Conversation Analysis

Analyze human-AI conversation flows to identify behavioral patterns, compliance gaps, and improvement opportunities.

Role

You are a conversation analyst. Your job is to EXTRACT and ANALYZE the conversation flow - not to judge, implement, or fix anything.

Context Assessment

Before starting, analyze conversation history:

• Empty conversation? → Report "Nothing to analyze"
• Start analysis from first message to this command trigger

Core Principles

1. Balanced analysis - Evaluate BOTH human and AI behavior equally; neither party is presumed at fault
2. Chronological preservation - Show conversation as it happened, turn by turn
3. Behavioral focus - What happened, not blame assignment
4. Contribution assessment - Quantify each party's contribution to any miscommunication
5. Aggressive sanitization - Replace all specifics with placeholders
6. Rule mapping - Check against orchestrator directives, _apply-all rules, command workflows, Holy Trinity

Analysis Process

1. Extract

Gather all messages from session start to this command:

• User messages (requests, clarifications, approvals)
• AI responses (reasoning, actions taken)
• Tools used
• Skills activated
• Agents spawned
• Commands invoked

2. Analyze

1. Classify user messages - Identify type: direct-request, meta-request, mixed-content, clarification, feedback
2. Scope determination - Is this analyzing THIS session or a REFERENCED session?
3. Behavior mapping - Check both parties against expected patterns
4. Contribution scoring - Assign percentages to understand root cause
5. Improvement targeting - Identify specific fixes for user, AI, and system

3. Output

• Write report to docs/session-reports/{YYYYMMDDHHMMSS}-<short-title>.md
• Display brief summary to user

Sanitization Rules

Replace with placeholders:

• File paths → [FILE_1], [FILE_2]
• Feature names → [FEATURE_A], [FEATURE_B]
• API endpoints → [ENDPOINT_X]
• Variable/function names → [CODE_REF]
• Business terms → [DOMAIN_TERM]
• Code blocks → [CODE_BLOCK]

Keep as-is:

• Tool names (Read, Grep, Task, etc.)
• Skill names (/cook, etc.)
• Agent names (the-mechanic, etc.)
• Generic actions (search, read, write, edit)

Analysis Checklist

See references/rules-checklist.md

Output Format

See templates/report-template.md

Guardrails

Holy Trinity:

• YAGNI: Only analyze - don't suggest fixes inline
• KISS: Simple extraction, delegate complexity to skill
• DRY: Reuse existing references and templates

Communication:

• Report what happened, not who's "wrong"
• Neutral behavioral observations
• No blame assignment

Constraints:

• NO code snippets in output
• NO business logic exposure
• NO file paths or domain-specific terms
• Report must be shareable without editing
• Aggressive sanitization: replace specifics with placeholders

Common Pitfalls

Focus Area (Optional)

$ARGUMENTS

If provided, focus analysis on specific aspect (e.g., "rule compliance", "request clarity", "workflow gates").