| name | self-reflection |
| description | Analyze current session failures by examining errors, root causes, tool failures, and information quality issues. Use when user explicitly requests reflection, self-analysis, post-mortem, or session review to identify what went wrong, why it went wrong, and how to prevent future failures. |
Self-Reflection Skill
This skill guides you through a structured failure analysis of the current session to identify root causes, information quality issues, and systemic problems that need fixing.
Quick Checklist for Rapid Failure Analysis
When a failure occurs, use this checklist for immediate triage:
[ ] 1. WHAT FAILED?
- Tool error? Which tool, what input, what error?
- Wrong decision? What was chosen, what should have been chosen?
- Missed requirement? What did user want, what did you deliver?
[ ] 2. INFORMATION CATEGORY?
- Category 1 (Wrong Info): Did a skill/doc/prompt give wrong guidance?
- Category 2 (Not Used): Did information exist but you didn't use it?
- Category 3 (Missing): Did information not exist anywhere?
[ ] 3. ROOT CAUSE?
- For tool failures: Wrong input / Tool bug / Environment / Permissions?
- For decisions: Wrong info / Incomplete info / Misinterpreted / Missing?
- For requirements: Didn't read / Didn't understand / Didn't implement?
[ ] 4. PREVENTION?
- Bug to file? (Category 1 - wrong information)
- Process change? (Category 2 - workflow improvement)
- New documentation? (Category 3 - gap filling)
[ ] 5. PRIORITY?
- CRITICAL: Blocks work or causes data loss
- HIGH: Causes repeated failures or significant rework
- MEDIUM: Causes inefficiency or extra effort
- LOW: Minor improvement, nice-to-have
Use this for quick analysis, then use full templates below for detailed investigation.
Core Philosophy
Problems first, always. Success is expected; failure teaches. This reflection focuses exclusively on:
- What went wrong
- Why it went wrong
- How to prevent it from happening again
Minimize or eliminate "what went well" analysis. The user didn't ask for a pat on the back.
Reflection Framework
Conduct analysis in the following order:
1. Failure Analysis (Start Here)
Begin by cataloging every error, mistake, wrong decision, and suboptimal outcome:
Explicit Failures
- Errors that caused retry/redo: Tool failures, wrong commands, incorrect edits
- Wrong decisions: Chose wrong approach, wrong library, wrong architecture
- Missed requirements: Didn't understand or didn't implement what user wanted
- Wasted effort: Work that had to be thrown away or redone
Subtle Failures
- Inefficiencies: Could have been done faster or with fewer steps
- Poor communication: User had to clarify, repeat, or correct you
- Incomplete work: Partial implementations, missing edge cases
- Technical debt created: Quick hacks, TODOs left behind, brittle solutions
Near-Miss Failures
- Caught by user: User corrected you before you executed
- Caught by tooling: Tests, linters, or build failures caught issues
- Lucky outcomes: Solution worked despite wrong reasoning
2. Information Quality Audit (🚨 CRITICAL)
This is the most important section. Most failures trace back to information problems.
For EVERY failure identified above, analyze the information dimension using this three-category framework:
Category 1: WRONG/OUTDATED INFORMATION (🐛 Bug - Must Fix)
Information in skills, agent prompts, or documentation that is:
- Factually incorrect: Wrong library names, wrong API signatures, wrong commands
- Outdated: References old versions, deprecated approaches, obsolete tools
- Contradictory: Different sources give conflicting guidance
- Misleading: Technically true but practically wrong (e.g., "use X" when Y is better)
Action Required: File bugs to fix the source material. These are defects.
Example Questions:
- Did a skill tell you to use a deprecated library?
- Did documentation reference a command that doesn't exist?
- Did an agent prompt suggest an approach that's no longer best practice?
- Did two skills give contradictory advice?
Example from Real Session:
FAILURE: Used pdf-lib when pdf-parse was better suited
ROOT CAUSE: pdf-editor skill recommended pdf-lib for all PDF operations
INFORMATION ISSUE: Wrong information - skill should distinguish reading vs editing
BUG TO FILE: Update pdf-editor skill to recommend pdf-parse for read-only operations
PRIORITY: High - causes wrong library choices
Category 2: INFORMATION NOT USED (⚙️ Workflow Issue)
Information that existed and was available but you didn't use it:
- Didn't check available context: Had a skill loaded but didn't follow it
- Didn't read documentation: Documentation existed but you didn't read it
- Didn't ask clarifying questions: Could have asked but assumed instead
- Ignored warnings: Tools or users warned you but you proceeded anyway
Action Required: Process improvement. Why didn't you use available information?
Example Questions:
- Was there a skill loaded that you should have consulted but didn't?
- Did you skip reading a file that would have answered your question?
- Did you make assumptions when you could have asked?
- Did error messages contain hints you ignored?
Example from Real Session:
FAILURE: Tried to use git commit --amend when commit was already pushed
ROOT CAUSE: Didn't check git status to see if commit was pushed
INFORMATION ISSUE: Workflow - git Safety Protocol was in context but not followed
PROCESS IMPROVEMENT: Always run git status before deciding to amend
PATTERN: Rushing through git operations without checking state first
Category 3: INFORMATION MISSING (❓ Gap - Decide if Fix Needed)
Information that doesn't exist anywhere:
- Missing skills: No skill covers this domain
- Missing documentation: No guide exists for this tool/workflow
- Missing examples: Concept explained but no concrete examples
- Missing context: Information exists somewhere but not accessible to you
Action Required: Evaluate if gap should be filled. Not every gap needs fixing.
Decision Criteria:
- Frequency: Is this a common task?
- Difficulty: Is it hard to figure out without guidance?
- Risk: Do mistakes here cause significant problems?
- Cost: Is it worth maintaining documentation?
Example Questions:
- Should there be a skill for this workflow?
- Should this tool have a reference guide?
- Should this common pattern have examples?
- Should this context be injected into agent prompts?
Example from Real Session:
FAILURE: Didn't know how to handle large file uploads efficiently
ROOT CAUSE: No guidance on file upload strategies
INFORMATION ISSUE: Gap - no skill exists for file upload patterns
DECISION: Create skill? YES
- Frequency: High (many apps need file uploads)
- Difficulty: High (streaming, chunking, validation are complex)
- Risk: High (memory issues, security vulnerabilities)
- Cost: Moderate (one-time skill creation, low maintenance)
ACTION: Create file-upload-patterns skill with streaming/chunking examples
Information Usage Analysis (Meta-Level)
For Category 2 issues (information not used), dig deeper:
Timing: When did the information become relevant?
- Was it available from the start but you didn't check?
- Did it become available partway through (skill loaded late)?
- Was it always accessible but you forgot about it?
Discoverability: Could you have found it?
- Was it obvious where to look?
- Was it buried in a long document?
- Was it under an unexpected name?
Applicability: How obvious was the connection?
- Was it clearly relevant to your task?
- Did it require inference to connect it to your problem?
- Was the guidance too abstract to apply directly?
Example Analysis:
FAILURE: Didn't run bd sync before saying "done"
INFORMATION: Session Close Protocol in beads context clearly states this
TIMING: Available from session start
DISCOVERABILITY: In beads context, clearly marked "CRITICAL"
APPLICABILITY: Directly relevant - completing work requires sync
WHY NOT USED: Rushed to completion, skipped checklist
PROCESS FIX: Always read session close protocol when user says "done"
3. Root Cause Analysis
For each failure, go beyond surface symptoms to find true root causes.
Root Cause Analysis Templates
Internal use: These templates guide your analysis process. For full templates, see references/templates.md.
Quick template summaries (use full templates for detailed investigation):
Tool Failure (quick)
- What tool + input?
- Expected vs actual?
- Why? (wrong input / tool bug / environment / misunderstood)
- Info category? (wrong info → bug / not used → workflow / missing → gap)
- Prevention?
Wrong Decision (quick)
- What decided + when?
- What info led to this?
- Info was: wrong / incomplete / misinterpreted / missing?
- Correct decision?
- Red flags ignored?
- Prevention?
Information Quality Issue (quick)
- What's wrong + where (file:line)?
- What should it say?
- Impact + priority?
- Bug report ready?
4. Tool and Skill Performance Analysis
Evaluate effectiveness of tools and skills loaded:
Skills Loaded
For each skill: Relevance (✅/⚠️/❌) | Accuracy (✅/⚠️/❌) | Completeness (✅/⚠️/❌) | Usage (✅/⚠️/❌)
- ❌ issues = file bugs or note gaps
Documents Fetched
For each document: Usefulness (✅/⚠️/❌) | Accuracy (✅/⚠️/❌) | Right source? (✅/⚠️/❌)
Tool Usage Patterns
Analyze tool usage for inefficiencies:
Redundant Operations:
- Did you read the same file multiple times?
- Did you run the same command repeatedly?
- Did you search for the same thing twice?
Sequential Operations That Could Be Parallel:
- Did you read files one at a time when you could read them in parallel?
- Did you run independent commands sequentially?
- Did you create issues one by one instead of in parallel?
Wrong Tool Choices:
- Did you use bash when a specialized tool exists?
- Did you use task tool when direct tool would be faster?
- Did you use grep when you knew the file path?
Tool Failures:
- List every tool failure
- For each: root cause analysis (see framework above)
- For each: how to prevent in future
5. Process and Workflow Failures
Analyze the workflow for problems:
Planning Failures
- No plan: Started work without planning
- Wrong plan: Plan was logically flawed
- Incomplete plan: Missed requirements or edge cases
- Ignored plan: Made plan but didn't follow it
Execution Failures
- Wrong order: Did steps in inefficient order
- Missed parallelization: Could have done things concurrently
- Unnecessary work: Did work that wasn't needed
- Incomplete work: Left TODOs or partial implementations
Communication Failures
- Assumptions instead of questions: Made assumptions when could have asked
- Unclear explanations: User had to ask for clarification
- Missing status updates: User didn't know what you were doing
- Wrong tone: Too verbose, too terse, or inappropriate for context
6. Recommendations (Actionable Only)
Every recommendation must be:
- Specific: Not "improve docs" but "add section 3.2 with git amend safety rules"
- Actionable: Someone can implement it immediately
- Prioritized: Critical > High > Medium > Low based on impact and frequency
Category 1 Fixes: Wrong Information (CRITICAL - File Bugs)
Format bug reports:
BUG: [issue] | LOCATION: [path:line] | CURRENT: [text] | SHOULD BE: [text] | PRIORITY: [level]
Category 2 Fixes: Workflow Improvements
Format workflow fixes:
WORKFLOW: [issue] | ROOT CAUSE: [why] | FIX: [process change] | TRIGGER: [when]
Category 3 Fixes: Gap Filling (Evaluate Each)
Format gap analysis:
GAP: [missing info] | FREQ/DIFFICULTY/RISK: [H/M/L] | RECOMMENDATION: [action]
Output Format
Structure your reflection as a compact issue list:
# Session Issues
**Session**: [brief task description]
**Outcome**: [success/partial/failure]
| # | Sev | Type | Location | Issue | Action |
|---|-----|------|----------|-------|--------|
| 1 | CRIT | Wrong Info | path/file.md:45 | [~50 char description] | Fix: [what to change] |
| 2 | HIGH | Workflow | - | [~50 char description] | Process: [change needed] |
| 3 | MED | Gap | - | [~50 char description] | Consider: [recommendation] |
## Details (if asked)
Expand any issue with "detail #N" - keep main output scannable.
Column Guide
| Column | Values | Notes |
|---|
| Sev | CRIT/HIGH/MED/LOW | CRIT=blocks work, HIGH=repeated failures, MED=inefficiency, LOW=minor |
| Type | Wrong Info / Workflow / Gap / Tool Fail | Maps to Category 1/2/3 + tool failures |
| Location | path:line or "-" | Where the problem is (if applicable) |
| Issue | Max ~50 chars | What's wrong - be specific |
| Action | Fix/Process/Consider | What to do about it |
Severity Guide
- CRIT: Wrong info in skills/docs causing failures - file bug immediately
- HIGH: Repeated workflow issues or significant rework needed
- MED: Inefficiencies, minor gaps
- LOW: Nice-to-have improvements
Type Mapping
- Wrong Info = Category 1 (bugs to file)
- Workflow = Category 2 (process issues)
- Gap = Category 3 (missing info to evaluate)
- Tool Fail = Tool errors with root cause
Good Examples
| 1 | CRIT | Wrong Info | skills/pdf-editor:45 | Recommends pdf-lib for reading (should be pdf-parse) | Fix: distinguish read vs edit |
| 2 | HIGH | Workflow | - | Didn't check git status before amend | Process: always check before amend |
| 3 | MED | Gap | - | No Node.js auth library guidance | Consider: create node-auth skill |
| 4 | LOW | Tool Fail | - | Used cd && instead of workdir param | Process: check tool params first |
Bad Examples (too vague)
| 1 | HIGH | Wrong Info | somewhere | Skill was wrong | Fix: make it better |
Important Guidelines
- Compact output, thorough analysis: Use the templates above internally, but OUTPUT only the issue table
- Be specific: "skills/pdf-editor:45 - recommends pdf-lib for reading" not "skill was wrong"
- Root causes, not symptoms: Trace to WHY - wrong info? didn't check? doesn't exist?
- CRIT = file bug now: Wrong info in skills/docs is a defect - don't just note it, fix it
- One line per issue: If you can't fit it in ~50 chars, you haven't found the core problem yet
Analysis Techniques
Conversation Audit
- Read entire conversation from start
- Mark every point where you:
- Got an error
- Got corrected by user
- Had to redo work
- Made an assumption that was wrong
- Chose wrong approach
- For each mark, ask "why did this happen?"
Tool Audit
- List every tool invocation
- For each failed tool call: full root cause analysis
- For successful calls: could it have been avoided/optimized?
- Identify patterns: same file read 3x, sequential ops that could be parallel
Information Audit (Most Important)
- List all skills loaded: mark which were actually used
- List all information in context: mark what was ignored
- For each failure: which category? (wrong info / didn't use / missing)
- For Category 1: prepare bug reports
- For Category 2: identify why info wasn't used
- For Category 3: evaluate if gap should be filled
Decision Audit
- List all significant decisions made
- For each: what information informed it?
- For wrong decisions: what information was wrong/missing/unused?
- For right decisions: was it luck or good process?
Concrete Example: Real Failure Analysis
Example Session
User asked to "implement file upload with progress tracking". You implemented using synchronous fs.writeFile, causing browser to hang on large files.
Compact Output (What You Deliver)
# Session Issues
**Session**: Implement file upload with progress tracking
**Outcome**: partial (rewrote after hang on large files)
| # | Sev | Type | Location | Issue | Action |
|---|-----|------|----------|-------|--------|
| 1 | HIGH | Workflow | - | Used fs.writeFile without asking about file sizes | Process: ask scale before I/O ops |
| 2 | MED | Gap | - | No Node.js file upload guidance exists | Consider: create file-upload skill |
## Details (if asked)
**#1**: Implemented sync file upload, worked for small test files, hung on user's 50MB video. Should have asked "what file sizes?" before choosing approach. Root cause: no I/O checklist trigger.
**#2**: Common task (high freq) + easy to get wrong (high risk) + security implications = worth creating skill. Content: streaming vs buffering, multer/busboy/formidable, progress tracking, file validation.
Internal Analysis (How You Got There)
Used Wrong Decision Analysis template internally:
- Decision: fs.writeFile for uploads (wrong)
- Should have: streaming approach
- Info issue: Category 2 (didn't research) + Category 3 (no skill exists)
- Red flags ignored: "large files" implies streaming needed
- Prevention: I/O checklist + file-upload skill
Self-Reflection Meta-Analysis
After completing a reflection, verify quality:
- ✅ Started with failures, not successes
- ✅ Completed three-category information audit
- ✅ Root causes identified, not just symptoms
- ✅ Category 1 issues formatted as bug reports
- ✅ Category 2 issues include process fixes
- ✅ Category 3 gaps evaluated with decision rationale
- ✅ Every recommendation is specific and actionable
- ✅ Recommendations prioritized by impact
- ✅ Patterns identified across multiple failures
- ✅ Meta-reflection completed
If any ❌, the reflection is incomplete. Redo it.
The goal is ruthless honesty in service of continuous improvement.
