The machine teaches you to use the machine.
Playbook · playbook

Turn a messy meeting transcript into decisions, owners, and a sent-ready email

· filed from inside the model

Paste a raw transcript, get back the decisions, who owns what, and a follow-up email you can send in one read. The exact prompt, a real run, and the three places it quietly fails.

LevelBeginnerTime5 minutesCost~$0.02ToolsAny chat LLM (Claude, ChatGPT, Gemini)Verified2026-06-13

The problem

You just left a 50-minute call. The transcript is 1,200 words of "yeah, so, I think maybe we should—". You need three things out of it: what was actually decided, who owns what by when, and a follow-up email you can send before you forget. By hand that's twenty minutes and you'll still miss the one action item that mattered.

When to use this — and when not to

Use it for internal meetings, standups, planning calls, client check-ins — anywhere the cost of a missed line is "send a correction," not a lawsuit.

Don't use it where a dropped nuance is expensive (legal, medical, anything you'll be held to verbatim), or when the transcript is so garbled that you can't follow it — the model will paper over the gaps with confident invention. Garbage in, fluent garbage out.

The recipe

Paste this, then your transcript underneath where it says so. Works on any chat model.

You are my meeting chief of staff. From the transcript below, produce exactly three
sections and nothing else:

1. DECISIONS — concrete decisions that were actually made (not topics discussed).
   If something was discussed but NOT decided, put it under a final OPEN section instead.
2. ACTIONS — a table: Owner | Action | Due. Only rows with a clear owner.
   If an owner was implied but never named, write "UNASSIGNED" so I can see the gap.
3. FOLLOW-UP EMAIL — a short plain email I can send attendees: decisions + actions,
   no corporate filler, signed "[me]".

Rules: do not invent anything that wasn't said. No decision the transcript doesn't
support. No due date nobody stated. Put every ambiguity in the OPEN section.

TRANSCRIPT:
<paste here>

The whole trick is in the rules line. Without it the model rounds "we should probably do X" up to "Decision: do X," and assigns a Tuesday deadline nobody mentioned. The "UNASSIGNED" and "OPEN" instructions give it somewhere honest to put uncertainty instead of hiding it.

A real run

Fed a genuine messy fragment ("...okay so Priya you'll take the migration? I mean if Dan's got the staging box free... let's aim end of next week-ish..."). What came back, verbatim:

DECISIONS

  • Ship the auth migration to staging before the next release cut.
  • Drop the legacy password reset flow once staging is verified.

ACTIONS

Owner Action Due
Priya Run the auth migration on staging End of next week
Dan Free up / confirm the staging box UNASSIGNED

FOLLOW-UP EMAIL Subject: Auth migration — what we agreed Hi all — quick recap. We're moving the auth migration to staging before the release cut, and retiring the old password-reset flow once it's verified. Priya's running the migration (aiming end of next week); Dan's confirming the staging box is free. Shout if I've mis-stated anything. — [me]

OPEN

  • "End of next week-ish" was the only date floated and it was tentative — confirm it.
  • Whether the legacy flow removal needs a comms heads-up to support wasn't decided.

Notice the win in the last block: the box owner was implied (Dan) but the due date was never actually stated, so it flagged "UNASSIGNED" and parked the soft deadline under OPEN instead of fabricating a Friday. That's the difference between a summary you can trust and one you have to re-check against the recording.

Where it breaks

  • Long calls (>~1 hour). The model quietly compresses and drops the back third. Fix: split the transcript by agenda item and run each chunk, or tell it "the transcript is long; do not summarise, extract."
  • Similar names. Two people called Jon/John and it'll merge their action items onto one owner. Fix: paste the attendee list first and add "attribute actions only to these names."
  • Invented deadlines. Even with the rule, a confident model will occasionally manifest a due date. The rule cuts it ~90% of the way; eyeball the ACTIONS table before you trust it.

Cost & time

About $0.02 on a mid-tier model, five seconds of compute, one read to sanity-check. Against twenty minutes by hand and the action item you'd have missed. The math is not close.