Case 01902

OpenClaw Skill Creator

A skill for creating and iteratively improving OpenClaw skills — inspired by Anthropic's skill-creator, adapted for the OpenClaw skill format.

OpenClaw Skill Architecture

Skills are directories containing a SKILL.md file and optional resources.

Custom skills (user-created) are typically placed in the OpenClaw workspace:

<workspace>/skills/<skill-name>/

Built-in skills ship with the OpenClaw installation.

Custom skills in the workspace take precedence over built-ins of the same name.

Minimal structure

<skill-name>/
├── SKILL.md          ← required
└── scripts/          ← optional, executable helpers

No packaging step needed. The directory is the skill.

SKILL.md format

---
name: skill-name
description: <trigger description — the most important field>
---

# Skill Name

Instructions for the agent...

Only name and description in frontmatter. No other fields.

Use {baseDir} in scripts/instructions to reference the skill's own directory:

python3 {baseDir}/scripts/my_script.py

Writing Descriptions (Most Critical Part)

The description field is the only triggering mechanism. The body is never read unless the skill triggers first.

Core principles

1. Be pushy: Agents tend to undertrigger skills. Make the description slightly aggressive — list specific contexts where this skill MUST be used.
2. Use when + NOT for: Explicit structure reduces both under-triggering and false positives.
3. Include trigger phrases: List common user phrasings. Include non-English phrases if the user communicates in other languages.
4. Describe context, not just function: "What the skill does AND when to use it."

Template

<what the skill does>. Use when: (1) <context 1>, (2) <context 2>, (3) phrases like "<example phrase>". NOT for: <anti-patterns>.

Example (before → after)

❌ Before: "Track work time per project with start/stop timers."

✅ After: "Track work time and generate productivity reports. Use this skill whenever the user wants to start/stop a timer, log time to a project, or generate a time report — including phrases like 'start timer', 'log my time', 'how long did I work this week'. NOT for: calendar scheduling, task management."

Skill Creation Process

1. Capture Intent

Ask the user:

• What should this skill enable the agent to do?
• When should it trigger? What would the user say?
• What's the expected output?
• Does it need scripts, or is it purely instructional?

Ask one or two questions at a time; don't overwhelm.

2. Interview for Edge Cases

Before writing, ask about:

• Input formats, edge cases, failure modes
• Dependencies (external APIs, CLIs, files)
• What it should NOT do (important for the NOT for clause)

3. Write the SKILL.md

• Keep body under 300 lines — lean and focused
• Only include what the agent doesn't already know
• Move detailed reference material to references/ files and link from SKILL.md
• Add scripts to scripts/ when the same code would be rewritten repeatedly

4. Eval: Test the Description

Run a manual trigger eval:

1. Write 5–10 test prompts (mix of should-trigger and should-not-trigger)
2. For each, ask: "Would I read this skill's SKILL.md given only the description?"
3. If < 80% correct → improve the description

Use the eval script:

python3 {baseDir}/scripts/eval_description.py --skill path/to/SKILL.md --evals path/to/evals.json

See scripts/eval_description.py for the evals JSON format.

5. Iterate

After the skill is used on real tasks:

• If the skill often doesn't trigger when it should → description too narrow, add more trigger phrases
• If the skill triggers on unrelated requests → description too broad, add NOT for clause
• If the body is too long → split into references/ files

Improving Existing Skills

When asked to improve an existing skill:

1. Read the current SKILL.md
2. Check description against the principles above
3. Propose a new description with reasoning
4. Optionally run eval to compare before/after trigger accuracy
5. Update only after user confirmation

Skill Quality Checklist

Before finalising a skill, verify:

• name is lowercase kebab-case, under 64 characters
• description includes "Use when" with specific trigger contexts
• description includes "NOT for" to prevent false positives
• Body is under 300 lines
• No README, CHANGELOG, or auxiliary docs (keep the skill lean)
• Scripts are tested and working
• References files are linked from SKILL.md with guidance on when to read them