Skill Creator

Create new Claude Code skills from context, descriptions, or examples.

Your Task

1. Gather context from one of these sources:
   • Current conversation (default if no argument provided)
   • A file path passed as argument (e.g., /create-skill ./notes/workflow-idea.md)
   • A description passed as argument (e.g., /create-skill "A skill to run tests and format output")
2. Ask the user using AskUserQuestion:
   • Skill name (propose one based on context, lowercase with hyphens)
   • Brief description (one sentence for the frontmatter)
   • Target location: user-level (~/.claude/skills/) or project-level (./.claude/skills/)
   • Any specific tools or MCP servers the skill should use
3. Determine skill type based on context:
   • Workflow skill: Multi-step process with clear phases
   • Utility skill: Single focused task
   • Integration skill: Interacts with external services via MCP

4. Generate the SKILL.md following the structure below

5. Write the file to the appropriate location. Filename MUST be SKILL.md (uppercase) — the GitHub Pages deploy looks for that exact name.

6. Sync the docs if this is a user-level skill in ~/dev/claude/: run python3 scripts/sync_docs.py to regenerate the skill tables in README.md and skills.md.

7. Remind the user to commit the new skill and the regenerated tables together — CI rejects PRs where the tables drift from the filesystem.

Skill File Structure

---
name: skill-name
description: One sentence describing what this skill does.
# Jekyll nav (only needed for user-level skills published at nicolasrouanne.github.io/claude):
title: /skill-name
parent: Skills
permalink: /skills/skill-name/
---

# Skill Title

Brief overview of what this skill accomplishes.

## Your Task

[Numbered steps for what the skill should do]

1. **First step** - what to do and why
2. **Second step** - continue the workflow
3. **Ask the user** using AskUserQuestion if decisions needed
4. **Execute** the main action
5. **Report results** back to user

## Guidelines

[DO and DON'T lists if applicable]

**DO:**
- Specific behaviors to follow
- Quality standards

**DON'T:**
- Anti-patterns to avoid
- Common mistakes

## Input Handling

[Document how arguments are interpreted]

- No argument: Use current conversation context
- File path: Read and process the file
- Description: Use as the basis for the task

## Tools Used

[List any specific tools, MCP servers, or integrations]

- Tool/MCP name: What it's used for

## Example Usage

/skill-name /skill-name ./path/to/file.md /skill-name “description of what to do”

Guidelines

DO:

• Keep skills focused on a single purpose
• Include clear step-by-step instructions
• Document which tools/MCPs are needed
• Add example usage patterns
• Use AskUserQuestion for user decisions

DON’T:

• Create overly complex multi-purpose skills
• Hardcode values that should be configurable
• Skip the frontmatter (name and description are required)
• Assume tools are available without checking

Placement Decision

User-level (~/.claude/skills/) for:

• Generic utilities useful across all projects
• Personal workflow preferences
• Skills that don’t reference project-specific paths

Project-level (./.claude/skills/) for:

• Project-specific workflows (deploy, release, test)
• Skills that use project paths, scripts, or tools
• Team processes that should be shared via version control

After Creation

Ask the user:

• “Would you like to test the skill now?”
• If user-level: “Would you like to commit and push to your claude repo?”
• If project-level: “Don’t forget to commit this to your project repo.”

Example Prompts

From conversation:

User: I keep having to run these 5 commands every time I deploy
Claude: [discusses the workflow]
User: /create-skill

Creates a skill based on the deployment workflow discussed.

From description:

/create-skill "A skill to summarize meeting notes and extract action items"

Asks clarifying questions, then generates a meeting-notes skill.

From file:

/create-skill ./docs/release-process.md

Reads the document and creates a skill that automates the described process.