create-skill

name	create-skill
description	Create, edit, review, evaluate, package, and optimize skills. Use when users ask to create a skill, revise skill instructions, review a skill, run skill evals, benchmark skill performance, package a skill, or optimize a skill description for trigger accuracy.
license	Apache-2.0
tags	["creator","skills","authoring"]
metadata	{"author":"Anthropic","version":"1.9.0","source":"github.com/anthropics/skills","catalog":"utility","category":"meta"}

create-skill

Create new skills, review and improve existing skills, evaluate outputs, optimize trigger descriptions, and package final skill folders.

---

Workflow

1. Route the request. Read only the reference that matches the user's current task:

   User intent	Read
   Create a new skill or revise skill instructions	references/authoring.md
   Review a created or revised skill	references/review.md
   Build eval cases, run iterations, benchmark outputs, or collect human feedback	references/evaluation.md
   Optimize a skill description for trigger accuracy	references/description-optimization.md
   Adapt the workflow for agents without subagents, Claude Code, generic CLIs, or Cowork	references/agent-compatibility.md
   Validate eval YAML or grading, benchmark, and feedback JSON structures	references/schemas.md

   If the request spans multiple phases, read the references in workflow order: authoring, review, evaluation, description optimization, then agent compatibility only when platform details matter.

2. Clarify activation and behavior. Identify what the skill should do, which user phrases or contexts should trigger it, what output it should produce, and whether objective evals are useful.

3. Write or revise the skill. Name new skills using the <verb>-<subject>[-<variant>] convention or a concise <verb> format (e.g., code-tests, ask). Follow references/authoring.md for metadata, trigger descriptions, SKILL.md body format, reference file format, section delimiters, scan anchors, examples, helper scripts, portability, and validation. Always bump metadata.version using semantic versioning upon any material change to a skill's files.

4. Test behavior. Run this skill's scripts/validate.py against the target skill when available. For router skills, confirm routed eval cases use expect.routing.reference; for objectively testable skills, run skill-enabled outputs against a meaningful baseline.

5. Show evidence. Share validation output, eval results, benchmark summaries, and relevant diffs before making another revision.

6. Iterate deliberately. Continue until feedback is resolved or further changes stop improving behavior.

7. Package last. Package the final skill only after the user is satisfied with behavior and trigger accuracy.

---

Output

• For skill edits: summarize changed files, behavior changes, and validation results.
• For reviews: lead with correctness, trigger, structure, safety, and test coverage findings.
• For evals or benchmarks: report the command, dataset or eval set, pass/fail counts, variance notes, and recommended next change.
• For packaging: report the generated artifact path and any remaining manual checks.

---

Boundaries

• Use the reference as source of truth. Do not duplicate detailed authoring, review, evaluation, schema, or compatibility rules in this file.
• Keep skills focused. Center each skill on one trigger and one workflow; route broad domains through references instead of expanding the main body.
• Avoid placeholders. Add scripts/, references/, assets/, or evals/ only when the skill actually uses them.
• Package after behavior. Do not package a skill before the user is satisfied with behavior and trigger accuracy.

---

Bundled Resources

• Trigger optimization: scripts/run_eval.py, scripts/run_loop.py, and scripts/improve_description.py
• Validation: scripts/validate.py
• Benchmark summaries: scripts/aggregate_benchmark.py
• Packaging: scripts/package_skill.py
• Human review UI: eval-viewer/generate_review.py
• Review agents: agents/grader.md, agents/comparator.md, agents/analyzer.md, and agents/benchmark-analyzer.md

---

Verification

• The selected reference matches the user's current task
• SKILL.md and reference edits follow references/authoring.md
• Validation, eval, benchmark, or packaging commands were run when applicable
• Results and remaining risks are reported to the user