Prompts That Turn Meeting Transcripts Into Real Action Items (Not Slop)

Every transcription tool now generates summaries and action items. The output is usually generic. The real action items are not where the bullet points live.

Here are the prompts that find them.

The problem with default summaries

Default summaries optimize for "I covered everything." They include the warm-up chat, the irrelevant tangent, the consensus statement that's actually a fight.

What you want is "what changed because of this meeting." That's a different prompt.

Prompt 1: The "what actually moved" extractor

``` You are summarizing a meeting for decisions and changes, not topics covered.

Transcript: {TRANSCRIPT}

Return: 1. Decisions made (with owner and any commitment dates) — only decisions, not discussions 2. Changes in plan from before the meeting (what was assumed entering, what's true now) 3. New questions that emerged that weren't asked entering the meeting 4. Disagreements that were not resolved (with the two positions)

Do NOT include: - Topics discussed without conclusion - Pleasantries or chitchat - Statements of intent without specific commitments

If a section is empty (no decisions made, etc.), say "none" rather than padding. ```

The "none" instruction is critical. Default LLM behavior is to find something for every section, which creates fake action items.

Prompt 2: The hidden-commitment scanner

``` Scan this transcript for soft commitments that the speaker may not realize they made.

Transcript: {TRANSCRIPT}

A soft commitment is a phrase like: - "I'll get back to you on that" - "Let me look into that" - "I'll send you that doc" - "I'll talk to [name] about it" - "I should be able to do that by [vague time]"

Return each soft commitment with: - Who said it - What they implicitly committed to - The vagueness of the timing ("by Friday" is specific, "soon" is not) - A suggested explicit follow-up to convert it to a hard commitment

Skip if no soft commitments are present. ```

This is the most valuable prompt I have. Soft commitments are where real work falls through cracks.

Prompt 3: The "what didn't come up" prompt

``` Compare the meeting agenda to the transcript.

Intended agenda: {AGENDA} Transcript: {TRANSCRIPT}

Identify: 1. Agenda items that did not get discussed (with reason if apparent) 2. Topics that came up that were not on the agenda 3. Whether the meeting time was used proportional to the agenda's stated priorities

Output as a 3-bullet summary, one per category. If a category is empty, say "all agenda items covered" or "no off-agenda topics." ```

For recurring meetings (weekly all-hands, partner meetings), this is the audit you'd never do manually.

Prompt 4: The relationship-status read

For sales / client meetings specifically.

``` Read this meeting transcript as a relationship-status indicator.

Transcript: {TRANSCRIPT} Context: {RELATIONSHIP_CONTEXT}

Score 1-10 each: - Engagement level (are they leaning in or coasting?) - Trust signal (are they sharing real internal info or staying surface?) - Urgency signal (are they treating this as time-pressured or not?) - Decision-power signal (is the right person in the room?)

For each score under 7, give one specific quote from the transcript that supports the score.

End with one sentence: "Recommended next step: [advance / deepen / put on hold / disqualify]." ```

The quote requirement keeps the AI grounded. Numbers without evidence are theater.

Prompt 5: The async update generator

For meetings where stakeholders weren't in the room.

``` Generate an async update for a stakeholder who wasn't in this meeting.

Transcript: {TRANSCRIPT} Stakeholder role: {ROLE} What they care about most: {THEIR_PRIORITIES}

Produce a 5-bullet update: - The single most important thing they need to know - Any decision that affects their work - Any new dependency on them - The next thing they should do - The next thing happening that they should be aware of

Skip anything that doesn't affect them. Don't pad to fill bullets. ```

The "skip anything that doesn't affect them" instruction prevents async updates from becoming meeting recaps. Different output.

My workflow

Every meeting I record gets: - Whisper transcript (or Otter / Fireflies / Read.ai) - Prompt 1 run automatically → decisions and changes - Prompt 2 run automatically → soft commitments - Prompt 3 run when there was an agenda → coverage check - Prompts 4 and 5 run on demand → relationship and async outputs

The first three are automated. They land in my Notion within 5 minutes of the meeting ending.

The last two are on-demand when I need them.

What changed when I started doing this

My weekly hours dropped 6-8. Mostly because soft commitments stopped falling through cracks (so re-do work disappeared) and because async updates stopped requiring a 20-minute write per stakeholder.

The transcription tool is the easy part. The prompts are the part most people skip and the part that earns the time back.

What I'd build first

Prompt 2 (soft commitments). It pays for itself in week one. Set it up to run on every meeting. Pipe the output to your task manager.

Everything else is incremental. The soft-commitment scanner is the keystone.