Meta Create Skill

Use this guide to add a new capability (skill/workflow) that follows Anthropic's "Complete Guide to Building Skills for Claude" playbook. Treat this document as the canonical example—any updates to it must satisfy the same standards described below.

Quality Standards

Structure & Progressive Disclosure

• Keep the three-layer model: concise YAML frontmatter → actionable SKILL.md body → optional linked references.
• Maintain composability; every skill must coexist with others without assuming exclusivity.
• Keep the body under ~500 lines. Move deep dives into references/ or templates/ and link from the main file.

Frontmatter Requirements

• name: prefix__slug folder name, lowercase letters/numbers/hyphens, ≤64 characters, no claude/anthropic.
• description: ≤1024 characters describing what the skill does, when to use it (trigger phrases/file types), and key capabilities in third-person voice.
• Optional fields (license, compatibility, metadata) clarify distribution licensing, runtime expectations, or MCP dependencies. Avoid XML brackets to keep the system prompt safe.

Folder Hygiene

• Directory format from the guide:

  .claude/skills/{prefix__slug}/
  ├── SKILL.md
  ├── templates/      # optional code or doc templates
  ├── references/     # optional deep guidance
  ├── assets/         # optional fonts/icons/mockups
  └── scripts/        # optional automation helpers
  

• Never add README.md inside the skill directory; human-facing docs live elsewhere.
• Reference template files (e.g., templates/Command.cs) instead of inlining large artifacts.

Description & Trigger Checklist

• Lists both positive triggers ("Use when the user says…") and negative guards ("Do not use when…") to reduce under/over-triggering.
• Mentions relevant files/APIs (.fig, Linear, etc.) so Claude can match user intent accurately.
• Passes the self-test: Ask Claude "When would you use the {skill}?" and ensure the echoed description matches your intent.

Instruction Crafting

• Begin with an opening statement summarizing the workflow outcome.
• Present numbered steps with validation gates, decision points, and references to reusable templates.
• Highlight error handling, rollbacks, and quality gates (e.g., iterative refinement loops) consistent with the guide's pattern library.
• Use // turbo callouts for steps that automation can perform without human approval.

Planning Checklist

1. Document 2–3 concrete use cases with trigger phrases, tool requirements (built-in or MCP), and the "definition of done" for each.
2. Map each use case to one of Anthropic's categories: document creation, workflow automation, or MCP enhancement.
3. Establish success criteria before authoring:
   • ≥90% trigger rate on relevant prompts vs. manual enabling.
   • Clear metrics (tool calls, token budget, error rate) to compare with/without the skill.
4. Decide how progressive disclosure will split content between SKILL.md and linked references.

Skill Creation Steps

1. Plan Naming & Scope

   • Reserve a unique prefix__slug and ensure the folder mirrors the name exactly.
   • Capture user outcome statements ("Help me plan this sprint") and negative cases upfront.

2. Author Frontmatter

   • Use templates/SKILL.md as the scaffold.
   • Populate description with outcome + trigger phrases + relevant assets/APIs.
   • Add optional metadata such as compatibility: "Claude Code + MCP-Linear" when the runtime matters.

3. Draft Instructions

   • Follow the recommended structure: overview, numbered workflow, examples, troubleshooting.
   • Link to reusable templates (e.g., templates/Endpoint.cs) instead of pasting long code blocks.
   • Include decision trees, validation steps, and error handling patterns pulled from the Anthropic guide.

4. Add Supporting Resources

   • Create templates/, references/, assets/, or scripts/ as needed; use placeholders ({ResourceName}) for user-supplied values.
   • Keep references one level deep to respect progressive disclosure.

5. Register in AGENTS.md

   • Document the /prefix__slug command with a concise description and placement inside the Skills table.

6. Test & Iterate

   • Design trigger tests (should trigger vs. should not trigger lists), functional tests (end-to-end workflow success), and performance comparisons (token/tool-call deltas) per the guide.
   • Iterate descriptions when you detect under- or over-triggering; add negative triggers or clarifications as needed.

// turbo 7. Automated Verification

• Lint YAML/Markdown, confirm template placeholders resolve, and ensure linked files exist.
• Validate that SKILL.md stays under size limits and references valid paths.

8. Distribute & Document
   • Provide installation guidance (zip folder → upload via Claude.ai Settings > Skills or place in Claude Code skills directory).
   • Position the skill around outcomes ("set up project workspaces in seconds") and highlight the MCP + skill story when applicable.

Testing Playbook

• Trigger Coverage: Maintain a regression list of prompts that must load the skill and another list that must not.
• Functional Runs: Exercise representative workflows ("Create project with 5 tasks") and log API/tool outcomes and error handling paths.
• Performance Deltas: Compare baseline vs. skill-enabled runs for conversation turns, tool invocations, and token counts.
• Feedback Loop: Capture under/over-triggering and execution issues; feed them back into the description or instruction updates.

Distribution & Positioning

• Host skills in a public repo with a README for humans (outside the skill folder) plus screenshots or GIFs.
• Cross-link from your MCP documentation explaining how the skill and connector combine for faster workflows.
• Provide a short installation checklist so users can enable and test the skill immediately after download.

Pattern Reference

• Sequential Workflow Orchestration: Stepwise actions with validation and rollback guidance.
• Multi-MCP Coordination: Phase-by-phase coordination across multiple services, with explicit data handoffs.
• Iterative Refinement: Draft → validate → refine loops with stopping criteria.
• Context-Aware Tool Selection: Decision trees that choose the right MCP/tool based on file type, size, or risk tier.
• Domain-Specific Intelligence: Embed compliance/intelligence rules that run before calling external tools.

Related Skills

Prerequisites:

• Foundational skills such as /meta__cheat_sheet for quick conventions.

Next Steps:

• After scaffolding, use /test__verify_feature or relevant automation to validate broader workflows.

See Also:

• /meta__write_agents_md for documentation guidance.
• Other .claude/skills/ examples for reference implementations.