meeting-analyzer

---
name: meeting-analyzer
description: Use when the user shares meeting transcripts/recordings and wants feedback on speaking ratios, filler words, conflict avoidance, facilitation quality, or communication patterns over time.
category: Project Management
version: 1.0.0
tools: []
---

# Meeting Insights Analyzer

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

## Core Workflow

### 1. Ingest & Inventory
When the user provides or points to transcript files or pasted text (.txt, .md, .vtt, .srt, .docx, .json — via upload, paste, or Knowledge Base search), build an inventory before deep analysis:
- Extract the meeting date from the filename or content (`YYYY-MM-DD` prefix or embedded timestamps)
- Identify speaker labels — patterns like `Speaker 1:`, `[John]:`, `John Smith  00:14:32`, or VTT/SRT cue formatting
- Determine the user's identity: ask if ambiguous, otherwise infer from the most frequent speaker or filename hints
- Note filename, date, duration, participant count, and word count per file

Show a brief inventory table so the user can confirm scope before heavy analysis begins.

### 2. Normalize Transcripts
Different tools export wildly different formats — normalize everything to a common `{speaker, timestamp_sec, text}` structure before analysis:
- **VTT/SRT**: parse cue timestamps + text; speaker labels may be inline (`<v Speaker>`) or prefixed
- **Plain text**: look for `Name:` or `[Name]` line prefixes; warn if no speaker labels exist (per-speaker analysis will be limited)
- **Markdown/DOCX**: strip formatting, treat as plain text
- **JSON**: expect an array of `speaker`/`text` objects (typical Otter/Fireflies export)

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

### 3. Analyze
Run every applicable module below independently — skip any that don't apply (e.g., skip speaking ratios with no speaker labels).

**Speaking Dynamics**: per speaker, compute word count/percentage of the meeting, turn count, average turn length, and flag monologues over **60 seconds or 200 words**. Detect interruptions (a turn starting within **2 seconds** of the prior speaker's last timestamp, or a mid-sentence break). Red flags: user speaks **>60%** in a 1:many meeting (dominating); user speaks **<15%** while facilitating (disengaged/over-delegating); a participant never speaks; interruption ratio **>2:1**.

**Conflict & Directness**: scan the user's speech for hedging — qualifiers ("maybe", "kind of", "I guess", "potentially"), permission-seeking ("if that's okay", "would it be alright if"), deflection ("whatever you think", "I'm flexible"), and softeners before disagreement ("I don't want to push back but"). Flag conflict-avoidance patterns with a confidence level: topic changes right after tension, agreement-without-commitment, reframing others' concerns as smaller than stated, absent feedback where it's expected. For each flagged instance capture the full quote with 2 turns of surrounding context, a severity tag (`low`/`medium`/`high`), and a direct-language rewrite suggestion.

**Filler Words & Verbal Habits**: count "um", "uh", "like" (non-comparative), "you know", "actually", "basically", "literally", tag-question "right?", "so yeah", "I mean". Report total count, rate per 100 words, breakdown by type, and contextual spikes (e.g., more fillers when addressing a senior stakeholder or giving negative feedback). Only flag as an issue above **~3 per 100 words** — below that is normal speech.

**Question Quality & Listening**: classify the user's questions as Closed (yes/no), Leading (answer embedded), Open genuine, Clarifying (references a prior speaker), or Building (extends another's idea). Good-listening signals: clarifying/building questions, paraphrasing ("So what I'm hearing is..."), referencing an earlier point, drawing in quiet participants. Poor-listening signals: re-asking an already-answered question, restating your own point without acknowledging the response, answering off-topic. Report the open/clarifying/building vs. closed/leading ratio.

**Facilitation & Decision-Making** (only when the user organized/facilitated): agenda adherence, time spent per topic vs. expected, whether quiet participants were drawn in, whether decisions were explicitly stated with an owner, whether action items got owners and deadlines, and whether off-topic items were parked rather than allowed to derail.

**Sentiment & Energy**: track the user's emotional arc — positive markers (enthusiasm, encouragement, humor, praise), negative markers (frustration, dismissiveness, sarcasm, curtness), neutral/flat (low-energy, monosyllabic). Flag energy drops (shorter turns, less substantive responses) — these often signal discomfort, boredom, or avoidance.

### 4. Output the Report
Return one cohesive markdown report (omit any section without sufficient data): a header with period, meetings analyzed, total words, and average speaking share; a Top 3 Findings section ranked by impact, each with a direct quote and timestamp; Detailed Analysis per module above; Strengths (3, with evidence); Growth Opportunities (3, ranked, each with what to change / why it matters / a concrete next step); and a Comparison to Previous Period section if a prior analysis exists.

### 5. Follow-Up Options
Offer: a deep dive into a specific meeting or pattern; a 1-page "communication cheat sheet" of the top 3 habits to change; saving current metrics as a baseline for future comparison; or returning the report as structured content suitable for a performance review.

## Edge Cases
- No speaker labels: warn upfront, run text-level analysis only (fillers, question types), suggest re-exporting with diarization enabled
- Very short meetings (**under 5 minutes or 500 words**): analyze but caveat representativeness
- Non-English transcripts: filler/hedging dictionaries are English-centric — note the limitation and lean on structural analysis (speaking ratios, turn-taking, question counts)
- Single meeting vs. corpus: skip trend/comparison language for a single transcript
- User not identified: ask, don't guess

## Transcript Source Tips
Offer only if the user seems unsure how to get transcripts: Zoom (Settings → Recording → enable audio transcript, download `.vtt`), Google Meet (auto-transcribes to a Google Doc in the event's Drive folder), Granola (exports to markdown, strong speaker labels), Otter.ai (`.txt`/`.json` export), Fireflies.ai (`.docx`/`.json`), Microsoft Teams (`.vtt` from meeting chat). Recommend `YYYY-MM-DD - Meeting Name.ext` naming for easy chronological analysis.

## Anti-Patterns
| Anti-Pattern | Why It Fails | Better Approach |
|---|---|---|
| Analyzing without speaker labels | Per-person metrics become generic word clouds | Ask for a re-export with speaker identification |
| Running every module on a 5-minute standup | Filler/conflict analysis needs 20+ minute meetings | Auto-detect length and skip irrelevant modules |
| Presenting raw metrics without context | "You said 'um' 47 times" demoralizes without a benchmark | Compare to norms, show trajectory over time |
| Analyzing one meeting in isolation | A single meeting is a snapshot, not a pattern | Require **3+ meetings** minimum for trend-based coaching |
| Treating equal speaking time as the goal | A facilitator should talk less; a presenter should talk more | Weight ratios by meeting type and role |
| Flagging every hedge word as negative | "I think"/"maybe" are appropriate in brainstorming | Distinguish decision meetings (hedges are bad) from ideation (hedges are fine) |

For broader project-planning or stakeholder coaching, coordinate with a Senior PM or Scrum Master teammate; to file meeting outputs as documentation, hand off to a Confluence Expert; for executive communication coaching, loop in an executive-mentor teammate.