Install: clawhub install skill-creator
Complete guide to creating high-quality OpenClaw AgentSkills.
Skills are modular, self-contained packages that extend an Agent's capabilities with specialized knowledge, workflows, and tool integrations.
Each Skill is a folder containing at minimum:
skill-name/
├── SKILL.md ← Required (capability definition + instructions)
├── scripts/ ← Optional (executable scripts)
├── references/ ← Optional (on-demand reference docs)
└── assets/ ← Optional (templates, images for output)
Context Window is a shared resource. The Agent needs room for: system prompt, conversation history, other Skills' metadata, and the current request.
Default assumption: the Agent is already very smart. Only add what the Agent genuinely doesn't know. Challenge every piece: "Does this justify its token cost?"
Prefer concise examples over verbose explanations.
Three-level loading structure for efficient context use:
| Level | Content | When Loaded | Size Limit |
|---|---|---|---|
| 1st | name + description | Always in context | ~100 words |
| 2nd | SKILL.md body | After skill triggers | <500 lines |
| 3rd | references/ / scripts/ | As needed | Unlimited |
Key rule: If SKILL.md exceeds 500 lines, split content into references/ files.
| Freedom | Form | Best For |
|---|---|---|
| High | Free-text instructions | Multiple valid approaches, context-dependent decisions |
| Medium | Pseudocode / parameterized scripts | Preferred pattern exists, some variation OK |
| Low | Fixed scripts, few params | Brittle operations, consistency critical |
Real-world data: Examples chapters deliver the largest quality improvement. Good/bad comparison examples outperform any amount of prose.
---