meeting-analyzer

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: meeting-analyzer
description: Analyzes meeting transcripts and recordings to surface behavioral patterns, communication anti-patterns, and actionable coaching feedback. Use this skill whenever the user uploads or points to meeting transcripts (.txt, .md, .vtt, .srt, .docx), asks about their communication habits, wants feedback on how they run meetings, requests speaking ratio analysis, mentions filler words or conflict avoidance, or wants to compare their communication across time periods. Also trigger when users mention tools like Granola, Otter, Fireflies, or Zoom transcripts. Even if the user just says "look at my meetings" or "how do I come across in meetings" — use this skill.
---

# Meeting Insights Analyzer

> Originally contributed by [maximcoding](https://github.com/maximcoding) — enhanced and integrated by the claude-skills team.

Transform meeting transcripts into concrete, evidence-backed feedback on communication patterns, leadership behaviors, and interpersonal dynamics.

## Core Workflow

### 1. Ingest & Inventory

Scan the target directory for transcript files (`.txt`, `.md`, `.vtt`, `.srt`, `.docx`, `.json`).

For each file:
- Extract meeting date from filename or content (expect `YYYY-MM-DD` prefix or embedded timestamps)
- Identify speaker labels — look for patterns like `Speaker 1:`, `[John]:`, `John Smith  00:14:32`, VTT/SRT cue formatting
- Detect the user's identity: ask if ambiguous, otherwise infer from the most frequent speaker or filename hints
- Log: filename, date, duration (from timestamps), participant count, word count

Print a brief inventory table so the user confirms scope before heavy analysis begins.

### 2. Normalize Transcripts

Different tools produce wildly different formats. Normalize everything into a common internal structure before analysis:

```
{ speaker: string, timestamp_sec: number | null, text: string }[]
```

Handling per format:
- **VTT/SRT**: Parse cue timestamps + text. Speaker labels may be inline (`<v Speaker>`) or prefixed.
- **Plain text**: Look for `Name:` or `[Name]` prefixes per line. If no speaker labels exist, warn the user that per-speaker analysis is limited.
- **Markdown**: Strip formatting, then treat as plain text.
- **DOCX**: Extract text content, then treat as plain text.
- **JSON**: Expect an array of objects with `speaker`/`text` fields (common Otter/Fireflies export).

If timestamps are missing, degrade gracefully — skip timing-dependent metrics (speaking pace, pause analysis) but still run text-based analysis.

### 3. Analyze

Run all applicable analysis modules below. Each module is independent — skip any that don't apply (e.g., skip speaking ratios if there are no speaker labels).

---

#### Module: Speaking Dynamics

Calculate per-speaker:
- **Word count & percentage** of total meeting words
- **Turn count** — how many times each person spoke
- **Average turn length** — words per uninterrupted speaking turn
- **Longest monologue** — flag turns exceeding 60 seconds or 200 words
- **Interruption detection** — a turn that starts within 2 seconds of the previous speaker's last timestamp, or mid-sentence breaks

Produce a per-meeting summary and a cross-meeting average if multiple transcripts exist.

Red flags to surface:
- User speaks > 60% in a 1:many meeting (dominating)
- User speaks < 15% in a meeting they're facilitating (disengaged or over-delegating)
- One participant never speaks (excluded voice)
- Interruption ratio > 2:1 (user interrupts others twice as often as they're interrupted)

---

#### Module: Conflict & Directness

Scan the user's speech for hedging and avoidance markers:

**Hedging language** (score per-instance, aggregate per meeting):
- Qualifiers: "maybe", "kind of", "sort of", "I guess", "potentially", "arguably"
- Permission-seeking: "if that's okay", "would it be alright if", "I don't know if this is right but"
- Deflection: "whatever you think", "up to you", "I'm flexible"
- Softeners before disagreement: "I don't wan