Skill Creator

About

A skill is a folder containing SKILL.md (required instructions with YAML frontmatter) and optional bundled resources. Skills teach Claude how to handle specific tasks or workflows, eliminating the need to re-explain preferences, processes, and domain expertise in every conversation.

Skills are powerful when you have repeatable workflows — generating consistent documents, conducting research with a standard methodology, orchestrating multi-step processes, or providing workflow guidance on top of MCP integrations.

Use Case Categories

Skills fall into three common categories:

Category 1: Document & Asset Creation — creating consistent, high-quality output (documents, presentations, code, designs).

• Key techniques: embedded style guides, template structures, quality checklists
• No external tools required; uses Claude's built-in capabilities

Category 2: Workflow Automation — multi-step processes that benefit from consistent methodology.

• Key techniques: step-by-step workflow with validation gates, iterative refinement loops
• This skill (skill-creator) is a Category 2 example

Category 3: MCP Enhancement — workflow guidance layered on top of an MCP server's tool access.

• Key techniques: multi-MCP call coordination, embedded domain expertise, error handling for MCP failures

Anatomy of a Skill

skill-name/
├── SKILL.md         # Required — instructions + YAML frontmatter
├── scripts/         # Optional — executable code (Python, Bash, etc.)
├── references/      # Optional — documentation loaded into context as needed
└── assets/          # Optional — templates, fonts, icons used in output

SKILL.md (required)

The YAML frontmatter is how Claude decides whether to load your skill. Get it right.

Minimal required format:

---
name: your-skill-name
description: What it does. Use when user asks to [specific phrases].
---

Full example with optional fields:

---
name: pdf-rotator
description: Rotates and reorients PDF pages. Use when the user says "rotate PDF", uploads a .pdf and asks to fix page orientation, or needs portrait/landscape conversion.
license: MIT
compatibility: Requires Python 3.9+ and pypdf
metadata: {version: "1.0.0", author: YourName, tags: [pdf-processing, file-manipulation]}
---

Description field structure: [What it does] + [When to use it] + [Key capabilities]

Good descriptions — specific, trigger-focused:

# Good — includes trigger phrases users would actually say
description: Manages Linear project workflows including sprint planning and task creation. Use when user mentions "sprint", "Linear tasks", "project planning", or asks to "create tickets".

# Good — mentions relevant file types
description: Analyzes Figma design files and generates developer handoff docs. Use when user uploads .fig files or asks for "design specs" or "design-to-code handoff".

Bad descriptions — vague or missing triggers:

# Bad — too vague
description: Helps with projects.

# Bad — missing triggers
description: Creates sophisticated multi-page documentation systems.

Security restrictions:

• Never include XML angle brackets (< >) — YAML is injected into Claude's system prompt; malicious content can hijack instructions
• Never use claude or anthropic in the name field (reserved words)
• Do not embed executable code or dynamic expressions in YAML values

Bundled Resources (optional)

scripts/

Executable code for tasks requiring deterministic reliability or repeated execution.

• Include when the same logic is rewritten across conversations or exact accuracy is required
• Code is deterministic; language instructions are not — prefer scripts for critical validations
• Reference with exact commands: python scripts/validate.py --input {filename}

references/

Documentation loaded into context as needed.

• Include schemas, API docs, domain knowledge, policies, workflow guides
• For files exceeding 10k words, add grep/search patterns in SKILL.md to guide context loading
• Keep references one level deep — no nested subdirectories
• Avoid duplicating content between SKILL.md and reference files

assets/

Files used in output, not loaded into context.

• Include templates, images, icons, boilerplate code, fonts, sample documents
• Example: assets/report-template.md, assets/logo.png, assets/frontend-template/

Core Design Principles

Progressive Disclosure — Three loading levels minimize token usage:

1. YAML frontmatter — always loaded; enables Claude to decide when to trigger the skill
2. SKILL.md body — loaded when skill is relevant; contains full instructions (< 5k words)
3. Bundled resources — loaded only as needed; scripts can execute without being loaded

Composability — Claude can load multiple skills simultaneously. Design skills to work alongside others rather than assuming exclusive context.

Portability — Skills work identically across Claude.ai, Claude Code, and the API without modification. Document required dependencies in the compatibility field.

Skill Creation Process

Follow the steps in order; skip only when clearly not applicable.

Step 1: Define Use Cases

Identify 2–3 concrete use cases before writing any instructions. For each use case, answer:

• What does the user want to accomplish?
• What multi-step workflow does this require?
• Which tools are needed (built-in, scripts, or MCP)?
• What domain knowledge or best practices should be embedded?

Trigger-example pairs:

• Trigger: "Rotate this PDF 90 degrees." → Expected: run scripts/rotate_pdf.py, return updated file.
• Trigger: "Create a dashboard starter in React." → Expected: copy boilerplate from assets/, explain how to run.
• Trigger: "Summarize this vendor contract." → Expected: load references/policies.md, produce structured summary.

Define success criteria before building:

• Quantitative: skill triggers on 90%+ of relevant queries; workflow completes in target tool calls; 0 failed API calls per workflow
• Qualitative: users don't need to redirect Claude; consistent results across sessions; new users succeed on first try

Step 2: Plan Skill Contents

Identify the skill's category (Document & Asset Creation, Workflow Automation, or MCP Enhancement) and list reusable resources:

1. Walk through each use case and identify what scripts, references, or assets it needs
2. Flag steps requiring deterministic accuracy → bundle as scripts
3. Flag knowledge reusable across conversations → bundle as references

Step 3: Initialize the Skill

Run the initialization script for new skills:

python scripts/init_skill.py <skill-name> --path <output-directory>

The script creates the skill directory with a SKILL.md template and example resource directories. Remove unused example files after initialization.

Step 4: Write the Skill

Implement Resources First

Build scripts/, references/, and assets/ before finalizing SKILL.md. Request user-provided assets or documentation explicitly before proceeding.

Write SKILL.md Body

Answer these questions to complete SKILL.md:

1. What is the purpose of the skill?
2. When exactly should it trigger? (Include specific phrases users would say.)
3. What are the step-by-step instructions?
4. How should bundled resources be invoked?

Writing style:

• Frontmatter: third-person ("This skill should be used when...")
• Body: imperative, verb-first ("Fetch", "Validate", "Create" — not "You should" or "Remember to")
• Put critical instructions at the top; mark must-follow steps with CRITICAL:
• Reference scripts with exact commands: python scripts/validate.py --input {filename}
• Reference documentation explicitly: "Consult references/api-patterns.md for rate limiting guidance before writing queries."
• Move detailed docs to references/; keep SKILL.md focused on core workflow

Include error handling:

## Common Issues

### [Error Message]
Cause: [Why it happens]
Solution: [How to fix]

Step 5: Test

Triggering Tests

Run 10–20 queries to verify when the skill loads:

Debug under-triggering: ask Claude "When would you use the [skill name] skill?" — it quotes the description back. Adjust based on missing keywords.

Debug over-triggering: add negative triggers to the description:

description: Advanced data analysis for CSV files. Use for statistical modeling. Do NOT use for simple data exploration (use the data-viz skill instead).

Functional Tests

Verify the skill produces correct outputs:

• Confirm valid outputs are generated for the defined use cases
• Confirm tool/MCP calls succeed without errors
• Confirm error handling works for defined edge cases

Step 6: Validate and Package

Run the validator before packaging:

python scripts/validate_skill.py <path/to/skill>

Then package into a distributable zip:

python scripts/package_skill.py <path/to/skill-folder>

# Optional output directory:
python scripts/package_skill.py <path/to/skill-folder> ./dist

The packaging script runs validation then creates dist/<skill-name>.zip.

Common validation errors:

• Description missing trigger keywords → add "use/when" phrasing
• Name not kebab-case or doesn't match folder name → rename both
• Unsupported frontmatter keys → only name, description, license, compatibility, allowed-tools, metadata are allowed
• SKILL.md exceeds 500 lines → move content to references/

Step 7: Distribute

Host the skill on GitHub with a public repo and a clear README at the repo level. Do not put a README.md inside the skill folder — all documentation goes in SKILL.md or references/.

For MCP-enhanced skills, add a section to your MCP documentation that:

• Links to the skill
• Explains the value of using MCP + skill together
• Provides a quick-start example prompt

Step 8: Iterate

Skills are living documents. Iterate based on real usage signals:

Iterate on a single challenging task until Claude succeeds, then extract the winning approach into the skill. This gives faster signal than broad testing.

Validation Rules

Run python scripts/validate_skill.py <path/to/skill> to check all rules before packaging.

Best Practices

• Triggers in description: the description is the only content always in context — make it precise with specific phrases users would say
• Progressive disclosure: keep SKILL.md focused on core workflow; move detailed docs to references/
• Specificity: avoid vague descriptions — include actions, file types, and concrete trigger phrases
• Determinism: for steps where exact accuracy matters, bundle scripts rather than relying on language instructions
• No duplication: put content in SKILL.md or references/, not both
• No time-sensitive content: avoid timestamps or "as of [date]" phrasing
• Flat references: keep references/ one level deep — no nested subdirectories
• No README inside skill folder: all documentation goes in SKILL.md or references/
• Negative triggers: if the skill over-triggers, add "Do NOT use for..." to the description
• Composability: assume other skills may be active simultaneously — don't write instructions that conflict with general Claude behavior

Template Usage

• Use templates/basic.md for instruction-only skills (no scripts needed)
• Use templates/advanced.md for skills that bundle scripts or MCP workflows
• Customize placeholders and remove unused sections after initializing

Additional Links

• Templates: templates/basic.md, templates/advanced.md
• Example skill: ../test-skill/SKILL.md
• Skill creator skill: https://github.com/ComposioHQ/awesome-claude-skills/tree/master/skill-creator
• Public skills repository: https://github.com/anthropics/skills