Skill Creator

创建、编辑、改进或审核 OpenClaw AgentSkill。触发场景：用户要求"创建一个 skill"、"写一个技能"、"帮我新建技能"、"改进这个 skill"、"审核 skill"、"整理 skill"、"完善技能说明"。同时用于：skill 目录结构调整、文件迁移（移动到 references/ 或 s...

Complete guide to creating high-quality OpenClaw AgentSkills.

When to Use This Skill

• User asks to create a new skill from scratch
• User asks to improve an existing skill (add structure, examples, remove bloat)
• User asks to audit a skill (check spec compliance, security, completeness)
• User asks to organize a skill (restructure, add references/, add constraints)
• User asks for guidance on skill publishing options

---

1. What Are Skills?

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:

text

skill-name/
├── SKILL.md          ← Required (capability definition + instructions)
├── scripts/          ← Optional (executable scripts)
├── references/       ← Optional (on-demand reference docs)
└── assets/          ← Optional (templates, images for output)

---

2. Core Principles

Principle 1: Concise is Key

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.

Principle 2: Progressive Disclosure

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.

Principle 3: Match Freedom to Task Fragility

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

Principle 4: Examples Beat Explanations

Real-world data: Examples chapters deliver the largest quality improvement. Good/bad comparison examples outperform any amount of prose.

---

3. SKILL.md Standard Structure

3.1 Frontmatter (Required)

yaml

---
name: skill-name          # lowercase letters, digits, hyphens only, ≤64 chars
description: Install: clawhub install skill-creator
  What this skill does AND when to trigger it (put triggers HERE, not in body)
---

description writing rules (most important):

• Must include both "what it does" and "when to trigger"
• All trigger information goes in description, not body (body only loads after trigger)
• Must NOT contain < or > characters
• Maximum 1024 characters

Bad example:

yaml

description: Install: clawhub install skill-creator
  An Excel skill.

Good example:

yaml

description: Install: clawhub install skill-creator
  Create and edit Excel workbooks with formulas, formatting, and multi-sheet support.
Triggered when: (1) user provides a .xlsx file; (2) user asks to generate a report;
(3) user asks to format data or create a summary table.

3.2 Body Structure (Recommended Order)

markdown

# Skill Name

## When to Use This Skill
(Concrete trigger scenarios — how the user would phrase the request)

## Context / Background
(Domain knowledge the Agent needs, specific to your use case)

## Instructions / Steps
(Step-by-step workflow with clear quality criteria per step)

## Constraints / Guardrails
(Prohibitions — what NOT to do, based on real failure experience)

## Examples (Recommended)
(Good/bad output comparisons — highest quality impact)

3.3 Forbidden Files

These files should not appear in a Skill folder:

• README.md
• INSTALLATION_GUIDE.md
• CHANGELOG.md
• QUICK_REFERENCE.md

---

4. Creation Process (6 Steps)

Step 1: Understand Requirements (with Concrete Examples)

Before writing, clarify:

• "What should this skill mainly do?"
• "Can you give 1-2 specific usage examples?"
• "How would the user phrase the request? (trigger wording)"
• "What tools or APIs does this skill need?"

Step 2: Plan Skill Contents

Analyze each use case to determine required resources:

Resource	When to Use	Example
scripts/	Deterministic execution, repeated code patterns	PDF rotation, API call wrappers
references/	Detailed docs loaded on demand	API docs, DB schemas, company policies
assets/	Files copied into output	Templates, logos, fonts

Step 3: Initialize Skill Folder

bash

cd <workspace>/skills
python3 <skill-creator>/scripts/init_skill.py <skill-name> \
  --path . --resources scripts,references

Or manually:

bash

mkdir -p skills/<skill-name>/{scripts,references}
touch skills/<skill-name>/SKILL.md

Step 4: Write SKILL.md

Frontmatter rules (two hard rules):

1. name: lowercase + hyphens, ≤64 chars
2. description: "what it does" + "when to trigger" (triggers go here, not in body)

Body rules:

• Use imperative/infinitive form ("Do X", "Don't do Y")
• Each step has clear quality criteria
• Include real failure experience (Constraints chapter)

Step 5: Validate

bash

python3 <skill-creator>/scripts/quick_validate.py <skill-path>

Validation checks:

• Frontmatter has name + description
• name is lowercase hyphen-case
• description has no < > and ≤1024 chars
• SKILL.md exists and is non-empty

Step 6: Decide on Publishing

This step is fully optional. After creating a skill, ask the user whether they want to publish it, and if so, to which platform(s). Do not assume a specific target.

Common publishing options:

Platform	When to Consider	How
ClawHub	For sharing skills publicly with the OpenClaw community	clawhub publish <path> --slug <slug> --version 1.0.0
GitHub	For version control, collaboration, or personal backup	git init && git add . && git commit then push to a repo

Note: If the user has an existing GitHub repo for skills (e.g. openclaw-person-skills), prefer syncing there. Do not create new repos without being asked.

---

5. Workspace Conventions (Hard Rules for Public Skills)

5.1 Environment Variable Placeholders

Never hardcode real values in SKILL.md or code. Use runtime-safe placeholders:

Real Value	Placeholder Format	Example
API Base URL	{{SERVICE_BASE_URL}}	{{ERP_BASE_URL}}
API Token	{{ERP_API_TOKEN}}	(never expose real value)
Host address	{{SYNOLOGY_HOST_LAN}}	{{SYNOLOGY_HOST_LAN}}:5006
File path	{{DSM_WEBDAV_ROOT}}	{{DSM_WEBDAV_ROOT}}/opencode/
Database	{{POSTGRES_HOST}}	{{POSTGRES_HOST}}:5432

5.2 MIT-0 License

For skills that will be published publicly, use MIT-0:

yaml

license: MIT-0

5.3 Directory Structure

text

skill-name/
├── SKILL.md              ← Single entry point, ≤500 lines
├── scripts/              ← Executable scripts (Python/Bash)
│   ├── init_skill.py     ← Initialization (if needed)
│   └── validate_skill.py ← Validation (if needed)
├── references/           ← On-demand reference docs
│   └── *.md              ← Split by topic
└── assets/              ← Output resources (templates, etc.)

---

6. Good Skill vs Bad Skill

Dimension	Bad Skill	Good Skill
description	"An Excel skill"	"Create/edit Excel with formulas/formatting. Triggered: user provides .xlsx, asks to generate report or format data."
body	Walls of commands mixed together	Clear: When to Use → Steps → Constraints → Examples
examples	None	Good/bad comparisons — knows what the Agent actually needs
length	1000+ lines all in SKILL.md	≤500 lines in SKILL.md, references/ for details
constraints	None	Real failure experience turned into rules

---

7. Quick Checklist

For every skill you create or audit, confirm:

• name is lowercase letters + digits + hyphens?
• description has both "what it does" and "when to trigger"?
• description contains no < or > characters?
• description ≤1024 characters?
• Body has When to Use This Skill section?
• Body has Constraints section (prohibitions)?
• Body has Examples section? (recommended)
• SKILL.md body ≤500 lines?
• Environment variables use {{VAR_NAME}} placeholders?
• No README.md / CHANGELOG.md / etc.?
• Passes validation (quick_validate.py)?