skill-creator（技能创建器）

Skill Creator: The Design System

This skill uses a "Design First, Code Second" approach. Your goal is not just to create a folder, but to design an effective tool for an autonomous agent.

Core Philosophy

1. Concise is Key: The context window is precious. Do not overload SKILL.md.

- Limit: Keep SKILL.md under 500 lines. - Strategy: Move large prompt chains, examples, or schemas to resources/ or references/.

1. Progressive Disclosure:

- Level 1: Metadata (Name/Description) - Agent sees this first. - Level 2: SKILL.md - Agent reads this only when activated. - Level 3: Scripts/References - Agent reads/runs these only when needed.

1. IDE Awareness:

- Always use absolute paths. - Use task_boundary for long-running workflows. - Assume the agent knows nothing about the file structure until it runs ls.

1. Artifact Persistence (Dual-Write Pattern): [!IMPORTANT] Phase 1: Draft in Brain (Ephemeral) Create artifacts in brain/ directory. Iterate with user via notify_user. All Q&A, sketches, and intermediate versions stay here. Phase 2: Persist on Approval (Permanent) ONLY after user says "Looks good" → copy final content to project/docs/ path. The project/docs/ folder is the only memory that survives between sessions.

- Why Two Phases? Brain = safe iteration space (no git pollution). Docs = approved truth. - Requirement: Every skill MUST save its final output to project/docs/ before handoff. - Registry: Update project/docs/ARTIFACT_REGISTRY.md to signal the baton is ready.

Repository Structure

This repository is a Skill Factory that separates the creator from the products:

ag-skill-factory/
├── .agent/skills/         # Factory tooling (only skill-creator)
│   └── skill-creator/     # This skill - creates other skills
├── squads/                # Generated skills live here
│   ├── backend-go-expert/
│   ├── frontend-nuxt/
│   └── ...
└── Makefile               # Links skills to global brain

IMPORTANT: All new skills MUST be created in squads/, NOT in .agent/skills/ or directly in the global brain.

Skill Creation Workflow

Phase 1: Design (Mental Step)

Before running any script, answer these questions:

1. What is the Trigger? What user intent triggers this skill? (Write this in the Description).
2. What is the Decision Tree? Does the skill have one path or multiple? (If multiple, plan a "Decision Tree" section).
3. What are the Resources? Do I need a Python script for logic? A Markdown template for output?

Phase 2: Scaffold

Run the initialization script to create the standard structure in squads/:

# ALWAYS create skills in squads/ directory
python3 .agent/skills/skill-creator/scripts/init_skill.py <skill-name>

This creates:

squads/<skill-name>/
├── SKILL.md          # The "Brain": Logic, Decisions, Workflow.
├── scripts/          # The "Hands": Python/Bash scripts for execution.
├── resources/        # The "Tools": Templates, Configs (e.g., Dockerfile).
├── examples/         # The "Demos": Usage examples.
└── references/       # The "Library": Documentation, Cheatsheets.

Phase 3: Refine

Edit the generated files:

1. Fill the Decision Tree: Map user inputs to actions.
2. Write the Workflow: Step-by-step instructions.
3. Clean up: Remove unused sections from the template.
4. Adapt the checklist: Rewrite references/checklist.md for this skill's domain.

- Mandatory: Keep the 🚨 Document Persistence section! - The template checklist is generic — make it specific! - Example: For an MCP skill, add checks for "stdio transport", "tool descriptions", etc.

1. Define Artifact Ownership: What files does this skill own in project/docs/?

Phase 4: Verify

Run the validation script:

make validate SKILL=<skill-name>

This checks:

• Frontmatter, length (<500 lines)
• Team Collaboration & When to Delegate sections
• Customized checklist in references/
• No large embedded code blocks

Fix all errors before installing.

Phase 5: Install (Link to Global Brain)

After creating and refining the skill, link it to the global brain:

make install    # Validates all, generates TEAM.md, then links

This copies skills from squads/<skill-name>/ → ~/.gemini/antigravity/global_skills/<skill-name>/.

Makefile Commands

Anti-Patterns

❌ DO NOT create skills directly in ~/.gemini/antigravity/global_skills/ ❌ DO NOT use the --global flag in init_skill.py ❌ DO NOT place new skills in .agent/skills/ (reserved for factory tooling)

✅ DO create skills in squads/ ✅ DO use make install-squads to link them globally

Content Organization Rules

SKILL.md should contain:

• Decisions, workflows, and logic (the "brain")
• Artifact Ownership: What docs this skill creates/updates
• Handoff Protocol: Rules for passing the baton
• Brief inline examples (max 5-10 lines)
• References to detailed examples: See examples/python-server.py

examples/ should contain:

• Full working code examples (complete files)
• Configuration samples
• Complete templates

references/ should contain:

• Cheatsheets and quick reference guides
• External documentation links
• Troubleshooting guides

resources/ should contain:

• Templates for generation
• Config files
• Static assets

CRITICAL: Do NOT embed large code blocks (>10 lines) in SKILL.md. Instead, create files in examples/ and reference them.

Language Requirements

CRITICAL: All skill files MUST be written in English.

Why English?

• Skills are consumed by AI agents globally
• English ensures consistent parsing and understanding
• Easier to maintain and contribute

Localization rule:

• SKILL.md, references/, resources/, examples/ → English only
• Agent runtime communication → User's language (agent adapts automatically)

Validation will check for Cyrillic characters and fail if found in skill files.

Team Collaboration

Every skill should know about the team. Add these sections to SKILL.md:

Team Collaboration — List related skills by role:

## Team Collaboration
- **<Role>**: `@<skill-name>` (Brief description of collaboration)
- **<Role>**: `@<skill-name>` (Brief description of collaboration)

When to Delegate — When to hand off to another skill:

## When to Delegate
- ✅ **Delegate to `@<skill-name>`** when: <Condition for handoff>
- ⬅️ **Return to `@<skill-name>`** if: <Condition to return>

Team Reference: See squads/TEAM.md for the full list of available skills.