agent-skills/create-skill/SKILL.md

name	description	license	metadata
create-skill	Creates, rewrites, reviews, and improves concise, unambiguous, progressively disclosed, and agent-actionable Agent Skills. Use this skill when creating, reviewing, or revising a SKILL.md, its routing description, references, assets, scripts, or reusable workflow guidance. Do not use for installing third-party skills or capturing one-off preferences that do not warrant a reusable skill.	MIT	short-description Create and evaluate concise Agent Skills

Create Skill

Required resources and checks

• Before using a starter, read and adapt assets/skill-template.md. Replace every sentinel and remove unused sections.
• Before finalizing or reviewing, read references/skill-quality-checklist.md. Record applicable evidence.
• Run repository-provided skill validators when they exist.
• Otherwise perform equivalent checks manually. Report the unavailable gate.

Follow user and higher-priority constraints. Keep one coherent job. Preserve behavior outside an evidence-backed requested change.

Scale the work

Match evidence and validation to the requested outcome:

• For a review, inspect and report without editing or running execution evaluations unless requested or needed to verify a material risk.
• For a narrow revision, make the requested change and run applicable static checks.
• For creation or substantial revision, define relevant use cases, make the requested changes, and run applicable static checks.
• When evaluating or claiming improvement, test representative behavior and compare with the previous version when feasible.

Do not turn a narrow review into an unsolicited rewrite or benchmark.

Skill contract

• Put all activation guidance in frontmatter description. The body loads only after activation.
• Use: [Third-person capability]. Use this skill when [intents/triggers]. Do not use when/for [material near misses].
• Keep the description under 1024 characters. Focus on intent, outcomes, and adjacent boundaries—not mechanics.
• Keep SKILL.md concise and actionable. Put critical constraints, mandatory resources, fragile sequencing, validation, and safety early.
• Default to short headings, compact bullets, numbered procedures, and small tables. Use prose paragraphs only when they communicate necessary nuance more clearly.
• Keep each bullet focused on one instruction or tightly coupled behavior. Prefer direct wording; remove throat-clearing, repeated rationale, and paragraph-length bullets.
• Match the surrounding skill catalog's density. A longer main file requires execution-critical justification; move conditional detail into references/.
• Use progressive disclosure deliberately. Do not squeeze useful reusable detail into SKILL.md merely to avoid a resource.
• Use references/ for detailed workflows, schemas, examples, tool guidance, and gotchas.
• Use assets/ for reusable templates and static output shapes.
• Use scripts/ for repeated or fragile deterministic operations.
• Add resources when they improve repeated execution, clarity, consistency, or maintenance.
• Keep references one level deep.
• Add a resource pointer only after its file exists. Give its exact path, purpose, and read/use/run condition.
• Repeat mandatory resource loading at the relevant workflow step. Delete stale pointers and empty directories.
• Refer to other skills by name, such as create-plan. Do not use relative paths to their SKILL.md files.
• Ground changes in relevant evidence: user requirements or corrections, project conventions, static inspection, real tasks or artifacts, recurring failures, or evaluation results. Match the strength of evidence to the claim.
• Reject one-off or speculative preferences. Prefer the smallest durable correction.
• Check tools, schemas, commands, and examples for staleness. Ask when reusability is unclear.
• Prescribe fragile, destructive, security-sensitive, and validation-critical sequences. Allow judgment elsewhere.

Workflow

1. Define observable use cases

For creation, substantial revision, or behavioral evaluation:

• Record two or three realistic tasks.
• For each, record inputs, expected output, success criteria, and constraints.
• Record two or three strong near misses. Name the correct adjacent skill or direct workflow.
• For an existing skill, record reusable change evidence and the smallest durable correction.
• Do not broaden scope because related guidance is available.

2. Inspect the ecosystem

• Read the target skill and project instructions.
• Inspect nearby global and project skills, conventions, scripts, tools, and schemas.
• Identify duplicated policy and resource boundaries.
• Reuse good patterns. Do not copy generic advice or another skill's whole workflow.

3. Draft frontmatter first

For creation or a requested routing change:

• Write the description before the body.
• Include realistic user phrasing, useful co-activation, and material near misses.
• Use lowercase kebab-case for name. Keep it under 64 characters and match the directory.

4. Build the post-activation contract

For creation or a requested body or resource change:

• Include only rules needed after activation.
• Define critical constraints, sequence, resources, validation, failure handling, and safety.
• Remove redundant explanation, excessive menus, stale examples, vague guidance, and unrelated content.
• Perform a compression pass: shorten wording, split overloaded bullets, convert scannable prose to lists, and remove content that does not change execution.
• Create supporting resources when detail is reusable or would crowd the main file.
• Follow the target repository's established language, runtime, dependency, packaging, and test conventions. Do not introduce a runtime solely for a helper script when existing project tooling is sufficient.
• Make command-line scripts automation-safe: non-interactive by default, bounded output, meaningful exit statuses, and safe idempotent defaults.
• Provide --help or the ecosystem-equivalent. Document prerequisites, working directory, inputs, outputs, and side effects.
• Send primary or machine-readable output to stdout and diagnostics to stderr where the environment supports that convention.
• For stateful or destructive operations, default to a dry run or require an explicit opt-in; add confirmation where appropriate.

5. Evaluate routing and execution

When evaluating behavior, finalizing a substantial skill, or claiming improvement:

• Test three to five realistic should-trigger prompts and two to three strong near misses.
• Judge routing from the description alone. Expand cases only for broadly distributed or high-risk skills.
• Run one representative task with realistic input. Compare its output with the recorded success criteria.
• Prefer a fresh, safely available subagent for unbiased execution or review.
• Otherwise record the direct fallback and its independence limit. Treat child claims as evidence, not acceptance.
• Revise only material evidence-backed gaps. Rerun affected routing and output checks once.
• Do not build a benchmark framework for a small skill.

6. Validate and finalize

Run applicable checks from the repository root using its documented commands and established runtime.

• Run an available Agent Skills specification validator when the repository provides or permits one.
• Otherwise check specification compliance manually and report the unavailable automated gate.
• Run each added or changed command-line script's --help or ecosystem-equivalent, focused repository-native tests, and one representative safe execution.
• Before finalizing, read and complete references/skill-quality-checklist.md against the final revision.
• Report changed files, evaluated cases, exact results, skipped checks, reasons, and remaining risks.
• Never claim resources or artifacts that do not exist.