Skill Creator

Skill Creator

Create effective skills through a structured 6-step process.

Core Principles

1. Context window is shared - Only include what Claude doesn't already know. Challenge each paragraph: "Does this justify its token cost?"
2. Progressive disclosure - Metadata (~100 words, always loaded) -> SKILL.md body (<5k words, on trigger) -> references/ (as needed, unlimited)
3. Match freedom to fragility - Narrow bridge = specific guardrails (scripts), open field = flexible guidance (text)
4. Concise examples over verbose explanations - Show, don't tell

Skill Structure

skill-name/
├── SKILL.md          # Required - YAML frontmatter + Markdown instructions
├── scripts/          # Optional - Executable code (deterministic, token-efficient)
├── references/       # Optional - Docs loaded into context as needed
└── assets/           # Optional - Templates, fonts, icons used in output

No README.md, CHANGELOG.md, or auxiliary documentation inside the skill folder.

Creation Process

Step 1: Understand Use Cases

Ask the user for 2-3 concrete examples:

• "What would a user say to trigger this skill?"
• "What multi-step workflows does this require?"
• "Which tools are needed (built-in or MCP)?"
• "What domain knowledge or best practices should be embedded?"

Skip only when usage patterns are already clear.

Step 2: Plan Reusable Contents

For each example, identify reusable resources:

• scripts/: Code rewritten repeatedly (e.g., rotate_pdf.py, validate_fields.py)
• references/: Documentation needed during execution (e.g., schema.md, api-docs.md)
• assets/: Files used in output (e.g., template.pptx, boilerplate/)

Step 2b: Determine Scope (Global vs Project-Local)

Ask the user or determine from context where the skill should live:

Decision questions:

• "Does this skill reference project-specific services (Supabase, Cloud Run, specific APIs)?" → Project-local
• "Would this skill work identically in any TypeScript/Python/Go repo?" → Global
• "Is this a team policy that should apply to all members?" → Organization

Why this matters: Global skills count toward the per-session limit (recommended 20-50). Over-installing globally leads to context window pressure and description conflicts. When in doubt, start project-local and promote to global only after confirmed reuse.

Step 3: Initialize

Run the initialization script to scaffold:

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

Skip if updating an existing skill.

Step 4: Edit

4a. Write Frontmatter

The description is the primary trigger mechanism. Follow this structure:

[What it does] + [When to use it] + [Key capabilities]

CRITICAL rules:

• Include specific trigger phrases users would say
• Include negative triggers ("Do NOT use for...")
• Under 1024 characters, no XML angle brackets
• Only name and description are required
• Optional fields: license, allowed-tools, metadata, compatibility

See references/description-guide.md for good/bad examples and the full checklist.

4b. Implement Resources

1. Create scripts, references, assets identified in Step 2
2. Test scripts by actually running them
3. Delete example files not needed (init creates placeholders)

For scripts: ensure deterministic reliability, test with real inputs. For references: if >10k words, include grep search patterns in SKILL.md. For assets: verify file paths and formats work as expected.

4c. Write Instructions

• Imperative form: "Run X", "Check Y" (not "You should run X")
• Specific and actionable: Run python scripts/validate.py --input {filename} not "Validate the data"
• Critical info first: Use ## Critical or ## Important headers at the top
• Under 500 lines: Move details to references/
• Error handling: Include what to do when things fail
• Examples: Add input/output pairs for complex operations
• No duplication: Information lives in SKILL.md OR references, not both
• Conditional attention: Wrap task-specific sections in <important if="condition"> tags so Claude focuses on them only when relevant. Use specific conditions (e.g., if="you are deploying to production"), not broad ones. Leave universal rules unconditional. See references/attention-tags.md for details and examples.

Pattern guides for reference:

• Multi-step processes: See references/workflows.md
• Output formats: See references/output-patterns.md

Step 5: Validate and Package

python scripts/package_skill.py <path/to/skill-folder> [output-dir]

Validates automatically, then creates a .skill file. Fix errors and re-run if validation fails.

The validator checks: frontmatter format, naming conventions, description quality, body length, file structure, and common mistakes.

Step 6: Iterate

Test with real tasks. Watch for signals:

• Undertriggering (doesn't load when it should) -> Add more trigger phrases to description
• Overtriggering (loads for irrelevant queries) -> Add negative triggers, narrow scope
• Instructions not followed -> Move critical info to top, use bullet points, be more specific
• Inconsistent results -> Add examples, improve error handling

See references/testing-guide.md for the full testing approach. See references/troubleshooting.md for common issues and fixes.

Performance Notes

• Take your time to do this thoroughly
• Quality is more important than speed
• Do not skip validation steps