AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people who paste a meeting transcript into an AI ask for "a summary", get a tidy page of bullet points, and feel organised. Then the summary sits in a tab while the actual follow-ups — the email, the client note, the document someone asked for — wait until Thursday. The meeting was never the problem. The homework it creates is, and a summary does none of it.
The rule: a meeting is processed only when you have the decisions, the owned actions, the open questions and a first draft of the quickest follow-ups in hand. Anything less is summary theatre.
The mechanism below is a standing Claude Project with fixed instructions, so every transcript you drop in comes back in the same five-part shape. Call it the Five-Part Meeting Return.
A Project in Claude is a separate workspace with its own instructions, its own uploaded reference files ("project knowledge") and its own chat history. The instructions you set apply to every chat started inside it, so you write the brief once rather than re-pasting it after every call.
Two details matter for this use:
| Step | What you do | Where |
|---|---|---|
| 1 | Open Projects | Left sidebar, "Projects", or go to claude.ai/projects |
| 2 | Click "+ New Project" | Top right of the Projects page |
| 3 | Name it, e.g. "Meeting homework" | Claude does not read the name or description, so they are for you |
| 4 | Click "Set project instructions" and paste the instructions below | Project page |
| 5 | Optionally add reference files with the "+" button | Project knowledge panel |
| 6 | For each meeting, start a new chat in the project and paste or upload the transcript | Project page |
Useful reference files for step 5: a who's-who list (names, roles, which client), your email sign-off, and any notes template the team expects. Project files are capped at 30MB each at time of writing, far more than these need.
This is the core of the playbook. Paste it into "Set project instructions" and fill in the two placeholders once.
ROLE
You are a meeting follow-up assistant for [YOUR_NAME], who works as [YOUR_ROLE] at [YOUR_ORGANISATION_OR_"freelance"].
CONTEXT
In each chat I will give you one meeting transcript or set of rough notes. Treat that text as the only source of facts about the meeting. Project knowledge files may tell you who people are and how I write; use them for names and tone, never as evidence of what was agreed.
GOAL
Turn the meeting into work I can act on today. A good result lets me close the chat and send something within ten minutes, without rereading the transcript.
STEPS
1. Read the whole transcript before writing anything.
2. If the text is too fragmentary to tell who said what, or appears to be missing a section, say so first and ask me to confirm before continuing.
3. Produce the five sections below, in this order, with these headings.
OUTPUT
A. Decisions — one line per item that was clearly agreed. If something was discussed but not settled, put it under C, not here. If nothing was decided, write "No firm decisions recorded."
B. Actions — a table with columns: Task | Owner | Due | What finished looks like. Write "Not stated" for any owner or date the transcript does not give.
C. Unresolved — questions or disagreements left open, and who would need to answer each one, if that is known.
D. First drafts — choose the two or three actions that are quickest to start and write the actual first version of each (the email, the message, the outline). Label each draft with the action it belongs to. Do not describe the task; do it.
E. Do this first — one line naming the most time-sensitive item, with a one-sentence reason drawn from the transcript.
CONSTRAINTS
- Do not invent decisions, names, figures, dates or commitments. If a detail is ambiguous, mark it [UNCLEAR] and quote the words that caused the doubt.
- Keep everything specific to this meeting. No general productivity advice.
- Write in British English. Drafts should match the tone in [STYLE_FILE_NAME_OR_"plain and polite"].
- Do not send, schedule or share anything; you only draft.
CHECK BEFORE ANSWERING
Confirm silently that every decision, owner and date in your answer can be traced to the transcript. Remove or mark [UNCLEAR] anything that cannot.
Fill in: your name, role and organisation, plus the name of a style file in project knowledge (or leave "plain and polite").
For a single meeting, the chat message itself can be short:
Meeting: [MEETING_NAME], [DATE], attendees [ATTENDEE_NAMES_AND_ROLES].
My side of the follow-up is [WHAT_YOU_PERSONALLY_OWE_OR_"unsure"].
Transcript below. If the attendee list does not match the speakers in the transcript, ask me before you assign owners.
[PASTE_TRANSCRIPT]
Fill in: meeting name, date, who was there, what you think you owe, and the transcript.
Picture a freelance designer, Priya, on a 40-minute call with a small bakery chain about a menu redesign. She exports the transcript from her call software and pastes it into a new chat in the project.
Priya edits two sentences and sends the email. Nobody misses the summary.
The trap is trusting the confident tone of section B. Transcription software mishears names and numbers, and speaker labels are often wrong in group calls. A model handed a transcript where "Sam" is labelled as "Speaker 3" may assign Sam's task to the wrong person, and it will look perfectly tidy. The instructions force "Not stated" and [UNCLEAR] markers to reduce this, but they cannot fix a bad transcript. Read the Actions table against your own memory of the call before anything leaves your outbox.
The second trap is filling project knowledge with old meeting transcripts "for context". The instructions tell Claude not to treat them as evidence, but stale decisions sitting next to fresh ones is how an old deadline ends up in a new email. Keep project knowledge to stable reference material.
It will not chase anyone or know anything said off-transcript, such as the thing agreed in the corridor afterwards. It drafts; it does not send. And it cannot tell a firm decision from a polite "sounds good" any better than the words on the page allow.