Skill Creator

Existing Skills (for pattern reference)

!ls .claude/skills/ 2>/dev/null

---

Process

Step 1: Gather Requirements

From the user's description, determine:

1. Name: lowercase-with-hyphens, becomes the /slash-command
2. Purpose: What does this skill do?
3. Trigger: When should someone use it?
4. Input: What arguments does it need?
5. Output: What should it produce?
6. Tools needed: Which Claude Code tools does it require?

Step 2: Generate SKILL.md

Create the file following this template:

---
name: skill-name
description: "One sentence: what it does + when to use it. Be specific enough for auto-invocation."
argument-hint: "[argument description]"
allowed-tools: Read, Grep, Glob
---

# Skill Title

## Context Detection
!`relevant shell command for auto-detecting project context`

---

## Instructions

### Step 1: [First action]
Detailed instructions for what Claude should do.

### Step 2: [Second action]
...

### Step N: [Final action]
...

## Output Format
Describe the expected structure of the output.

## Rules
- Rule 1
- Rule 2

Step 3: Validate the Skill

Check the generated SKILL.md against these requirements:

• Frontmatter has name and description (required)
• Description is under 150 characters and specific
• allowed-tools lists only what's needed (principle of least privilege)
• Shell injection uses !command`` syntax correctly
• Instructions are numbered steps, not vague paragraphs
• Output format is defined
• Rules section exists with boundaries

Step 4: Save and Test

1. Create directory: .claude/skills/<skill-name>/
2. Save SKILL.md to that directory
3. Instruct user to test: open Claude Code, type /<skill-name>

Frontmatter Reference

Field	Required	Description
name	Yes	Slash command name (lowercase, hyphens)
description	Yes	When to use (specific, <150 chars)
argument-hint	No	Autocomplete hint, e.g. [filename]
allowed-tools	No	Tools allowed without asking
disable-model-invocation	No	true = manual only
user-invocable	No	false = hidden from menu
model	No	Override model
effort	No	low, medium, high, max
context	No	fork for subagent isolation

Rules

• Follow existing skill patterns in this project
• Keep SKILL.md under 200 lines
• Use shell injection for context detection, not hardcoded assumptions
• Description must be specific enough for Claude to auto-invoke correctly