AI prompts for Chiefs of Staff

Most prompt lists for Chiefs of Staff open with "Act as a world-class chief of staff" and get back advice that fits any company on earth. The model was told to play a role. Nobody told it about your principal, this quarter, or what your leadership team decided last Tuesday.

Each prompt below is built around a real output of the job, names the sources the model may use, sets the format and ends with a check. They're grouped by the four core responsibilities, work in ChatGPT and Claude, and run best inside a project with your context file, which our beginner's guide to Claude sets up in an hour. The index at the end lists the 40 prompts in our other guides.

Prompts for priorities and decisions

These three prompts lay out a decision so your principal can make it faster: options from a messy thread, a pre-mortem on a plan, and a weekly sort of incoming requests against the quarter's priorities. Each gathers what a decision needs and leaves the choice with you and your principal. The full decision memo format is in our guide to briefing, and a prompt that drafts the memo itself is in our AI guide.

Lay out the options from a messy thread

Most decisions arrive as a 40-message email chain or a Slack thread where three leaders talk past each other. Before you write anything for your principal, get the positions straight.

Below is a thread about [topic]. I'm Chief of Staff to [principal], who will decide.

[Paste thread]

Using only the thread:
1. State the decision in one sentence, phrased as a question.
2. List each option that's been proposed, including doing nothing. For each: who supports it, their main argument, and the cost or risk they've raised.
3. List the facts people disagree about, separately from the opinions.
4. List what's missing to decide: data, a person not yet consulted, a deadline.
5. Say who should own the decision, if the thread suggests it, and who only needs to be consulted.
Quote short phrases where a position matters. Don't recommend an option. Mark anything unclear as [CHECK].

Step 5 matters more than it looks. I'd bet a lot of the stuck decisions in your company are stuck because nobody agreed who decides. Our guide to the Chief of Staff mandate shows how to write down the three tiers we use, Decide, Recommend and Not own.

Run a pre-mortem on a plan

A pre-mortem asks the team to imagine the plan has failed and explain why. AI is good at the first pass, because it isn't attached to the plan and won't worry about upsetting its author.

Here's a plan we're about to commit to: [paste plan, or attach it]. Context: [company stage, the goal, the deadline, who owns delivery].

Imagine it's [date six months out] and the plan has clearly failed.
1. Write the five most likely reasons it failed, most likely first. For each, name the early warning sign we'd see in the first month.
2. Name the assumption the plan depends on most, and how we could test it in two weeks.
3. List which teams or people the plan needs that aren't named in it.
4. Suggest one change that would reduce the biggest risk without changing the goal.
Be specific to this plan. Skip generic risks like "poor communication" unless you can say where and why.

A list of failure modes from a model reads as criticism if it lands in a leadership meeting unannounced. Take the two or three that ring true to the plan's owner first, one to one.

Sort requests against the quarter

Your principal will ask for more than the quarter can hold. A weekly pass gives you a fair basis for saying no.

Here are the requests [principal] and the leadership team sent me this week: [list each request, who asked, and any deadline]. Use the priorities and the "off the list" section in our context file.

For each request:
1. Which priority it serves, or "none".
2. My suggested action: do it, delegate it (to whom), park it until [date], or decline.
3. For any decline or park, a two-sentence reply I could send that names the priority it competes with.
Then list any request that conflicts with something on the "off the list" section, and anything that looks like a new priority we haven't agreed.

The last line is the useful one. When the same kind of request keeps arriving, it's usually a priority nobody has named, and it belongs in the next planning conversation as a question for your principal.

This week, run the request sort on Friday afternoon and bring the "new priority" list to your next 1:1.

Prompts for the operating rhythm

These two prompts catch what slipped this week and prepare a quarterly business review. Running the rhythm is a large part of the job: 61% of Chief of Staff job postings ask for it, through the operating cadence, OKRs or KPIs, according to the Chief of Staff Network's 2026 Pathways Report (more than 1,800 job postings).

Prompts for the weekly update and the leadership meeting agenda live in those guides, and our guide to the operating cadence sets out the full rhythm. All of them are in the index below.

Find what slipped this week

A promise made in Tuesday's meeting has no owner by Thursday, and by the next leadership meeting nobody remembers who said it. Run this every Friday before you write your update.

Here are this week's meeting notes and my commitments list: [paste or attach]. Today is [date].

1. List every commitment with a due date on or before today that the notes don't show as done. For each: owner, what was promised, to whom, and where it was made.
2. List actions from this week's meetings that have no owner or no date.
3. List decisions that were made but not yet sent to the people affected.
4. For each item in 1, draft a one-line nudge I could send the owner, friendly and specific.
Use only the material provided. If an item might be done but the notes don't say, put it under "check with owner" instead of "slipped".

Send the nudges yourself, and choose which to send. A tool that chases a senior leader on your behalf can cost you more goodwill than the slipped item was worth.

Prepare a quarterly business review pack

A quarterly business review asks each team the same questions, and the Chief of Staff usually builds the pack. AI can draft the questions per team and check each submission before it reaches the room.

We run a quarterly business review on [date] for [teams]. Attached: last quarter's plan for each team, their current metrics, and our priorities for next quarter from the context file.

1. For each team, draft five review questions: two on results against plan, one on the biggest miss, one on what they need from other teams, and one on next quarter's priority.
2. For each team, list the data I should ask for before the review, with the metric and the period.
3. Flag any team whose metrics and stated status don't match (for example, "on track" with a falling number).
4. Draft a one-page pre-read template every team fills in the same way.
Use only the attached material. Mark missing data as [CHECK].

Step 3 is the one leaders feel. A status of "on track" next to a falling number is the conversation the review exists for, and finding it in advance lets you raise it with the team lead privately first.

Prompts for communication

These three prompts turn leadership decisions into a company update, prepare your principal for the hard all-hands questions, and rewrite a message for a different reader. 90% of the 1,200+ senior leaders in our AI Readiness Diagnostic had used AI to draft internal communications in the past 30 days.

That saves hours, and a company that receives five AI-drafted updates a week learns to skim them. What keeps an update sounding like your principal is real examples of their writing, the company's own principles, and a person who edits the result. Our guide to communications covers writing in your principal's voice and stress-testing a hard announcement.

Turn this week's decisions into a company update

Abhishek Vora, who was Chief of Staff to the CEO at LaunchDarkly when we spoke, said in his episode on Chiefs of Staff leading AI adoption that company updates used to take him about ten hours. Most of it went on gathering inputs and matching the message to the company's operating principles and the CEO's tone.

Giving AI a few examples of that tone plus the operating principles turned it into a first draft in seconds; gathering the inputs from people still took time. Run this the afternoon of the leadership meeting, while the decisions are fresh.

Here are this week's leadership meeting decisions and the "what we tell the company" notes: [paste]. Attached: three recent company updates from [principal] and our operating principles.

Draft a company update from [principal]:
1. Open with the one change people will notice most, and why.
2. Cover each decision we agreed to share, in plain words: what changes, for whom, from when, and who to ask.
3. Link each decision to an operating principle or priority where it honestly fits. Don't force it.
4. Leave out anything not marked for sharing.
Match the tone, length and sign-off of the attached updates. Under 350 words. Then list any decision where the reason given in the notes is thin, so I can ask [principal] before this goes out.

The last instruction is the safeguard. An update that announces a decision without a reason invites the hallway version, and AI will happily invent a plausible reason if you don't stop it. Take a thin reason to your principal as a question.

Prepare for the hard all-hands questions

When an all-hands goes wrong, it's usually in the Q&A, when someone asks the question the leadership team hoped to avoid. Your principal can rehearse the talk alone; the questions need someone to play the room.

[Principal] is presenting [topic] at the all-hands on [date]. Attached: the slides or talking points, and recent context (team changes, results, anything people are worried about).

1. List the 10 questions employees are most likely to ask, including the uncomfortable ones. For each, say which group is most likely to ask it.
2. For each question, draft a short, honest answer using only the attached material. Where the material doesn't support an answer, write "we don't know yet" and suggest what we can commit to instead.
3. Flag any talking point that could be read as a promise we can't keep.
Plain language. No spin.

Go through the uncomfortable questions with your principal. The value is in hearing which ones they'd rather not take, because those are the ones that will come up.

Rewrite a message for a different reader

The same decision goes to the board, the leadership team and the company in different shapes, and each version has to say the same thing.

Here's a message written for [original audience]: [paste]. Rewrite it for [new audience, e.g. the board / a frontline team / a partner company].

Keep every fact and commitment exactly as written. Change what this reader needs: the order, the level of detail, the jargon, and what they're being asked to do. Tell me in a short list what you cut, what you added and why. Don't add new claims.

The "keep every fact" line stops the rewrite from softening a commitment on its way from the board note to the team update.

Prompts for cross-functional projects

These three prompts bring five teams' notes into one status, read a spreadsheet before the leadership meeting, and draft a job description for a role outside your expertise. AI helps here because it reads five documents in the time it takes you to open them, and it doesn't skip the badly written one.

Research and competitor scans have their own prompts in our guide to research, and stakeholder maps in our stakeholder map guide.

Build one status view from every team's notes

A cross-team project status is only useful if every team is measured the same way. Ask for that explicitly, or the model will repeat each team's own optimism.

Project: [name]. Goal: [one line]. Deadline: [date]. Attached: this week's notes or updates from [teams].

1. Make a table: team, what they committed to this period, what's done, what's late, status (on track, at risk, off track) by your reading of the facts. Where a team's own status differs from yours, show both.
2. List dependencies between teams that are late or unclear, and who needs to talk to whom.
3. List decisions the project needs from [principal] or the leadership team, with a date for each.
4. Write a three-line summary for [principal]: overall status, the biggest risk, the one ask.
Use only the attached notes. Mark anything you inferred as [INFERRED].

A team that reports "on track" with half its commitments late needs a conversation, and it's better that you have it than the CEO.

Read a spreadsheet before the leadership meeting

Finance, hiring and pipeline spreadsheets arrive the night before a meeting. Know what changed before you walk in.

Attached is [the spreadsheet] for [meeting] on [date]. Last period's version is attached too, if I have it.

1. What changed since last period? List the five biggest changes with the numbers.
2. What looks wrong or inconsistent (totals that don't add up, a number that moved without explanation, a missing month)? Show your calculation for each.
3. What are the three questions the leadership team is most likely to ask about it?
4. One line I could open the discussion with.
Show every calculation so I can check it. Don't round unless I ask.

Check every number it flags yourself before you raise it. A model that misreads a column flags a problem with the same confidence as a real one, and a false alarm in front of the CFO costs more than a missed one.

Draft a job description for another team's role

Michael Finkbeiner, who was Chief of Staff at Vertosa when we spoke, explained his method in his episode on using AI for hiring. The model doesn't know your company, but it knows a lot about what a sales account executive generally does, so start there and add what's specific to you.

I need a first draft of a job description for a [role] on our [team]. The hiring manager is [name, title].

Step 1: Describe what this role usually does at a company like ours ([stage, size, industry]): typical responsibilities, the skills that matter, and how success is usually measured in the first year.
Step 2: Now adapt it using what's specific to us: [what the team needs, the problems this hire solves, who they work with, the level, location, and anything the hiring manager told me].
Step 3: Write the job description: summary, the three outcomes for the first year, responsibilities, must-have and nice-to-have skills.
Then list the five questions I should ask the hiring manager before this is posted.

The questions at the end matter most, because the hiring manager's corrections tell you what the role is. For a Chief of Staff role itself, use our job description guide.

How do you check AI output before it reaches your principal?

Check every name, number, date, quote and commitment against the source, then read the draft once as the person receiving it. Ask the model for a first pass, and treat its check as a list of places to look.

AI's errors don't look like errors. A wrong total sits in a clean table, and a misattributed quote sounds exactly like the person. Our AI guide lists the real mistakes one Chief of Staff's system made in its first months: a task assigned straight to his boss, one colleague's update credited to another, notes prepared on someone who had already left.

Once your principal notices one AI error in something you sent, they'll reread everything you send for a month. Run this on anything that will reach your principal, the leadership team or the board.

Here's a draft I'm about to send to [reader]: [paste draft]. Here are the sources it was based on: [paste or attach notes, data, documents].

1. List every factual claim in the draft (names, numbers, dates, quotes, commitments) and mark each as: matches the source, differs from the source (quote both), or not in the sources.
2. List anything about a named person that could be read as a judgment of them.
3. List any sentence that commits [principal] or the company to something.
4. Read it as [reader]: what's the first question they'll ask, and is it answered?
Don't rewrite the draft. Just give me the lists.

Then do your own pass on the first two lists, against the source, every time. Anything about a person stays a draft until you've verified it yourself, whatever the model says.

What a check looks like on real work

Abhishek Vora described the far end of this in the same episode. After a run of acquisitions and funding rounds among LaunchDarkly's competitors, his team ran a ten-minute research report and got seventeen pages back. It sped up their response plan, and it still had to be read "with a fine-tooth comb", because some of it was wrong.

A weekly update is quicker, and it follows a pattern. As an illustration, the check lists fourteen claims and eleven match the source. Two differ: a launch date that moved last week, and a headcount from an old plan.

The fourteenth isn't in the sources at all: a sentence saying the sales team "agreed" to a new process, when the notes only show they discussed it. You fix the date and the number in a minute. The invented agreement matters more, because it commits a team to something it never agreed, and people remember that kind of error.

How much checking each draft needs

Check in proportion to who reads it and what it commits. Anything for your principal, the board or the whole company gets the full check, every time. Prep notes only you will read need a glance at the names and numbers you'll repeat out loud, and board material gets a second reader, because by the time you've built a deck you can no longer see its errors.

Keep a short log of what the checks catch. After a month it shows where your setup is weak: wrong dates usually mean a stale context file, and invented agreements usually mean the prompt lacks the "only use the sources" rule. Fix the cause in the project or the skill, and once a job has run cleanly for a month, move it to spot checks on names and numbers.

Treat the model's own check as a list of places to look, and verify every name, number and commitment against the source yourself.

This week, run the check on the next draft you send upward, and note how many claims it marked as "not in the sources".

What makes a good AI prompt for a Chief of Staff?

A good prompt for this job has six parts: who you are and who the output is for, the job and the decision it serves, the sources it may use, the format, the standard it must meet, and what to check. Most of the words go on context. The instruction itself is usually one line.

A model already knows how to write a memo. It can't know that your CEO reads only the first two lines, that the pricing decision was made last week, or that one VP will object to anything that touches her budget. Each part closes one of those gaps, and the sources rule stops the model filling the rest with plausible guesses.

In our AI Readiness Diagnostic, 54% of the 1,200+ senior leaders surveyed write detailed prompts and iterate on what comes back, which leaves almost half who don't. When we asked companies for our Chief of AI Report what they'd learned, one respondent named exactly these two skills: proper prompting and source checking.

A weak prompt, rebuilt step by step

Take the most common request after a leadership meeting, "Summarize these notes", and add the six parts one at a time, in this worked illustration.

1. The weak version returns a tidy page of everything discussed, in the order it was discussed. Your principal reads the first two lines, which are about start time and attendance.
2. Add the role, reader and job: "I'm Chief of Staff to our CEO, who reads two lines. Pull out the decisions, owners and dates for the follow-up." Now the page leads with decisions, but it lists a debate about pricing as a decision.
3. Add the sources rule: "Use only these notes. Mark anything unclear as [CHECK]." The pricing item now reads "[CHECK] discussed, no decision recorded", which is true and useful.
4. Add the format and standard: "A table of decision, owner, date, then three bullets on open questions. Match the follow-up I sent last week," with that email attached.
5. Add the check: "List what you assumed, and any action with no owner." It flags two actions nobody took.

Each step fixed one failure with something you knew and the model didn't. The judgment calls sit in steps 2 and 3. People overcorrect by pasting forty pages of history, and the model weighs a two-year-old plan the same as this week's decision, so give it the context file and this task's material only.

The other overcorrection is asking the model to recommend before you've decided what you think. For anything with judgment in it, write your own view in one line and ask the model to build or attack it.

The reader changes the prompt more than the task does. A board member needs context they don't have, a team lead needs only their actions, and a principal reading on a phone needs five lines at most. If you support a functional leader, put the function's vocabulary in the context file, or the model will translate "the migration" into something vague.

End every prompt with a checking step: ask the model to list what it assumed and what you should verify against the source.

Show the standard with examples

The standard is where most prompts fall short. "Direct and warm" means something different to everyone; three real examples mean one thing. That's what took Abhishek Vora's company updates from ten hours to a first draft in seconds.

[SCOTT: the one prompt you'd keep if you could keep only one]

How do you save and share prompts so you stop retyping them?

Put recurring context in a project, turn any prompt you've run three times into a skill, and keep a short shared library with an owner for the prompts your team uses.

Project instructions hold how you work for one principal. A skill holds how to do one job, and works everywhere. Our guide to Claude for Chiefs of Staff includes a ready chief of staff skill and the install steps, and our guide to context management covers the context file the project holds.

Both products change these menus often, so check their help centers if a step doesn't match what you see [CHECK SOURCE: help.openai.com/en/articles/10169521-projects-in-chatgpt].

A team library that stays useful

If an EA, a chief of staff team or the leaders you support use your prompts, keep them in one shared document. Give each a title that names the output, one line on when to use it, the date it was last tested and an owner. Remove any prompt nobody has used in a quarter, because a long library of stale prompts teaches people to distrust all of it.

Shared prompts also make outputs consistent: when every EA and Chief of Staff on the leadership team uses the same meeting summary prompt, leaders read the summaries faster. Our older post on moving from radar jammer to air traffic controller has inbox triage and response prompts that fit well in a shared library.

This week, find the prompt you've typed most often and save it as a skill or in your project instructions.

The full library: every prompt in our guides

Our other guides hold 40 more prompts, grouped here by job. Each lives on one page only, next to the explanation of the job it serves, so there's one version to improve, and each has the same Chat with AI button as the prompts above.

If a job you do every week isn't on this list, tell us, and describe what you do on a Friday afternoon.