Skill Creator

Skill Creator

A skill for creating and iteratively improving JavaClaw skills.

Skills in JavaClaw live at workspace/skills/<skill-name>/SKILL.md and are loaded dynamically by the SkillsTool at runtime. The agent picks them up automatically — no restart needed.

The core loop:

1. Understand what the skill should do
2. Write a draft SKILL.md
3. Walk through 2-3 test scenarios and evaluate the output inline
4. Refine based on feedback
5. Repeat until the user is happy

Jump in wherever the user is. If they already have a draft, skip straight to testing. If they just have a vague idea, interview them first.

Creating a skill

Capture Intent

Start by understanding what the user actually wants. If the conversation already shows a workflow the user wants to capture (a sequence of steps, tools used, corrections made), extract the answers from that — then fill gaps with the user.

Key questions:

1. What should this skill enable the agent to do?
2. When should it trigger? (what kinds of user messages)
3. What does a good output look like?

Interview

Ask about edge cases, expected inputs, success criteria, and any context the agent will need. Don't start writing until you have enough to write something useful — but don't over-interview either. If you can make a reasonable assumption, make it and note it.

Write the SKILL.md

Create workspace/skills/<skill-name>/SKILL.md with:

• name: kebab-case identifier matching the directory name
• description: The primary triggering mechanism. Include what the skill does AND when to use it. The agent decides whether to load a skill based solely on this field — so be specific and make it slightly "pushy". Instead of "Helps with data analysis", write "Helps with data analysis. Use this skill whenever the user asks about datasets, CSV files, charts, or wants to understand or transform data — even if they don't say 'analysis'."
• Body: Step-by-step instructions for what the agent should do. Explain the why behind important steps so the agent can apply judgment, not just follow rules mechanically.

Anatomy of a skill

workspace/skills/
└── skill-name/
    └── SKILL.md   ← required; keep under ~300 lines

JavaClaw skills are single-file. The agent reads SKILL.md when the skill triggers. Keep it focused and readable — a dense wall of text is harder to follow than clear, structured instructions.

Writing patterns

Use imperative form. Explain reasoning where it matters.

Output format example:

## Report structure
Use this template every time:
# [Title]
## Summary
## Details
## Next steps

Example pattern:

## Commit message format
User said: "Added login with Google"
Write: feat(auth): add Google OAuth login

Principles

• Lean over comprehensive. A shorter, well-reasoned skill beats a long checklist.
• Explain the why so the agent can adapt to situations the skill didn't explicitly anticipate.
• Avoid ALL CAPS MUSTs and rigid structures where possible — trust the agent's judgment when given good context.
• Skills must not contain malware, exploit code, or content that would surprise the user given the skill's stated purpose.