Most "AI meeting notes" are 4,000-word transcripts that nobody reads. The useful pattern is different: a structured summary that captures decisions, action items, open questions, and tone — in a format that becomes the canonical record of what happened. Here's exactly how to set it up.
Use whatever transcription source you already have — Gong, Chorus, Otter, Fireflies, Granola, native Zoom/Google Meet transcripts. The specific tool matters less than the discipline of always having a transcript available.
For internal meetings where no transcription tool is running, just record audio on your phone or use a quick browser-based transcriber. The transcript is the input. Everything else is downstream of having it.
One-time setup. Create a dedicated Claude Project ("Meeting Summaries") and load:
— Your standard meeting summary template (sections: Context, Decisions Made, Action Items, Open Questions, Tone & Risk Signals)
— 3–5 examples of summaries that worked well
— Your team's terminology guide (project names, internal acronyms)
System prompt should instruct Claude to: prioritize decisions over discussion, attribute action items to specific people, flag anything that sounded like disagreement or ambiguity, and never invent attendees or details.
A useful prompt template:
Summarize the attached transcript using our standard structure. Focus on decisions and action items. Attribute each action to a person with an owner and a target date (use [TBD] if no date discussed). Flag any disagreement or ambiguity. Maximum 600 words.
For a 60-minute meeting, this produces a 400–600 word structured summary in about 15 seconds.
Read once. Verify the decisions section matches what you actually agreed. Check action items have the right owners. Edit anything that sounds wrong.
Then ship to a known location — Slack channel, Notion page, dedicated email. Same place every time. The discipline of consistent distribution is what makes meeting summaries actually become referenceable artifacts.
The single most-skipped step. Each action item from the summary needs to end up somewhere it gets reviewed. Could be Asana, Linear, a Notion database, or just a recurring Slack reminder.
Without this step, summaries become writing exercises. With it, they become operational artifacts that compound — the team starts to trust that decisions get tracked, which changes how people engage in meetings.
1. Pasting raw transcript without structure. You get a generic narrative summary. Always include the structured template in the prompt.
2. Skipping attribution. An action item without an owner is just a wish. Force Claude to attribute every action to a specific person.
3. Not building a Project. A one-off chat works for one meeting. By meeting #10, you're re-pasting your template every time. Put it in a Project.
4. Using AI on confidential meetings without policy. Some meetings — legal, M&A, HR — should not go through external AI. Have a written policy on what types of meetings get AI-summarized.
Claude is best for the synthesis pass (turning transcript into structured summary). Granola, Otter, or Fireflies are best for the capture pass (recording + transcribing). The pattern is: capture in your transcription tool, then summarize in Claude with your template.
For most 30–60 minute meetings, 400–800 words. Long enough to capture decisions and actions, short enough that people actually read it.
Yes, but with two caveats: (1) Use Custom GPTs to persist your template — equivalent to Claude Projects. (2) ChatGPT tends to be more verbose; budget more editing time.
Either don't use external AI, or use Claude Enterprise / API with zero-retention configured. Set a written policy.