Meeting Decision Extractor
15 views · 0 copies · 0 likes
Meeting summaries are close to worthless. They tell you what was discussed, which nobody needs, and bury the two things that actually matter - what was decided and who owes what. A week later the summary is unreadable, and the team relitigates a decision that was already made because nobody can find it.
The failure is that summarisation optimises for coverage. A good summary represents the whole meeting proportionally, so ten minutes of tangent gets ten minutes worth of text. But meeting value is wildly unevenly distributed. One sentence of commitment matters more than forty minutes of context.
This prompt discards summarisation entirely. It extracts four things and refuses to produce prose: decisions, commitments, open questions, and - most importantly - the distinction between a real decision and a discussion that merely sounded conclusive.
That last distinction is where most value hides. Meetings routinely end with everyone believing something was settled when no one actually said so. Someone said we should probably move to the new system, three people nodded, and nothing was assigned. That is not a decision. This prompt is instructed to classify such moments as UNRESOLVED and flag them, which is frequently uncomfortable and always useful, because an unresolved item you know about costs far less than one you discover in three weeks.
Commitments require an owner and a date. If the transcript does not contain a named person, the prompt writes OWNER UNASSIGNED rather than guessing from context. Guessing here is actively harmful - an action item attributed to the wrong person is worse than one attributed to nobody, because it will not be questioned.
What to feed it. A raw transcript works better than notes, because notes have already discarded the ambiguity you want to detect. If you have a recording with automatic transcription, the messy verbatim version is the correct input, filler words and all.
A practical habit. Run it immediately after the meeting and paste the output straight into the channel while everyone still remembers. The unresolved section becomes the agenda for the next meeting, and the flagged items get corrected by participants within minutes - which is exactly when correction is cheap.
One honest caveat. Transcription errors on names and numbers are common, so verify owners and figures before circulating anything that people will act on.
You extract structure from meeting transcripts. You do NOT write summaries, narratives, or overviews. Never describe what was discussed. INPUT Transcript: [paste the raw transcript, verbatim is better than cleaned notes] Known attendees: [names, if the transcript uses initials or is unclear] Output exactly these five sections and nothing else. 1. DECISIONS MADE - only items where someone stated a choice as settled. For each: the decision in one sentence, who stated it, and the exact quote that establishes it. If no quote establishes it, it does not belong in this section. 2. UNRESOLVED BUT TREATED AS DECIDED - discussions that sounded conclusive but where no one actually committed. For each: what people probably believe was decided, why it does not qualify, and the specific question that must be answered to settle it. This section is important - be thorough rather than polite. 3. COMMITMENTS - a table with columns: what, owner, due date, and how explicit it was rated as EXPLICIT, IMPLIED, or ASSUMED. Write OWNER UNASSIGNED if no name was stated. Write NO DATE STATED if no date was given. Never infer an owner or a date from context. 4. OPEN QUESTIONS - questions raised and not answered. Note who raised each and whether anyone was asked to resolve it. 5. CONTRADICTIONS AND RISKS - places where two participants stated incompatible things, where a commitment conflicts with another, or where a decision rests on an unverified assumption. If none, write None found. RULES Quote the transcript to support anything in section 1 or 3. No quote means no entry. Never soften an unresolved item into a decision to make the output look tidier. If the transcript is too unclear to judge, say which part and why. Do not add recommendations, next steps of your own invention, or commentary.
Recommended AI Model
This prompt works best with Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro. Other capable models will also work - compare them on our AI Models page.
Prompt Guide
How to Use
Copy the prompt below and paste it into Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro. Replace anything written in [brackets] with your own details, then send it. The more context you give about your goal, audience and tone, the better the result will be for Business work.
Step-by-step Instructions
- 1 Click the Copy button on the prompt above.
- 2 Open Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro and start a new chat.
- 3 Paste the prompt and replace every [placeholder] with your own information.
- 4 Send the prompt and read the first draft carefully.
- 5 Ask follow-up questions such as "make it shorter", "change the tone" or "give me 3 variations" until the output matches what you need.
Recommended AI Model
Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro
Tips
- Be specific: mention your audience, goal and preferred length.
- Add an example of the style you like so the AI can match it.
- Ask for the output in a table or list when you need structure.
- If the first answer is generic, ask the AI to be more concrete and to avoid filler wording.
- Save the versions that work well so you can reuse them later.
Expected Output
A ready-to-use result for "Meeting Decision Extractor" that you can refine further with follow-up messages. Expect a structured, well-written answer that you can copy straight into your own workflow.
Related Prompts
Pricing Experiment Designer
Pricing decisions get made in the worst possible way. Someone looks at three competitors, picks a number in the middle,
Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro
Customer Interview Synthesizer
Customer interviews get summarised into themes, and themes are where the insight dies. Six conversations become four bul
Comments (0)
No comments yet. Be the first to share your results.