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
- Balanced analysis - Evaluate BOTH human and AI behavior equally; neither party is presumed at fault
- Chronological preservation - Show conversation as it happened, turn by turn
- Behavioral focus - What happened, not blame assignment
- Contribution assessment - Quantify each party's contribution to any miscommunication
- Aggressive sanitization - Replace all specifics with placeholders
- 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
- Classify user messages - Identify type: direct-request, meta-request, mixed-content, clarification, feedback
- Scope determination - Is this analyzing THIS session or a REFERENCED session?
- Behavior mapping - Check both parties against expected patterns
- Contribution scoring - Assign percentages to understand root cause
- 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
| Pitfall |
How to Avoid |
| Focusing only on AI rule violations |
Always analyze user message clarity first |
| Analyzing quoted/pasted content as primary subject |
Identify meta-requests and scope correctly |
| Assigning 100% blame to one party |
Use contribution percentages based on evidence |
| Missing buried requests in mixed content |
Parse each message for multiple intents |
| Skipping rule loading |
MUST read all rules BEFORE analysis - see rules-checklist.md |
| Success bias (completed = good) |
Check HOW it completed, not just that it completed |
| Surface-level analysis |
Check principles (delegation, YAGNI), not just workflow steps |
Focus Area (Optional)
$ARGUMENTS
If provided, focus analysis on specific aspect (e.g., "rule compliance", "request clarity", "workflow gates").
1---2name: convo-analysis3description: Analyze conversation flows for behavioral patterns. Use this skill when debugging AI compliance issues, reviewing human request clarity, or identifying root causes of human-AI miscommunication. Produces sanitized reports safe to share.4---5
6# Conversation Analysis
7
8Analyze human-AI conversation flows to identify behavioral patterns, compliance gaps, and improvement opportunities.
9
10## Role
11
12You are a conversation analyst. Your job is to EXTRACT and ANALYZE the conversation flow - not to judge, implement, or fix anything.
13
14## Context Assessment
15
16Before starting, analyze conversation history:
17- **Empty conversation?** → Report "Nothing to analyze"
18- Start analysis from first message to this command trigger
19
20## Core Principles
21
221. **Balanced analysis** - Evaluate BOTH human and AI behavior equally; neither party is presumed at fault
232. **Chronological preservation** - Show conversation as it happened, turn by turn
243. **Behavioral focus** - What happened, not blame assignment
254. **Contribution assessment** - Quantify each party's contribution to any miscommunication
265. **Aggressive sanitization** - Replace all specifics with placeholders
276. **Rule mapping** - Check against orchestrator directives, _apply-all rules, command workflows, Holy Trinity
28
29## Analysis Process
30
31### 1. Extract
32
33Gather all messages from session start to this command:
34- User messages (requests, clarifications, approvals)
35- AI responses (reasoning, actions taken)
36- Tools used
37- Skills activated
38- Agents spawned
39- Commands invoked
40
41### 2. Analyze
42
431. **Classify user messages** - Identify type: direct-request, meta-request, mixed-content, clarification, feedback
442. **Scope determination** - Is this analyzing THIS session or a REFERENCED session?
453. **Behavior mapping** - Check both parties against expected patterns
464. **Contribution scoring** - Assign percentages to understand root cause
475. **Improvement targeting** - Identify specific fixes for user, AI, and system
48
49### 3. Output
50
51- Write report to `docs/session-reports/{YYYYMMDDHHMMSS}-<short-title>.md`
52- Display brief summary to user
53
54## Sanitization Rules
55
56Replace with placeholders:
57- File paths → `[FILE_1]`, `[FILE_2]`
58- Feature names → `[FEATURE_A]`, `[FEATURE_B]`
59- API endpoints → `[ENDPOINT_X]`
60- Variable/function names → `[CODE_REF]`
61- Business terms → `[DOMAIN_TERM]`
62- Code blocks → `[CODE_BLOCK]`
63
64Keep as-is:
65- Tool names (Read, Grep, Task, etc.)
66- Skill names (/cook, etc.)
67- Agent names (the-mechanic, etc.)
68- Generic actions (search, read, write, edit)
69
70## Analysis Checklist
71
72> See [references/rules-checklist.md](./references/rules-checklist.md)
73
74## Output Format
75
76> See [templates/report-template.md](./templates/report-template.md)
77
78## Guardrails
79
80**Holy Trinity:**
81- **YAGNI**: Only analyze - don't suggest fixes inline
82- **KISS**: Simple extraction, delegate complexity to skill
83- **DRY**: Reuse existing references and templates
84
85**Communication:**
86- Report what happened, not who's "wrong"
87- Neutral behavioral observations
88- No blame assignment
89
90**Constraints:**
91- NO code snippets in output
92- NO business logic exposure
93- NO file paths or domain-specific terms
94- Report must be shareable without editing
95- Aggressive sanitization: replace specifics with placeholders
96
97## Common Pitfalls
98
99| Pitfall | How to Avoid |
100|---------|--------------|
101| Focusing only on AI rule violations | Always analyze user message clarity first |
102| Analyzing quoted/pasted content as primary subject | Identify meta-requests and scope correctly |
103| Assigning 100% blame to one party | Use contribution percentages based on evidence |
104| Missing buried requests in mixed content | Parse each message for multiple intents |
105| **Skipping rule loading** | MUST read all rules BEFORE analysis - see rules-checklist.md |
106| **Success bias** (completed = good) | Check HOW it completed, not just that it completed |
107| **Surface-level analysis** | Check principles (delegation, YAGNI), not just workflow steps |
108
109## Focus Area (Optional)
110
111<focus>$ARGUMENTS</focus>
112
113If provided, focus analysis on specific aspect (e.g., "rule compliance", "request clarity", "workflow gates").
majiayu000