What Should an AI Chief of Staff Produce Each Week?

Most "AI chief of staff" posts describe a role. This post describes a document. A chief of staff is only useful if the output is something a person can read in five minutes and correct in two.

Display is the docs workspace for agents and people. Any agent publishes docs and files to a link. Your team opens the link with the account they already use. They comment and edit in-line. Any agent or teammate can publish the next version to the same link.

1) What an AI chief of staff does and does not do

An AI chief of staff does four jobs:

• It collects facts from the tools you already use.
• It compares this week with last week.
• It lists what needs a human decision.
• It writes all of this in one document.

It does not make the decisions. It does not send messages for you unless you ask. It does not replace a human chief of staff who negotiates, mediates and reads a room.

Your assistant does the collecting. Claude, ChatGPT and Codex each reach calendars, ticket trackers and CRMs through their own connectors. Which connectors you get depends on your plan and your administrator. Check your assistant's connector list before you plan the brief.

2) The weekly brief: five sections, one page

Keep the brief to one page. Each section answers one question.

Write each item as one sentence with one link to the source. Do not paste the source into the brief. A reader who doubts a number opens the link.

3) Sources the agent needs, and the access it needs

Give the agent the least access that does the job.

1. Calendar. Read access to the next seven days.
2. Tickets. Read access to the boards you own.
3. CRM. Read access to deals above a size you choose.
4. Call notes. Read access to the folder where notes land.

Do not give write access to any of them. The brief is a report. If the agent finds a problem, it lists the problem and a human acts.

4) Publish to one link and let comments correct the facts

A brief in a chat window dies in the chat window. Ask the agent to publish the brief to Display with company access. The link then works for everyone in your organization. People sign in with the Google or Microsoft account they already use.

Keep the same link every week. The agent publishes each new brief as the next version. Last week's version stays in the history.

Now the brief can be corrected:

• A teammate selects a wrong sentence and writes a comment: "This deal closed on Friday."
• On Monday, the agent reads the open comments before it writes the new brief.
• It fixes the fact at its source, replies to the comment and resolves it.

This is the part a chat assistant cannot give you. The people named in the brief see it, and they correct it on the page. See review comments for how the agent reads and answers them.

If the brief is for investors rather than your team, start from a template.

5) Prompt and schedule

Copy this prompt. Change the names in brackets.

You are my chief of staff. Each Monday, write my weekly brief.
 
Read: my calendar for the next 7 days, the tickets I own,
deals above [amount] in the CRM, and last week's call notes.
 
Write one page with five sections:
1. Decisions waiting
2. Numbers that moved since last Monday
3. Slipped follow-ups
4. Meetings ahead
5. Risks
 
Rules:
- One sentence per item. One source link per item.
- Do not paste source text.
- Mark any fact you could not verify as "unverified".
 
Before you write, read the open comments on last week's brief
and correct the facts they point to.
 
Then publish the brief to Display with company access.
Update the same link each week. Reply with the link.

For the schedule, use the scheduler you already have. Claude Code routines run on a schedule on the Pro, Max, Team and Enterprise plans, per Anthropic's docs on 14 October 2026. Codex also has scheduled tasks, and its docs say availability depends on plan and workspace settings. A plain cron job that starts your agent also works. Display has no scheduler of its own.

6) An example brief

This sample uses invented data.

Decisions waiting

• Approve the Halden Bikes renewal at a 10 percent discount. Sales needs an answer by Wednesday. (source link)
• Choose between two launch dates for the mobile app. Engineering prefers 4 November. (source link)

Numbers that moved

• Trial starts rose from 112 to 131. (source link)
• Support backlog fell from 41 to 28. (source link)

Slipped follow-ups

• The security questionnaire for Polar Foods is five days late. (source link)

Meetings ahead

• Tuesday, 10:00: Lars at Polar Foods. He asked about data residency last time. (source link)

Risks

• One engineer is out for two weeks. The launch date depends on her review. (source link)

The sample is short on purpose. Eleven lines give a founder a full week.

7) When this does not work

• Your sources are empty. The agent cannot find slipped follow-ups if nobody writes call notes.
• Your assistant has no connector to a key tool. Then the brief lacks that section. Say so in the brief.
• The brief grows past one page. Cut sections. A long brief is not read.
• Some people must see parts of the brief but not all of it. Split it into two documents with two access rules.

We checked the product statements in this post against the Display docs on 14 October 2026. Connector availability differs by assistant and plan. Check your own account.