Claude Skill Builder

Build Claude Code skills following Anthropic's official methodology — from planning through distribution.

PRIMARY SOURCE: "The Complete Guide to Building Skills for Claude" (Anthropic, January 2026)

• PDF: https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf
• Local copy: ./claude-skill-builder-guide.pdf

Quick Start

To create a skill from scratch, follow the 4-phase process below. To improve an existing skill, read references/refining-skills.md. To validate a skill's structure, run scripts/audit-skill.sh.

Invoke with: /claude-skill-builder or ask about creating, improving, or auditing a skill.

How It Works

Four-phase methodology from the Anthropic guide:

Total: ~75-150 minutes for a complete, tested skill

Phase 1 covers Steps 1–2 below; Phases 2–4 correspond to Steps 3–5.

Skill Creation Workflow

Step 1: Discover Requirements

Research the domain first — before writing instructions, verify current best practices for the skill's subject. Use WebSearch and WebFetch to check official docs, community standards, and authoritative examples. Outdated or incorrect guidance in a skill is worse than no guidance: Claude will follow it confidently.

Retain all sources: every URL consulted must be recorded in the skill with its full link and a note of what it confirmed. Without sources, guidance cannot be verified or updated when the domain changes.

For research strategies, source quality standards, and the required Sources section format: references/research.md

1. Identify the problem and target users
2. Define 2-3 concrete use cases
3. Set measurable success criteria (time saved, errors reduced, quality improved)
4. Choose a skill category:

For in-depth category guidance: references/categories.md For interactive discovery questions: references/discovery.md

Step 2: Design Structure

Design using progressive disclosure — 3 loading levels:

For progressive disclosure writing tips and success metrics: references/best-practices.md

Step 3: Implement

Critical Rules:

• ✅ File MUST be named SKILL.md (not README.md)
• ✅ Folder name MUST be kebab-case — no spaces, no capitals (my-skill not My Skill)
• ✅ YAML frontmatter MUST include name, description, and version fields
• ✅ Description MUST include specific trigger phrases — what users SAY to activate the skill
• ✅ Description must be under 1024 characters (hard limit — longer descriptions are truncated)
• ✅ No XML angle brackets (< or >) in any frontmatter field
• ❌ NO README.md inside the skill folder — all docs go in SKILL.md or references/ (Exception: a README.md at the GitHub repo ROOT, outside the skill folder, is fine for GitHub.)
• ❌ NO spaces or underscores in folder names (my_skill → my-skill)

Frontmatter: required fields:

---
name: your-skill-name          # kebab-case only; no spaces, capitals, or underscores
description: What it does. Use when user asks to "specific phrase", "another phrase".
             # MUST include: what it does + when to use it (trigger conditions)
             # Under 1024 characters. No XML angle brackets.
             # Do NOT start with "claude" or "anthropic" (reserved namespaces).
version: 0.1.0                 # Required; use semantic versioning
---

Frontmatter: all optional fields:

---
name: your-skill-name
description: What it does and when. Use when user says "trigger phrase 1", "trigger phrase 2".
license: MIT                   # Optional: open-source license (MIT, Apache-2.0, etc.)
allowed-tools: "Bash(python:*) WebFetch"  # Optional: restrict which tools the skill can use
compatibility: Claude Code     # Optional: 1-500 chars; environment requirements
metadata:                      # Optional: custom key-value pairs
  author: Your Name
  version: 1.0.0               # Version inside metadata (Anthropic's spec); also accepted at top level
  mcp-server: your-server      # If skill requires a specific MCP server
  category: productivity
  tags: [automation, workflow]
  documentation: https://example.com/docs
---

Positioning language (outcome-focused: "generate tests 87% faster") belongs in README.md or GitHub landing page — NOT in the description field. See references/best-practices.md.

Folder structure (standalone skills at ~/.claude/skills/):

~/.claude/skills/your-skill-name/
├── SKILL.md                         # Required — loaded when skill triggers (<5k words)
├── references/                      # Docs Claude loads into context as needed
│   ├── detailed-guide.md            #   schemas, API docs, policies, detailed workflows
│   └── examples/                    #   working code users copy (subdirectory of references/)
│       └── working-example.sh
├── scripts/                         # Executables (run without loading into context)
│   └── validate.sh
└── assets/                          # Files used IN skill output (not loaded to context)
    └── template.html

Note: Plugin skills (in plugin-name/skills/) may place examples/ at the top level — that is plugin-dev convention. For standalone ~/.claude/skills/ skills, put examples inside references/.

To scaffold a new skill: bash scripts/scaffold-skill.sh my-skill-name To start from template: copy references/examples/SKILL-template.md

Step 4: Test

Three testing approaches (run in order):

1. Triggering tests — verify Claude activates on expected phrases; does NOT activate on unrelated requests
2. Functional tests — validate the skill's core workflow produces correct output for known inputs
3. Performance tests — measure improvement over baseline (time saved, error reduction, consistency)

Debugging trigger issues:

Ask Claude: "When would you use the [skill name] skill?"

Claude will quote its description back at you. Adjust based on what's missing or too vague.

Fixing undertriggering: Add more specific trigger phrases and relevant technical terms.

Fixing overtriggering: Add negative triggers to the description:

description: Processes PDF legal documents for contract review. Use for "review this contract",
  "analyze legal document", "extract contract clauses". Do NOT use for general PDF viewing,
  image extraction, or non-legal documents (use doc-converter skill instead).