Create Agent Skill

Create Agent Skill

This skill helps create concise, maintainable agent skills that future agents can reliably use. Treat a skill as reusable operating instructions for a recurring task, not as a transcript of one solution.

Workflow

1. Clarify the goal only as much as needed.

   • Identify the task the skill should help agents perform.
   • Ask for missing requirements only when a reasonable default would be risky.
   • Infer defaults when the user has already provided the purpose, scope, target location, or examples.
   • State assumptions briefly before implementing them.

2. Inspect local conventions before writing.

   • Look for existing skills in the requested skill root, commonly ~/.agents/skills for global agent skills.
   • Follow the target repository's ownership policy before editing a repository skill. Use tracked and ignored status, repository instructions, lockfiles, and other local evidence to distinguish source from runtime installs. Leave ownership unresolved when the available evidence is ambiguous.
   • Match the local file names, frontmatter style, optional metadata, resource folders, and validation tools already in use.
   • Do not invent a different skill format when a local convention exists.
   • Load documentation-engineering for skill-document quality. For a requested or final review, also load code-review and review-verification-protocol.

3. Define the skill contract.

   • Name the users, tasks, inputs, outputs, and success criteria.
   • List activation triggers: user phrases, file types, domains, tools, or workflows that should load the skill.
   • List non-goals so the skill does not silently expand into unrelated work.
   • Decide whether optional scripts/, references/, or assets/ resources are genuinely useful. Do not add empty or placeholder resource folders.

4. Choose the directory name.

   • Use lowercase letters, digits, and hyphens only.
   • Prefer short, verb-led names such as create-agent-skill, fix-ci-checks, or summarize-incident.
   • Keep the folder name and frontmatter name identical.
   • Avoid vague names such as helper, workflow, or assistant.

5. Write the skill.

   • Put required trigger information in the frontmatter description; the body loads only after the skill is selected.
   • Keep the body procedural and reusable. Prefer checklists, decision rules, and compact examples over long explanations.
   • Use imperative instructions that tell the agent what to do.
   • Include constraints, assumptions, and non-goals when they affect behavior.
   • Include examples only when they clarify activation, output quality, or common mistakes.
   • Avoid project-local paths, private repository assumptions, secrets, personal names, or one-off details unless the user explicitly wants a private local skill.

6. Validate and refine.

   • Confirm the required files exist and the directory lives in the requested skill root.
   • Confirm frontmatter parses as YAML and includes at least name and description.
   • Confirm the description explains both what the skill does and when it should activate.
   • Confirm first-party resource files are linked from SKILL.md, local Markdown links resolve, and there are no draft placeholders, stale examples, or unnecessary files.
   • Run any local validator or skill CLI if available. If not, perform a manual structure and content review.
   • Test the skill mentally against at least two realistic user requests: one that should trigger it and one that should not.

Minimal Structure

Follow the local convention if it differs, but a simple SKILL.md-based skill usually looks like this:

skill-name/
`-- SKILL.md

Use this frontmatter pattern unless the local format requires additional fields:

---
name: skill-name
description: What the skill helps with. Use when the user asks for specific trigger scenarios, task types, file types, or workflows.
---

Add optional resources only when they carry reusable value:

skill-name/
|-- SKILL.md
|-- scripts/      # deterministic commands or helpers the agent can run
|-- references/   # detailed docs loaded only when needed
`-- assets/       # templates or files used in generated output

Instruction Quality Rules

• Make instructions specific enough to guide behavior, but not so narrow that they only solve one prompt.
• Prefer "inspect existing files before editing" over "always use file X" unless the file is guaranteed by the skill.
• Prefer "ask if the target system is unknown" over asking for every detail up front.
• Use examples to distinguish good and bad behavior, not to restate the workflow.
• Avoid ambiguous verbs like "handle", "support", or "optimize" unless followed by concrete actions.
• Remove setup history, author notes, release notes, and general README content from the skill.

Good and Bad Patterns

Good description:

description: Create or update GitHub pull request review responses. Use when the user asks to address review comments, inspect unresolved PR threads, implement requested changes, or summarize reviewer feedback.

Bad description:

description: Helps with GitHub.

Good workflow instruction:

1. Fetch the PR metadata and unresolved review threads.
2. Separate actionable code-change requests from discussion-only comments.
3. Implement the smallest change that satisfies each actionable thread.
4. Run the relevant tests and summarize unresolved risks.

Bad workflow instruction:

Do a great job and make the PR better.

Good scope control:

Do not rewrite unrelated modules while addressing review comments unless the review specifically requires it.

Bad scope control:

Improve anything that looks wrong.

Maintenance Checklist

Before finishing a new or updated skill, verify:

• The skill name, folder name, and frontmatter name match.
• The frontmatter description contains activation guidance.
• The body is concise, procedural, and free of one-off context.
• Project names, modules, paths, recipes, frameworks, and validation commands come from target-repository evidence; examples use portable placeholders and do not encode source-project or machine-local assumptions.
• Optional resources are referenced from SKILL.md and are actually needed.
• Examples are realistic and transferable.
• Constraints and non-goals are explicit where they prevent misuse.
• The skill can be understood by a future agent without reading this creation conversation.
• Repository-maintained inventory or routing docs are updated when they exist and the skill set changes.
• Repository-specific review and validation guidance has been applied when working on first-party skills.