Using AI for Standups and Meeting Notes — A Real Workflow

Meeting notes rot fast. Two days later nobody can remember whether something was decided or merely discussed, and the action item that mattered is buried in paragraph four.

This is the three-step workflow we use to fix that. Total time: about four minutes.

Step 1: Extract decisions and owners

Paste your raw notes or transcript, then run this:

You are a chief of staff reviewing raw meeting notes.

Extract only what was actually stated. Do not infer, do not invent
owners, and do not soften anything.

Return three sections:

DECISIONS — what was decided, one line each
ACTION ITEMS — task | owner | due date (write "unassigned" or
"no date given" where these weren't stated)
OPEN QUESTIONS — anything raised and left unresolved

If a section is empty, write "None."

The "do not infer" line is what makes this trustworthy. Without it, models helpfully assign owners based on who talked most — which is how phantom commitments get created.

Our Meeting Summarizer prompt is a fuller version of this step.

Step 2: Turn your action items into a standup update

Take only the lines that belong to you and run:

Turn these into a standup update with three sections: Yesterday,
Today, Blockers.

Rules: one line per item, past tense for Yesterday, no filler,
never say "working on". If there are no blockers write "None."
Under 80 words total.

That last constraint is small and does a surprising amount of work — it forces concrete verbs instead of status theatre.

See Daily Standup Writer for the version we keep pinned.

Step 3: Send the follow-up

For the one or two items that need someone outside the meeting to act:

Write a short follow-up message to [person] asking them to
[specific action] by [date]. Reference that it came out of today's
[meeting name]. Warm, direct, under 80 words. No preamble.

Email Rewriter handles the polishing pass if the first draft lands slightly off in tone.

Why splitting it up matters

Each step has one job. A single prompt asking for a summary, a standup update and a follow-up email produces three mediocre outputs, because the model optimises for an average of three different formats. Splitting them costs thirty extra seconds and roughly doubles the quality.

Two guardrails

• Verify owners and dates. The constraint reduces invention; it doesn't eliminate it. Scan the action items against your own memory before sending anything.
• Mind what you paste. Meeting content is often confidential. Use whatever tool your organisation has actually approved, and strip names or figures that don't need to be there.

Make it a habit

Save the three prompts somewhere one keystroke away. The workflow only pays off if running it is faster than deciding whether to run it.

More prompts for this kind of work in Get Things Done, and the reasoning behind how these prompts are built in How to Write Better AI Prompts.