TL;DR
A “meeting summary” with ten bullets and no owners is a file nobody reopens. The output that survives is a short note where every action has an owner and a deadline, decisions are listed separately from actions, and hedged commitments are flagged. This page gives you a tool pick (transcription is the part AI does best), one copy-ready prompt, and a checklist to confirm the note is usable before you ship it. Do it within 24 hours of the meeting, while you can still verify what AI got wrong.
The job to be done
You sat through a 60-minute meeting and now you have a 14,000-character transcript. The default failure mode is a “summary”: ten bullets, no owners, no dates, no path to follow-up. The goal is the opposite. Strip the discussion, keep the outcomes, attach an owner and a deadline to every action item, and surface open questions so they reach the next meeting instead of disappearing.
Two parts of this are genuinely different tasks: getting accurate text out of audio (a transcription problem) and turning that text into structured outcomes (a reasoning problem). The first is best handled by a dedicated meeting tool; the second is where a general model like ChatGPT or Claude earns its keep.
Step 1: get a clean transcript with speaker labels
Action attribution falls apart without speaker labels, so the transcript step matters more than people expect. As of June 2026, these are the tools worth knowing, with prices in USD:
| Tool | Free tier | Paid entry | Best for |
|---|---|---|---|
| Otter.ai | 300 min/mo | Pro ~$8.33/mo annual (1,200 min/mo) | English meetings, live captions |
| Fireflies.ai | limited | Pro ~$10/mo annual (8,000 min storage) | Recording across Zoom/Meet/Teams |
| Notta | 120 min/mo, 3-min cap | Pro ~$8.17/mo annual (1,800 min/mo) | 58-language support, translation |
| Granola | Basic, limited history | Business $14/user/mo | Notes without a bot joining the call |
| Feishu Minutes (飞书妙记) | ~2 hrs/day free transcription | bundled with Feishu | Chinese meetings, speaker diarization |
| Apple Voice Memos + Whisper | free, local | free | Privacy-sensitive audio, offline |
A few decisions to make before you record. Otter and Notta cap monthly transcription minutes, so check the meeting length against your plan; Notta’s free tier rejects anything over a 3-minute clip, which makes it useless for real meetings until you upgrade. Granola records locally and does not send a bot into the call, which matters when a client or legal team objects to a visible “is recording” participant. If audio cannot leave your machine, run OpenAI’s Whisper locally on an exported recording.
Whichever tool you use, confirm the transcript names speakers (Speaker 1 / Speaker 2, or real names if the tool resolves them). A wall of unattributed text turns “Maya owns the migration” into “someone owns the migration.”
Step 2: feed the right context to the model
Paste the transcript into ChatGPT (GPT-5.5) or Claude (Sonnet 4.6), but include the context AI cannot infer from text alone:
- The transcript with speaker labels
- Meeting topic and purpose in one line
- Attendees with role and name, so the model can resolve “Speaker 3” to a person
- Decisions already made before the meeting, so they are not double-counted
- Your team’s convention for what counts as an action (only explicit verbs, or include implied) and what counts as a decision (passive consensus or an explicit vote)
- Where the output will live (a Notion page, a Slack thread, ticket descriptions)
Context window is rarely the bottleneck here. A 14,000-character transcript is roughly 4,000 tokens, well inside the in-app limits of ChatGPT Plus and Claude Pro. A full-day offsite with several hours of audio can run long; Claude Sonnet 4.6 and Gemini 3.1 Pro both carry a 1M-token context window as of June 2026, so paste the whole thing rather than chunking it and losing cross-section threads.
The prompt
You are an experienced meeting secretary at a tech company.
Meeting topic and purpose: [one line]
Attendees and roles: [list]
Pre-existing decisions (do not re-list): [list]
Action convention: [only explicit verbs / include implied]
Transcript:
"""
[paste]
"""
Return:
1. Meeting topic and purpose (one sentence)
2. Key decisions — each tagged with decision-maker and when it takes effect
3. Action items — each must include: action description, owner, deadline.
If owner or deadline is missing in the transcript, mark "needs confirmation."
4. Open / unresolved questions — with the person who can answer
5. Suggested agenda for the next meeting
6. A "watch out" list — anything that sounded like a commitment but was hedged
Do not restate the discussion. Outcomes only.
If an action was discussed but never agreed, list it under open questions, not actions.For a long meeting, run a second pass: “Now identify the 3 highest-leverage action items — the ones that, if dropped, would block the most other work.” This forces a priority order the raw list does not give you.
Step 3: structure and ship it
Format for one screen. A short header, decisions as a list with the decision-maker on each line, actions in a table (action / owner / due / status), open questions with an answerer, and a three-bullet next-meeting agenda. Then push it where work actually happens — Linear, Jira, Asana, Notion, or Feishu. Actions that stay in the notes die. The single highest-return habit is converting each action row into a ticket with the owner assigned, the same day.
What AI gets right and what it gets wrong
AI is reliably good at restructuring an unstructured transcript into decisions / actions / questions, and at attributing actions when the transcript names people. It is poor at three things: implied owners (“we’ll handle it”), commitments stated without a date, and the political context that decides whose action is actually load-bearing. Mark anything ambiguous as “needs confirmation” rather than letting the model guess — a confidently invented owner is worse than a blank one, because nobody questions it.
Checklist before you ship
- Every action has an owner and a deadline, or is explicitly marked “needs confirmation”
- Decisions are not duplicated as action items (they have different follow-up paths)
- The next-meeting agenda is short and specific
- The “watch out” list surfaces hedged commitments — the most commonly dropped balls
- Someone who missed the meeting can read the note and know what to do
Common mistakes
- Letting AI summarize the discussion. Nobody re-reads narrative summaries. Demand outcomes.
- Inventing owners or deadlines. Mark them “needs confirmation” instead.
- Mixing decisions and actions in one list. They follow up differently.
- Leaving actions in the doc. Push them to the task system the same day.
- Multilingual meetings without a language tag. Models sometimes drop the non-dominant language. Tell it explicitly which languages appear.
FAQ
Which tool should I start with if I don’t want to pay yet? For English meetings, Otter’s 300 free minutes per month cover a few short meetings. For Chinese meetings, Feishu Minutes gives you about two free hours of transcription per day. For privacy-sensitive audio, Whisper running locally costs nothing and never uploads.
Do I still need a general model if my transcription tool has built-in summaries? The built-in summaries are fine for a quick recap, but they rarely produce owner-plus-deadline action tables or flag hedged commitments. Paste the transcript into ChatGPT or Claude with the prompt above when the note has to drive follow-up.
What about multilingual meetings that code-switch between English and Chinese? Tell the model to keep quotes in their original language and translate only the summary lines. Without that instruction, models tend to flatten everything into the dominant language and silently drop nuance.
How soon after the meeting should I produce the note? Within 24 hours, ideally the same day. You are the verification layer for owners and deadlines AI marked “needs confirmation,” and your memory is the thing that decays.
Should AI also draft the next-meeting invitation? Yes. Have it pre-fill the agenda from the open-questions list, so unresolved items carry forward automatically instead of getting re-discovered three meetings later.
Related
- Study notes cleanup — the same approach applied to lectures
- Project plan draft — when the meeting produces a project
- Feature prioritisation — when meetings sort the backlog
- PRD draft — turning meeting outcomes into a PRD draft
- Meeting notes prompts — additional prompt variants
- AI meeting summary tutorial — the full workflow
- AI status update tutorial — turn notes into status updates
Tags: #AI writing #Prompt