Analyzing Meetings

Input Classification

Step 1: Detect Input Type

Detection Heuristics

Raw Transcript indicators:

• Speaker labels: [John]:, Speaker 1:, John Smith:
• Timestamps: [00:15:32], (15:32)
• Filler words: "um", "uh", "like", "you know"
• Incomplete sentences, interruptions marked with --

Meeting Notes indicators:

• Markdown headers: #, ##, ###
• Bullet points: -, *, •
• Action item markers: [ ], TODO:, Action:
• Structured sections: "Attendees:", "Decisions:", "Next Steps:"

Speaker Attribution Protocol

Input Assessment

Attribution Heuristics

Positional Inference:

• First speaker often sets agenda (likely meeting owner)
• Responses to questions indicate different speaker
• "I'll do X" vs "Can you do X" indicates speaker switch

Contextual Clues:

Conversation Flow:

• Question → Answer = speaker change
• Agreement ("Yes, and...") = new speaker
• Topic shift = possible new speaker

Confidence Levels

Output Formats

High confidence:

**[Name]**: "We should prioritize the API work for Q2"

Medium confidence:

**[Inferred: Engineering Lead]**: "The technical debt is blocking new features"
- *Attribution basis*: Speaker discussed technical architecture

Low confidence:

I couldn't determine who said this:
**Quote**: "We need to push back the launch date"
Who made this statement?
a) [Name 1]  b) [Name 2]  c) Someone else  d) Unknown

Critical Attribution Requirements

MUST attribute (ask if uncertain):

• Action item owners ("I'll handle X")
• Decision makers ("We've decided to...")
• Blockers/concerns raised ("I'm worried about...")
• Commitments made ("My team can deliver by...")

Name Verification Protocol

Purpose

Transcription services often misspell names. Use input/org/colleagues.json to verify.

When to Verify

Check a name against colleagues.json when:

• Name spelling looks phonetically plausible but unusual
• Name doesn't match any known colleague exactly
• Name appears in action item owner context

Verification Process

1. Extract all names mentioned
2. For each name:
   a. Check exact match in colleagues.json → Use as-is
   b. If no exact match:
      - Check commonAliases
      - Search for phonetically similar names
   c. If match found with high confidence → Auto-correct
   d. If uncertain → Flag for user verification

Correction Format

• Auto-corrected: Just use the correct name
• Uncertain: Note original: "[Corrected from: original transcription]"

For detailed name lookup protocols, see .claude/reference/name-verification.md.

Example Input

[00:15:32] Speaker 1: So I think we should move forward with the React migration.
[00:15:45] Speaker 2: I agree, but we need to consider the timeline. My team is already stretched.
[00:16:02] Speaker 1: Can you give me a realistic estimate?

Example Output

## Input Classification
**Type**: Raw Transcript
**Speaker Labels**: Partial (numbered speakers, timestamps)
**Duration**: ~1 minute segment

## Speaker Attribution
- **Speaker 1** [Inferred: PM/Lead]: Sets agenda, asks for estimates
- **Speaker 2** [Inferred: Engineering Lead]: References "my team", timeline concerns

## Names Verified
- No names mentioned directly in this segment
- "Speaker 2" likely engineering based on team reference

## Ready for Routing
- 1 potential decision: React migration
- 1 action item: Timeline estimate needed
- Attribution: Ask user to confirm speaker identities

Quality Gates

• Complete content read (beginning to end)
• Input type correctly classified
• Speaker label presence assessed
• Names verified against colleagues.json
• Attribution confidence levels assigned
• Action item owners identified or flagged

Success Criteria

1. Input type correctly identified
2. Speakers attributed with appropriate confidence
3. Names verified and corrected if needed
4. Ready for routing-brains skill