skill-creator

name: skill-creator description: How to create new agent skills following the open standard (agentskills.io). Use when the user wants to create, scaffold, or design a new skill for any AI coding agent.

Skill Creator

Meta-skill for creating new agent skills that follow the agentskills.io open standard. Works across Claude Code, OpenAI Codex, Google Gemini CLI, Cursor, and Windsurf.

Input

• A skill idea from the user (name, purpose, or general description)

Output

• A complete skill directory with SKILL.md and optional subdirectories
• Placed in .agent/skills/<skill-name>/

Process

Phase 1: Interview (Understand the Skill)

Ask these questions one at a time. Wait for each answer before continuing.

1. What should this skill do? (1-2 sentence summary)
2. When should an agent use it? (trigger conditions — be specific)
3. What inputs does it need? (files, user answers, config, etc.)
4. What does it produce? (deliverables, files, decisions)
5. Does it have phases/steps? (break it down if yes)
6. Does it need templates, scripts, or reference docs? (optional directories)

After all questions are answered, confirm with the user:

"Here's what I understood: [summary]. Shall I create it?"

Phase 2: Scaffold (Generate the Skill)

Create the directory and SKILL.md using the template at templates/SKILL_TEMPLATE.md.

.agent/skills/<skill-name>/
└── SKILL.md              # Required — use the template

Frontmatter Rules (from agentskills.io spec)

Name Validation

✅ Valid:   code-review, sprint-planner, github-flow
❌ Invalid: Code-Review (uppercase), -my-skill (leading hyphen), my--skill (double hyphen)

Body Structure (our convention)

Follow this pattern — consistent with our existing skills:

# Skill Name

Brief description.

## Input
- What the skill needs

## Output
- What the skill produces

## Process

### Phase 1: [Name]
1. Step 1
2. Step 2

### Phase 2: [Name]
...

## Rules
1. **ALWAYS** ...
2. **NEVER** ...

Phase 3: Enrich (Add Optional Directories)

Based on the interview answers, add subdirectories as needed:

.agent/skills/<skill-name>/
├── SKILL.md
├── templates/          # Optional
│   └── TEMPLATE.md
├── references/         # Optional
│   └── GUIDE.md
├── scripts/            # Optional
│   └── helper.py
└── assets/             # Optional
    └── config.json

Important: Keep SKILL.md under 500 lines (< 5000 tokens). Move detailed content to references/.

Phase 4: Validate (Completeness Checklist)

Before delivering, verify:

• □ name field is kebab-case and matches directory name?
• □ description field clearly states what + when?
• □ SKILL.md body has Input, Output, Process, and Rules sections?
• □ Body is under 500 lines?
• □ All referenced templates/scripts exist?
• □ No hardcoded paths — skill is portable?
• □ Rules section has at least 2 guardrails?
• □ Each phase produces something verifiable?

Progressive Disclosure (How Agents Load Skills)

Skills are loaded in 3 tiers to save context:

Tier 1: Discovery  → name + description only (~100 tokens)
Tier 2: Activation → Full SKILL.md body (< 5000 tokens)
Tier 3: Resources  → scripts/, references/, assets/ (on demand)

This means:

• Description is critical — it's the ONLY thing agents see at startup
• Keep SKILL.md focused — move extras to references/
• Templates are loaded only when used — don't inline them

Cross-Platform Compatibility

Skills created with this tool work across all major platforms:

For cross-platform details, see references/CROSS_PLATFORM_GUIDE.md.

Rules

1. ALWAYS ask the user before creating — never assume skill requirements
2. ALWAYS use kebab-case for directory and name field
3. ALWAYS write descriptions that include BOTH what + when
4. ALWAYS include a Rules section with guardrails
5. NEVER exceed 500 lines in SKILL.md — use references/ for overflow
6. NEVER hardcode absolute paths in skill content
7. NEVER skip the validation checklist
8. ALWAYS use English for all skill content