Conversation Analysis

Signal extraction → pattern detection → behavioral insights.

<when_to_use>

• User requests conversation analysis
• Identifying frustration, success, or workflow patterns
• Extracting user preferences and requirements
• Understanding task evolution and iterations

NOT for: real-time monitoring, content generation, single message analysis </when_to_use>

<signal_taxonomy>

Type	Subtype	Indicators
Success	Explicit Praise	"Perfect!", "Exactly what I needed", exclamation marks
Success	Continuation	"Now do the same for...", building on prior work
Success	Adoption	User implements suggestion without modification
Success	Acceptance	"Looks good", "Ship it", "Merge this"
Frustration	Correction	"No, I meant...", "That's wrong", "Do X instead"
Frustration	Reversion	User undoes agent changes, "Go back"
Frustration	Repetition	Same request 2+ times, escalating specificity
Frustration	Explicit	"This isn't working", "Why did you...", accusatory tone
Workflow	Sequence	"First...", "Then...", "Finally...", numbered lists
Workflow	Transition	"Now that X is done, let's Y", phase changes
Workflow	Tool Chain	Recurring tool usage patterns (Read → Edit → Bash)
Workflow	Context Switch	Abrupt topic changes, no transition language
Request	Prohibition	"Don't use X", "Never do Y", "Avoid Z"
Request	Requirement	"Always check...", "Make sure to...", "You must..."
Request	Preference	"I prefer...", "It's better to...", comparative language
Request	Conditional	"If X then Y", "When A, do B", situational rules

Confidence levels:

• High (0.8–1.0): Explicit keywords match taxonomy, no ambiguity, strong context
• Medium (0.5–0.79): Implicit signal, partial context, minor ambiguity
• Low (0.2–0.49): Ambiguous language, weak context, borderline classification </signal_taxonomy>

<phases> Track with TodoWrite. Phases advance only, never regress.

Phase	Trigger	activeForm
Parse Input	Session start	"Parsing input"
Extract Signals	Scope validated	"Extracting signals"
Detect Patterns	Signals extracted	"Detecting patterns"
Synthesize Report	Patterns detected	"Synthesizing report"

TodoWrite format:

- Parse Input { scope description }
- Extract Signals { from N messages }
- Detect Patterns { category focus }
- Synthesize Report { output format }

Edge cases:

• Small scope (<5 messages): Skip Extract Signals, jump to Synthesize
• Re-analysis: Resume at Detect Patterns
• Narrow focus (single signal type): Skip Detect Patterns

Workflow:

• Start: Create Parse Input in_progress
• Transition: Mark current completed, add next in_progress
• After delivery: Mark Synthesize Report completed </phases>

<workflow> 1. Define Scope - Message range (all, recent N, date range) - Actors (user only, agent only, both) - Exclusions (system messages, tool outputs, code blocks) - Mark Parse Input `completed`, create Extract Signals `in_progress`

2. Extract Signals

   • Scan messages for signal keywords
   • Match against taxonomy
   • Assign confidence (high/medium/low)
   • Record: type, subtype, message_id, timestamp, quote, context
   • Mark Extract Signals completed, create Detect Patterns in_progress

3. Detect Patterns

   • Group signals by type/subtype
   • Find clusters (3+ related signals)
   • Identify evolution (signal changes over time)
   • Track repetition (recurring themes)
   • Spot correlations (tool chains, workflows)
   • Mark Detect Patterns completed, create Synthesize Report in_progress

4. Output

   • Generate JSON with signals, patterns, summary
   • Include confidence, recommendations, action items
   • Append △ Caveats if gaps exist
   • Mark Synthesize Report completed </workflow>

<pattern_detection> Behavioral patterns from signal clusters:

Pattern	Detection	Confidence
Repetition	Same signal 3+ times	Strong: 5+ signals
Evolution	Signal type changes over time	Moderate: 3-4 signals
Preferences	Consistent request signals	Strong: across sessions
Tool Chains	Recurring tool sequences (5+ times)	High: frequent use
Problem Areas	Clustered frustration signals	Strong: 3+ in same topic

Temporal patterns:

• Escalation: Increasing frustration/stronger requirements
• De-escalation: Frustration → success transition
• Cyclical: Same issue recurs across sessions </pattern_detection>

<output_format> JSON structure:

{
  "analysis": {
    "scope": {
      "message_count": N,
      "date_range": "YYYY-MM-DD to YYYY-MM-DD",
      "actors": ["user", "agent"]
    },
    "signals": [
      {
        "type": "success|frustration|workflow|request",
        "subtype": "specific_subtype",
        "message_id": "msg_123",
        "timestamp": "ISO8601",
        "quote": "exact text",
        "confidence": "high|medium|low",
        "context": "brief explanation"
      }
    ],
    "patterns": [
      {
        "pattern_type": "repetition|evolution|preference|tool_chain",
        "category": "success|frustration|workflow|request",
        "description": "pattern summary",
        "occurrences": N,
        "confidence": "strong|moderate|weak",
        "first_seen": "ISO8601",
        "last_seen": "ISO8601",
        "recommendation": "actionable next step"
      }
    ],
    "summary": {
      "total_signals": N,
      "by_type": { "success": N, "frustration": N, ... },
      "key_insights": ["insight 1", "insight 2"],
      "action_items": ["item 1", "item 2"]
    }
  }
}

</output_format>

<rules> ALWAYS: - Create Parse Input at session start - Update todos at phase transitions - Include confidence levels for all signals - Support patterns with 2+ signals minimum - Mark Synthesize Report `completed` after delivery - Apply recency weighting (recent overrides old)

NEVER:

• Skip phase transitions
• Extract low-confidence signals without marking them
• Claim patterns from single occurrences
• Regress phases
• Deliver without marking final phase complete
• Over-interpret neutral language </rules>

<references> - [signal-patterns.md](references/signal-patterns.md) — extended taxonomy, edge cases - [extraction-techniques.md](references/extraction-techniques.md) — regex, heuristics - [sample-analysis.md](examples/sample-analysis.md) — complete walkthrough </references>