ABOUTME: Skill for analyzing Claude Code sessions to improve or create skills

ABOUTME: Extracts session transcripts and provides structured improvement workflows

Improve Skill

Analyze Claude Code session transcripts to enhance existing skills or create new ones.

When to Use This Skill vs Others

Use improve-skill When:

• You've just finished a debugging/automation session and want to analyze it
• You need to extract session transcript for offline review
• You want to identify what went wrong in your workflow
• You're creating new skills from successful workflows
• You want to improve existing skills based on what you learned
• Decision: Use AFTER you've completed a web-automation session

Use web-automation When:

• You're actively debugging or automating something NOW
• Decision: Use BEFORE you would use improve-skill (this is the "doing" skill)

Typical Workflow

1. Run web-automation to debug/automate
   (this is the "doing" phase)

2. Save session transcript

3. Use improve-skill to analyze what happened
   (this is the "learning" phase)

4. Update skill documentation based on findings

5. Next time, your updated skill helps you work faster

🔄 RESUMED SESSION CHECKPOINT

When a session is resumed from context compaction, verify skill improvement state:

┌─────────────────────────────────────────────────────────────┐
│  SESSION RESUMED - IMPROVE-SKILL VERIFICATION               │
│                                                             │
│  Before continuing skill improvement work:                  │
│                                                             │
│  1. Was I in the middle of session analysis?                │
│     → Check /tmp/ for extracted transcripts                 │
│     → Check summary for "analyzing", "extracting"           │
│                                                             │
│  2. Was I updating a skill?                                 │
│     → Check which skill was being modified                  │
│     → Run: uv run validate_skill.py on the skill            │
│                                                             │
│  3. Did the previous work complete?                         │
│     → Check todo list for pending analysis tasks            │
│     → Verify skill validation passed                        │
│                                                             │
│  If analysis was in progress:                               │
│  → Re-run the extraction/analysis scripts                   │
│  → Do NOT assume previous analysis is still valid           │
└─────────────────────────────────────────────────────────────┘

Quick Reference

Script	Purpose
extract_session.py	Extract session transcript to markdown
analyze_session.py	Analyze session for skill improvements

Run any script with --help for full options:

uv run ~/.claude/skills/improve-skill/scripts/extract_session.py --help

Workflows

1. Improve an Existing Skill

When a skill didn't work as expected, extract the session and analyze it:

# Step 1: Extract the session transcript
uv run ~/.claude/skills/improve-skill/scripts/extract_session.py \
    --session-id <session-id> \
    --output /tmp/session.md

# Step 2: Analyze for improvements
uv run ~/.claude/skills/improve-skill/scripts/analyze_session.py \
    --transcript /tmp/session.md \
    --skill ~/.claude/skills/<skill-name>/SKILL.md \
    --output /tmp/analysis.md

Then start a new Claude Code session with:

Review the analysis in /tmp/analysis.md and update the skill accordingly.

Why a new session? Starting fresh ensures only the transcript and skill are loaded, without prior context influencing the analysis.

2. Create a New Skill from Session

When you've done something successfully and want to capture it as a skill:

# Step 1: Extract the session
uv run ~/.claude/skills/improve-skill/scripts/extract_session.py \
    --session-id <session-id> \
    --output /tmp/session.md

# Step 2: Generate skill creation prompt
uv run ~/.claude/skills/improve-skill/scripts/analyze_session.py \
    --transcript /tmp/session.md \
    --create-skill \
    --output /tmp/new_skill_prompt.md

Then start a new session to create the skill based on the analysis.

3. Quick Session Review

List recent sessions and extract the most recent one:

# List recent sessions
uv run ~/.claude/skills/improve-skill/scripts/extract_session.py --list

# Extract most recent session for current directory
uv run ~/.claude/skills/improve-skill/scripts/extract_session.py --latest

# Extract most recent session for a specific project
uv run ~/.claude/skills/improve-skill/scripts/extract_session.py \
    --latest \
    --cwd /path/to/project

Finding Session IDs

Session IDs can be found:

1. In the Claude Code UI (shown in session info)
2. Using the /tasks command
3. By listing sessions: uv run ~/.claude/skills/improve-skill/scripts/extract_session.py --list

Analysis Framework

When analyzing a session transcript, look for:

Pattern	What to Look For
Confusion	Where did the agent struggle or misunderstand?
Missing Info	What information was needed but not in the skill?
Workarounds	What manual steps were needed that could be automated?
Errors	What errors occurred and how were they resolved?
Success	What approaches worked well and should be documented?
Wrong Skill	Was a different skill more appropriate?

Output Format

The analysis produces structured recommendations:

## Session Analysis

### What Worked Well
- [List of successful approaches]

### Issues Identified
- [List of problems encountered]

### Recommended Improvements
- [Specific changes to make to the skill]

### Missing Documentation
- [Information that should be added]

Help Output Examples

extract_session.py

usage: extract_session.py [-h] [--list] [--latest] [--session-id SESSION_ID]
                          [--cwd CWD] [--output OUTPUT]

Extract Claude Code session transcripts to markdown

options:
  -h, --help            show this help message and exit
  --list                List all available sessions
  --latest              Extract most recent session
  --session-id SESSION_ID, -s SESSION_ID
                        Extract specific session by ID
  --cwd CWD             Filter sessions by working directory
  --output OUTPUT, -o OUTPUT
                        Save transcript to file (default: stdout)

Examples:
  uv run extract_session.py --list
  uv run extract_session.py --latest --output /tmp/session.md
  uv run extract_session.py --session-id abc123 --output /tmp/session.md

analyze_session.py

usage: analyze_session.py [-h] --transcript TRANSCRIPT [--skill SKILL]
                          [--create-skill] [--quick] [--output OUTPUT]

Analyze Claude Code session transcripts for skill improvement

options:
  -h, --help            show this help message and exit
  --transcript TRANSCRIPT, -t TRANSCRIPT
                        Path to the session transcript markdown file (required)
  --skill SKILL, -s SKILL
                        Path to existing skill SKILL.md to improve
  --create-skill        Generate prompt for creating a new skill
  --quick               Quick analysis summary only (no full prompt)
  --output OUTPUT, -o OUTPUT
                        Output file path (default: stdout)

Examples:
  # Quick analysis
  uv run analyze_session.py --transcript /tmp/session.md

  # Improve existing skill
  uv run analyze_session.py --transcript /tmp/session.md \
      --skill ~/.claude/skills/web-automation/SKILL.md

  # Create new skill
  uv run analyze_session.py --transcript /tmp/session.md --create-skill

  # Save analysis to file
  uv run analyze_session.py --transcript /tmp/session.md --output /tmp/analysis.md

Output includes:

• Automatically detected struggles (missing files, permission issues, import errors, etc.)
• Missing information (implicit knowledge, missing examples, edge cases, etc.)
• Success patterns (what worked well, effective approaches)
• Session metrics (tool invocations, errors, retries, successes)
• Assessment (overall session quality and recommendations)

Tips

• Use a fresh session for analysis to avoid context pollution
• Focus on patterns, not individual instances
• Keep skills concise - only add essential information
• Include examples for complex workflows
• Update trigger words if the skill wasn't invoked when expected
• Start with quick analysis to get overview, then full analysis for details