Skill-JIT
Let Agents Learn and Persist Skills on the Fly
🐙 GitHub
|
📄 Blog
Installation
# Claude Code
claude plugin marketplace add china-qijizhifeng/skill-jit
claude plugin install skill-jit@skill-jit-marketplace
# OpenClaw
openclaw plugins install skill-jit --marketplace china-qijizhifeng/skill-jit
How It Works
Skill-JIT uses a two-agent architecture:
- Researcher — Deep-dives into documentation, tests commands, verifies behaviors, and returns structured knowledge with evidence.
- Writer — Decomposes tasks, orchestrates researchers, selects the right skill pattern, and generates production-ready skill files.
Workflow
User Request → Parse Input → Detect Mode (Create / Fix)
→ Capture Intent (what, when, output, edge cases)
→ Spawn Writer Agent → Writer spawns Researcher(s)
→ Combine Research → Select Pattern → Write Skill Files
Skill Patterns
Skills are generated following one of five proven patterns:
| Pattern | Use When | Example |
|---|
| Tool Wrapper | Teaching how to use a tool/API/CLI | ffmpeg, Stripe SDK |
| Generator | Producing artifacts with fixed structure | Changelogs, API docs |
| Reviewer | Checking/auditing against criteria | Code review, security audit |
| Inversion | Agent needs user decisions before acting | Project scaffolding, config wizards |
| Pipeline | Strict sequential workflow with gates | Build → test → deploy |
Patterns can be composed (e.g., Generator + Inversion, Pipeline + Reviewer).
Output Structure
Generated skills follow progressive disclosure:
.claude/skills/<skill-name>/
├── SKILL.md # Main skill (frontmatter + body, <500 lines)
├── references/ # Deep-dive docs, parameters, troubleshooting
├── templates/ # Output templates (Generator pattern)
├── rubrics/ # Checklists (Reviewer pattern)
└── scripts/ # Helper scripts
Usage
Create a new skill
/skill-jit Create a skill for ffmpeg video-to-GIF conversion
Fix an existing skill
/skill-jit Fix the ffmpeg-gif skill — palette generation fails on macOS
Dry run (preview without writing)
/skill-jit --dry-run Create a skill for Docker multi-stage builds
External Services
Skill-JIT integrates with MCP servers for enhanced research:
Design Principles
- Explicit by Default — Every step has one clear action; no "use your judgment"
- Research-Grounded — Trust test results over documentation
- Generalize, Don't Overfit — Skills are reusable, not task-specific
- Progressive Disclosure — Frontmatter → body → bundled resources
License
MIT
Author
KYLN24
