---
name: meeting-debrief
description: Produce a high-signal debrief of a meeting from a transcript. Use whenever someone uploads or points at a meeting transcript and asks for a debrief, recap, review, analysis, readout, or post-meeting notes. Also trigger on "what did I miss", "read between the lines", "what's really going on here", "debrief this call", or when someone shares a recording or transcript and asks what to do next. Works for any meeting type: sales calls at any funnel stage, client working sessions, partnership and cofounder exploratory calls, vendor pitches, negotiation rounds, internal strategy meetings. Auto-detects the meeting type and produces a one-page debrief whose weight sits on subtext, omissions, and concrete next actions rather than on re-summarising what was said.
---

# Meeting Debrief

Your notetaker already told you what was said. This tells you what happened.

The deliverable is a one-page markdown file. The chat reply is a short pointer to that file, never the file itself.

## What this does not do

It does not summarise the meeting. Gemini, Otter, Fathom, Fireflies, Zoom AI Companion and Copilot all produce a competent summary and a competent action list, and they are free with the call. Reproducing that work is the main way this skill fails.

This is the layer on top: what was revealed, what was avoided, and what to do about it.

## Inputs

Two files, ideally:

1. **The AI-generated summary or meeting notes** (Gemini "Notes by Gemini", Otter, Fathom, Fireflies, Zoom AI Companion, Teams Copilot, or a human's notes)
2. **The verbatim transcript** with speaker attribution and timestamps

The verbatim is where the signal lives. The summary is a cross-check: use it to orient in thirty seconds, never to source the subtext. Flattened summaries delete exactly the things this skill hunts for, so a debrief built only on a summary is a guess wearing a suit.

If only one file exists, work with it and say so inline. If it is only the summary, cap every subtext claim one tier lower than the evidence seems to justify and say why.

## The disclaimer that actually matters

**This skill gets sharply better when it runs against a knowledge base, and it is thin without one.**

Subtext is a claim about a *deviation*: someone answered differently than they usually answer, skipped a topic they always raise, or conceded a point they have defended for a year. A transcript on its own has no baseline, so the model can only compare a person against a generic idea of how people behave. That produces plausible, confident, unfalsifiable observations. It is the failure mode to fear here.

Give it a baseline and the same passage yields something real. What counts as a baseline:

- Notes on the participants: their role, what they want, how they negotiate, what they said last time
- Earlier debriefs from the same relationship, so a pattern across calls becomes visible
- The commercial context: pricing discussed, scope agreed, the actual state of the deal
- Your own prior positions, so it can catch where you drifted

Practically: run this inside a project, workspace, or folder that holds your notes on these people and the history of this relationship, and point it at that material explicitly. Cold, on a single transcript, treat every subtext line as a prompt for your own thinking rather than a finding.

## Workflow

### 1. Read everything end to end

Do not skim. The subtext layer lives in micro-moments: self-corrections, interruptions, filler clusters, flattery placement, topics parked and never picked back up. Skimming finds the plot and misses all of it.

Then check the knowledge base for anything on these people and this relationship before writing a line.

### 2. Classify the meeting

From first principles. Common types and what "success" means in each:

| Type | Primary target |
|---|---|
| Sales, first contact or discovery | Qualify, surface pain, establish credibility, earn the next call |
| Sales, mid-funnel | Deepen understanding, differentiate, handle objections, move to proposal |
| Sales, negotiation or close | Align terms, manage concessions, secure commitment |
| Client working session | Align on deliverables, unblock, reset expectations |
| Partnership or cofounder exploratory | Assess mutual fit, power gradient, who needs whom more |
| Informational or discovery, non-sales | Learn the domain, build the mental model, find the open questions |
| Vendor pitch, you as buyer | Assess capability, fit, risk, pricing posture |
| Internal strategy or decision | Surface options, make the call, assign owners |

If it does not fit cleanly, describe it in plain language instead of forcing a label. Meetings shift mode mid-call, discovery becoming scoping is the usual one. Flag it when it happens, because the mode it ended in is the one that governs the follow-up.

### 3. Write the file

Target 450 to 650 words total. One page. Be ruthless.

```markdown
# Meeting Debrief: [Counterparty / Topic]
**Date:** [YYYY-MM-DD] · **Type:** [detected type] · **Participants:** [names, roles]
**Sources:** [which files] · **Baseline:** [what knowledge base material was available, or "none"]

## Bottom line
One or two sentences. What actually happened and what it means for you.

## State change
What is now true that was not true before the call. Max 5 bullets, one line each, facts only.
Decisions, commitments, numbers, dates, names. If nothing changed, say that, because
"a pleasant call that changed nothing" is itself the finding.

## Reading between the lines
The main event. Max 5 items, highest conviction first, each one:

- 🟢/🟡/🔴 **[Label]**: what was observed, in one sentence, then what it implies, in one sentence.
  *Evidence:* the moment it rests on, quoted in under 15 words, with the timestamp if there is one.

Every item must name the moment it comes from. An observation without an anchor is a horoscope.
No item may rest only on tone unless the tone shift is specific and locatable.

## What was not said
Second main event, and usually the more valuable of the two. Max 5 items:

- 🟢/🟡/🔴 **[Label]**: what is missing, and why the absence matters.

Three kinds of silence, and they are not equally interesting:
1. **Steered away from.** A topic raised and visibly moved past. Highest value, name who steered.
2. **Structurally absent.** What any competent counterparty in this role would normally raise and did not.
   Sales: budget, timeline, decision process, other stakeholders, competing options.
   Partnership: commercials, equity, exclusivity, what happens if it fails.
   Client session: any pushback at all. A completely smooth call is a signal, not a success.
   Vendor pitch: references, failure cases, what the product does badly.
3. **Parked and never returned to.** "Let's come back to that" and nobody did. Cheap to spot, often the
   single most useful line in the debrief, and it is the thing summary tools reliably delete.

## Tone and dynamics
One or two sentences, and only if there is a real signal: a power gradient, unusual hesitation,
who set the pace, who deferred. Otherwise cut the section entirely. Do not pad it.

## Your next actions
Concrete deliverables only, each one something you can put a checkmark behind.
Sorted by urgency × importance, then by dependency order.
- Send [person] the [specific thing] covering [specific content]
- Put [specific question] to [person] before [date]
- Draft [specific artifact] and share it by [date]

Not "research X", not "think about Y", not "consider Z". Those are intentions, not actions.

## Open questions
Max 3. Things you now need an answer to and cannot get from the transcript. One line each.
```

Everything above the two inference sections should be compressed to the bone. If "State change" is longer than "What was not said", the debrief has drifted back into being a summary.

### 4. Conviction tiers, and use them honestly

- 🟢 **Strong**: direct textual evidence, more than one corroborating moment, near certain
- 🟡 **Moderate**: inferential, a pattern across the call, could be wrong and worth weighing
- 🔴 **Speculative**: one weak signal, flagged for awareness, do not act on it alone

Default to under-claiming. One thread of evidence means 🔴 no matter how good the insight feels. Marking everything 🔴 to look careful is its own failure: when the evidence is there, call it 🟢.

Without a knowledge base, cap subtext at 🟡. There is no baseline to be certain against.

### 5. Save it, then keep the chat reply short

Save as `meeting-debrief-YYYY-MM-DD-[counterparty-slug].md`, lowercase and hyphenated, in wherever meeting notes already live. If there is no such place, an `outputs/` folder in the working directory.

The chat reply is under 60 words: the bottom line, the file path, and at most one of either the top next action or the single sharpest observation. Never reproduce the sections in chat. If the chat reply reads like the file, it failed.

## Signal catalog

The hunting list for the verbatim. Not exhaustive, and not a checklist to run mechanically.

**Hesitation and discomfort**
- Filler clustering ("um, well, I mean") around one specific topic and nowhere else
- A topic change initiated by the speaker immediately after a question
- Vague answers to specific questions: "we're thinking about it", "it's on the roadmap"
- Self-corrections and hedges: "actually, that's not quite true"
- Long pauses, often visible as connectivity references or a stray "sorry, go ahead"

**Genuine interest**
- The same idea repeated in slightly different words, which means they are sold on it
- Sentences getting shorter and faster
- Unprompted elaboration well past what was asked
- Future-tense visioning: "where we want to take this is"
- Volunteering information nobody requested

**Power and status**
- Who interrupts whom, and whether the interrupted party defers or pushes back
- Who sets the agenda and controls pacing
- Flattery: who praises whom, how often, and at which point in the call it arrives
- Who cites external constraints (other meetings, conferences, a board) and when
- Whose questions get full answers and whose get deflected

**Language that marks a position shift**
- Pronoun drift: "we" becoming "they" about their own company, or "I" becoming "we" about a decision
- Modal softening across the call: "we will" becoming "we'd like to" becoming "we might"
- Timeframes quietly stretching: "next month" becoming "in the next quarter or so"
- Conditionals appearing on something previously stated flat

**Parked topics**
- Points you raised that were acknowledged and never engaged with
- Their own concerns mentioned once in passing and moved past
- Limitations admitted and not resolved
- Every "let's come back to this" that nobody came back to

## Failure modes

- **Re-summarising.** The notetaker did that. If the file could have been written from the summary alone, it is worthless.
- **Dumping the debrief into chat.** The file is the deliverable, chat is the pointer.
- **Confident subtext with no baseline.** The single worst outcome, because it is indistinguishable from insight while you read it. Cold transcripts get 🟡 at best.
- **Horoscope observations.** Anything that would sound true of any meeting between any two people. Cut it.
- **Evidence-free claims.** Every subtext item names the moment it comes from or it does not ship.
- **Listing everything.** If an observation is not actionable or leverage-generating, it is noise.
- **Vague actions.** "Think about the pricing" is not an action.
- **Bloat.** Over one page means a prioritisation failure, not a length problem.
- **Flattening the counterparty.** Use their actual first name. "The other party" produces generic reads.

## Style

Precision over narrative. No preamble, no flattery, no narrating the process. Quote briefly when evidence needs showing, under 15 words. If a section has nothing in it, write "Nothing notable" rather than padding it out.

---

*Install: save as `meeting-debrief.md` inside a folder named `meeting-debrief` in your skills directory, or drop the file into a Claude project and ask for a debrief. Point it at your notes on the people in the call.*