, and ask what changed, what looks wrong and what question the leadership team will ask. Ask for any calculation to be shown so you can check it.
5. Friday, your weekly update. Give Claude your notes from the week and ask for a first draft in your usual format. Then keep a short list of every edit you made to each day's draft.

That edit list is the most useful thing you'll make all week, because it tells you what to add to the project instructions and the skill.

Using only the files and instructions in this project, answer in under 300 words:
1. What do you think [principal]'s top three priorities are this quarter, in order? Quote the line in the context file each one comes from.
2. Who owns each priority?
3. If I asked you for my weekly update to [principal] right now, what would you need from me that isn't here?
4. List anything in the files that looks out of date or contradicts something else.
Don't guess. If the files don't say, write "not in the files".

Draft, react, redraft

Ashlea Holt described a good working rhythm in her episode with Tim Bristow. She treats AI as an analyst that gives her something to react to, "first draft, third draft", and if the second draft still isn't clear, she takes it as a sign her own thinking isn't there yet. Claude rarely nails a hard memo on the first pass, and it gets you to a version you can argue with much sooner than a blank page.

Aria Finger, who was Chief of Staff to Reid Hoffman when we spoke, told our podcast she uses an assistant constantly on her phone in voice mode, because talking gives it more context than typing. If your requests come out thin, try saying them out loud. Our prompt library has prompts for the rest of the job.

[SCOTT: one line on how you use Claude yourself, and the job you'd hand it first]

Measure week one by how many edits each draft needed, and whether that number fell from Monday to Friday. Time saved comes later, because you'll check everything at first. Once your principal sees faster turnarounds they'll ask for more drafts, sooner, so decide now which work stays slow on purpose, such as anything about people.

Which Claude plan should a Chief of Staff use, and is company data safe?

Use the plan your company has approved for work, which usually means a Team or Enterprise plan that IT administers. Keep company data off personal accounts, because the rules for your data, the admin controls and the contract all change with the plan.

Anthropic sells Claude on individual plans (Free, Pro, Max and Ultra) and on business plans (Team and Enterprise). On Free, Pro and Max, you choose in your privacy settings whether your chats can be used to train future models; if you allow it, Anthropic keeps those chats for up to five years, and if you don't, the standard period is 30 days. Team, Enterprise and the API run under commercial terms and are excluded from that training choice [CHECK SOURCE: anthropic.com/news/updates-to-our-consumer-terms].

ChatGPT is similar: Business and Enterprise content is excluded from training by default, and individual plans have a setting you control [CHECK SOURCE: openai.com/enterprise-privacy]. On Claude's Team and Enterprise plans, an owner decides which connectors exist, whether Claude can send email or move files without asking, and whether memory is on for members [CHECK SOURCE: support.claude.com/en/articles/10166901-use-google-workspace-connectors]. Those are the controls your security team will ask about.

What the hesitation is about

When we asked companies what slowed AI down for our Chief of AI Report, security, compliance and ethical concerns made up about 41% of the barriers named, ahead of adoption and change management at about 38%. One respondent's CTO stopped AI use in internal strategy meetings over privacy worries. That's a reasonable reaction to a tool nobody has reviewed, and a Chief of Staff who works around it on a personal account damages the trust the job runs on.

Company plan first, company data second. If your company hasn't approved a plan, learn on public material and your own writing, and bring IT the ready-made request to IT in our setup guide.

What to keep out, even on an approved plan

Most companies keep four things out, or allow them only under tighter rules:

1. Anything about a named person's performance, pay, health or employment.
2. Legal matters and anything under privilege.
3. Customer personal data, unless your policy covers it.
4. Board and investor materials before they're final, unless your principal agrees.

If your company's AI policy doesn't say this yet, the template in our guide to AI and the Chief of Staff is a fair starting point to bring to legal.

Claude remembers things across chats by default on individual plans, and each project keeps its own memory. You can read, edit or delete what it has saved under Settings, or open an incognito chat that isn't saved [CHECK SOURCE: support.claude.com/en/articles/11817273]. Memory can carry a detail from a project into your general chats, so keep anything sensitive inside the right project and look at what's been saved once a month.

Before Friday, ask IT which AI plan is approved for company data, then open your privacy settings and check the training choice and memory on whatever account you use today.

What else can Claude do for a Chief of Staff?

Beyond projects and skills, Claude can run multi-step research, create files and spreadsheets with working formulas, connect to your calendar, email and documents, and run saved jobs on a schedule. Each has a near equivalent in ChatGPT, so the habits carry over if your company standardizes on the other one.

Across the four core responsibilities in what a Chief of Staff does, that means a pre-mortem on a plan, next week's agenda from this week's notes, a first pass at your weekly update, a company update in your principal's voice or a status summary across five teams' notes. Claude speeds up the first draft and leaves you the judgment, the politics and the call.

Features and plan rules change often; check Claude's help center before you rely on any one of them [CHECK SOURCE: support.claude.com/en/collections/18031719-features-and-capabilities].

Claude or ChatGPT?

Use the one your company licenses. When we asked companies for our Chief of AI Report in early 2025, 34% named ChatGPT as their tool of choice, against 8% for Microsoft Copilot, 7% for Claude and 6% for Gemini. That mix has moved since, and both now offer projects, skills, connectors and scheduled tasks. If you have a choice, run the same three real tasks in each for a week and keep the one whose drafts you edit least.

When should you connect Claude to your calendar, email and documents?

Connect your tools once the project and skill are producing drafts you trust, usually after two to four weeks, and only on an approved plan. Connecting earlier just lets Claude make confident mistakes with more data.

With Google Workspace connected, Claude can search and read Gmail, draft replies, see and create calendar events, and find and read files in Drive; Docs, Sheets and Slides can be edited in a pane beside the chat. Sending, forwarding, sharing, moving or deleting needs your approval by default, and on Team and Enterprise plans an owner turns the connectors on and decides whether that approval can be skipped [CHECK SOURCE: support.claude.com/en/articles/10166901-use-google-workspace-connectors].

Microsoft 365 has its own connector [CHECK SOURCE: support.claude.com/en/articles/12542951-set-up-the-microsoft-365-connector]. ChatGPT's Google and Outlook apps work in a similar way, with write actions that workspace admins enable one by one [CHECK SOURCE: help.openai.com/en/articles/12512241].

Calendar first, email last

I'd connect the calendar first, because an error there costs little and the payoff is daily: meeting prep that knows who's in the room. Documents come next, so Claude can find last quarter's plan without you uploading it. Email comes last, read-and-draft only for at least a month, because an inbox holds the most sensitive material you touch and the most ways to embarrass your principal.

Scheduled tasks come after connectors. You describe a job and how often it should run, and Claude runs it in the cloud and reports back, even with your laptop closed; they're available on all paid plans [CHECK SOURCE: support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-claude-cowork]. ChatGPT has scheduled tasks too, with up to 10 active at once [CHECK SOURCE: help.openai.com/en/articles/10291617]. A Monday brief built from your calendar and the context file is the classic first one.

A wrong draft in a chat costs you a minute, while a wrong calendar invite reaches a leader, so add one connection at a time and run it for a week before the next. Once you're scheduling jobs, you need written rules for what Claude may do alone, which our guide to building an AI chief of staff covers; our guide to AI tools covers what to add around Claude, such as a notetaker.

What goes wrong in a Chief of Staff's first month with Claude?

Most early problems come from five habits: one endless chat, trusting numbers, work on a personal account, drafts that drift from your principal's voice, and handing over the thinking. Each has a simple fix.

Chiefs of Staff who describe themselves as AI-first average $189,128 in base salary, against $148,999 for self-described AI novices, according to the Chief of Staff Network's 2025 Salary Report (nearly 500 respondents). AI-first Chiefs of Staff also tend to sit in more senior roles, so the gap reflects more than tools, but employers are paying for people who use AI well.

Sort a week of your edits before you change anything. Mostly fixing facts means the context file is thin or stale; mostly fixing tone means the project needs better samples; mostly fixing structure means the instructions or the skill are vague. If you're rewriting the argument itself, you handed over the thinking.

The last one is the expensive one

The first four mistakes get caught. The fifth often doesn't, because the output looks finished: a memo with three balanced options and no view is easy to produce and useless to a leader who needed a recommendation. Claude is very good at building the case around a view, and a poor substitute for having one.

Write your recommendation in one sentence before you ask Claude for any draft that carries judgment, and put that sentence in the request.

If every memo starts from a model's draft, your own arguments get rusty, so once a week write one recommendation from scratch before you open Claude. To see where you stand against other Chiefs of Staff, take our AI Readiness Diagnostic.

When the project, the skill and a first connection are working, building an AI chief of staff turns them into a connected system with written rules.