
skill-manager
by amco3008
SKILL.md
name: skill-manager description: Manages the dynamic skill learning system. Use when you want to list skills, scan transcripts for patterns, learn preferences, approve/reject candidates, rollback versions, view skill history, or check usage statistics. Trigger phrases include "manage skills", "skill learn", "list skills", "approve skill", "skill stats", "skill history", "rollback skill", "scan patterns", "learn preferences". allowed-tools:
- Bash
- Read
- Grep
- Glob
- Write
Skill Manager
This skill manages the dynamic skill learning and generation system. It enables automatic skill creation from usage patterns, version control, and auto-improvement.
Core Capabilities
- Pattern Observation - Hooks capture tool usage and prompt patterns
- Skill Learning - Analyze patterns to generate skill candidates
- Version Control - Git-based versioning with rollback
- Auto-Improvement - Automatically enhance skills based on effectiveness
Commands
All commands use the manage.py CLI:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py <command> [args]
List All Skills
Show installed skills with status and effectiveness scores:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py list
Scan Transcripts
Scan Claude transcripts for usage patterns (run before learn):
python3 ~/.claude/skills/_skill-manager/scripts/manage.py scan
Learn Preferences
Learn user preferences from prompts and update CLAUDE.md:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py preferences --force
View Captured Prompts
View recent user prompts:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py prompts [N]
Learn from Patterns
Analyze collected patterns and generate skill candidates:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py learn
Show Candidates
List pending skill candidates awaiting approval:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py candidates
Approve a Skill
Approve a candidate skill for use:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py approve <candidate-id>
Reject a Skill
Reject and remove a candidate skill:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py reject <candidate-id>
View Skill History
Show version history for a skill:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py history <skill-name>
Rollback Skill Version
Restore a skill to a previous version:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py rollback <skill-name> <version>
View Statistics
Show usage analytics and effectiveness metrics:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py stats
Force Auto-Improvement
Trigger improvement analysis for a specific skill:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py improve <skill-name>
Refine a Skill
Interactively improve a skill with new patterns:
python3 ~/.claude/skills/_skill-manager/scripts/manage.py refine <skill-name>
How Pattern Learning Works
1. Observation Phase
Hooks automatically capture:
- Tool sequences - Which tools are used in what order
- Domain keywords - DevOps, security, frontend, etc.
- Success/failure signals - Task completion, errors, retries
2. Analysis Phase
The learning engine detects:
- Repeated workflows - Same tool chains used 2+ times
- Domain expertise - Concentration of domain-specific keywords
- Reusable patterns - Generalizable processes
3. Scoring Phase
Candidates are scored on:
| Factor | Weight | Criteria |
|---|---|---|
| Frequency | 30% | Pattern appears 2+ times |
| Complexity | 20% | 3-7 tool steps optimal |
| Domain clarity | 25% | Clear domain association |
| Distinctiveness | 25% | Different from existing skills |
4. Generation Phase
High-scoring patterns become skill packages with:
SKILL.md- Definition with YAML frontmatterskill.meta.json- Metadata and effectiveness trackingreferences/- Extracted documentationscripts/- Helper utilities (if applicable)CHANGELOG.md- Version history
Auto-Improvement System
Skills automatically improve when:
- Success rate drops below 70%
- New enhancing patterns detected
- 50+ usages since last improvement
- User manually refines skill (system learns from refinement)
Improvement Actions
- Add missing tool permissions
- Update reference documentation
- Add error handling for failure cases
- Refine trigger phrases
- Optimize workflow steps
Confidence-Based Application
| Confidence | Action |
|---|---|
| ≥90% | Auto-apply improvement |
| 70-89% | Queue for user approval |
| <70% | Suggest but don't queue |
Version Control
Each skill is a Git repository supporting:
# View version history
cd ~/.claude/skills/<skill-name>
git log --oneline
# Compare versions
git diff v1.0.0 v1.1.0
# Rollback
git checkout v1.0.0
Configuration
Edit ~/.claude/skills/.skill-system/config.json:
{
"observation": {
"enabled": true,
"pattern_retention_days": 90
},
"learning": {
"auto_suggest": true,
"min_frequency": 2,
"score_threshold": 0.6
},
"improvement": {
"enabled": true,
"auto_apply_threshold": 0.9,
"min_usage_before_improve": 10,
"success_rate_trigger": 0.7
}
}
Supported Domains
Pre-configured domain detection:
- devops - docker, kubernetes, ci/cd, terraform
- security - audit, vulnerability, owasp, encryption
- data_science - pandas, numpy, model, tensorflow
- frontend - react, vue, css, webpack
- backend - api, rest, database, orm
- git - merge, rebase, pr, branch
Custom domains can be added in config.
Data Storage
| Location | Contents | Persistence |
|---|---|---|
.skill-system/patterns/ | Raw observations | 90 days |
.skill-system/candidates/ | Pending approvals | Until approved/rejected |
.skill-system/config.json | Configuration | Permanent |
<skill-name>/ | Active skills | Git-versioned |
Example Workflow
- Normal usage - Use Claude Code normally
- Scan transcripts -
manage.py scanextracts patterns from transcripts - Learn preferences -
manage.py preferences --forceupdates CLAUDE.md with learned rules - Generate candidates -
manage.py learncreates skill candidates from patterns - Review candidates -
manage.py candidatesshows suggestions - Approve -
manage.py approve docker-compose-helper - Use skill - Skill is now active and tracking effectiveness
- Auto-improve - System enhances based on usage data
Troubleshooting
Hooks not capturing
Check hooks are registered:
cat ~/.claude/settings.local.json | grep -A5 hooks
No candidates generated
Ensure sufficient pattern data:
wc -l ~/.claude/skills/.skill-system/patterns/*.jsonl
Skill not improving
Check effectiveness metrics:
cat ~/.claude/skills/<skill-name>/skill.meta.json | jq .effectiveness
Score
Total Score
Based on repository quality metrics
SKILL.mdファイルが含まれている
ライセンスが設定されている
100文字以上の説明がある
GitHub Stars 100以上
3ヶ月以内に更新がある
10回以上フォークされている
オープンIssueが50未満
プログラミング言語が設定されている
1つ以上のタグが設定されている
Reviews
Reviews coming soon