How to use skill-creator

Anthropic's official skill for creating, evaluating, and iteratively optimizing Claude Code skills, from interview to benchmarked production.

A skill written once without evaluation tends to drift, firing when you do not want it to, failing silently on edge cases, or going stale when Claude Code updates. skill-creator runs a full RED→DRAFT→GREEN→REFACTOR→OPTIMIZE iterative loop: capture intent, structured interview, write draft, run quantitative evals via background subagents, review results in the eval-viewer, refactor against failures, then run the description optimizer to improve trigger accuracy. Non-technical users are explicitly supported. The skill is designed to work with anyone willing to open a terminal.

L3L4L5

Before you startClaude Code must be installed. You need at least one skill already installed or written. Skill-creator evaluates existing skills, not a blank starting point. If you have never written a skill, start with the [Anthropic Skills Library](/how-to/anthropics-agent-skills).

Steps

Install skill-creator from the Anthropic skills repo

Clone the Anthropic skills repo and copy the skill-creator folder into your `~/.claude/skills/` directory. Alternatively, use `claude plugin install` if skill-creator appears in your marketplace. Verify by running `/skill-creator --help` in a Claude Code session.

git clone https://github.com/anthropics/skills cp -r skills/skills/skill-creator ~/.claude/skills/ # Verify in a Claude Code session: /skill-creator --help

Understand the quality cycle before writing anything

skill-creator uses five stages: RED (write failing evals), DRAFT (write the skill), GREEN (all evals pass), REFACTOR (simplify without breaking evals), OPTIMIZE (tighten trigger and output). A skill is only done at GREEN. Anything before that is a draft.

Run the lightweight skill creation prompt before the formal cycle

Before you set up evals, ask Claude to interview you and walk through the first run. This is the no-code, no-eval version: it gives you a working skeleton you can then put through the RED-GREEN-OPTIMIZE cycle below. Useful for getting a skill from idea to first invocation in under fifteen minutes.

Create a reusable skill called weekly-recap. When I run it, it should: gather what I worked on this week from the files and notes I point it at, write a short summary grouped by project, and list anything still unfinished. Before you build it, ask me any questions you need so the skill matches how I actually work. After you build it, show me how to run it and walk me through what happens the first time.

Write evals before the skill: the RED phase

Write three eval cases before touching the skill content: one where the skill fires and produces the right output, one where it handles a tricky edge case, and one where it should stay silent. If you cannot write these three cases, you do not yet know what the skill should do.

I want to build a skill called [SKILL_NAME] that does: [SKILL_DESCRIPTION] Before writing the skill, help me define three evaluation cases: 1. Canonical case: skill fires and produces [EXPECTED_OUTPUT] 2. Edge case: skill fires and handles [EDGE_CASE] correctly 3. Non-trigger case: skill stays silent when [NON_TRIGGER_SCENARIO] Format each as: input context / expected behavior / grading criterion.

[SKILL_NAME]Your skill's name e.g. db-schema-reviewer

[SKILL_DESCRIPTION]What it does e.g. Reviews SQL schema files for normalization issues, missing indexes, naming inconsistencies

[EXPECTED_OUTPUT]What correct output looks like e.g. a structured report grouped by severity: critical / warning / suggestion

[EDGE_CASE]Tricky input the skill must handle e.g. a schema with no issues, the skill should output 'No issues found', not hallucinate problems

[NON_TRIGGER_SCENARIO]Case where the skill should stay silent e.g. a TypeScript file that contains a SQL string literal, not a schema file

Run evals and iterate to GREEN

Run your draft skill against all three eval cases. Any failing eval tells you exactly what to fix. Fix the skill, never the eval. When all three pass, you are at GREEN and the skill is ready to use.

/skill-creator eval [SKILL_NAME]

[SKILL_NAME]Your skill's name e.g. db-schema-reviewer

Run batch evals after any Claude Code update

When Claude Code updates, skills can break silently. Run all your skills through skill-creator's batch eval mode after any major version bump. Skills that fail after an update are telling you what changed.

/skill-creator eval-all

How do you know it worked?

`/skill-creator --help` runs in a Claude Code session without errors.
You wrote three eval cases for a skill before writing the skill itself.
Your skill reached GREEN, all three evals pass.
You ran `eval-all` and can interpret the pass/fail results.

If something looks off

Writing the skill before the evals

RED-before-DRAFT exists for a reason. A skill written without evals is defined by what it happens to produce, not by what you need it to produce. Those two diverge fast.

Fixing evals to make the skill pass

Evals are the contract. If the skill fails, fix the skill. Weakening an eval to get a pass lowers the bar instead of raising the skill.

Never running evals after Claude Code version bumps

Claude Code changes how it interprets skills as it updates. Run eval-all after any major version bump. Regressions surface automatically instead of surprising you mid-session.

Skills & Commands Reference

skill-creator is a single skill with a defined process. These are the stages you move through each time you use it. Understanding the stage helps you know what to provide and what to expect.

Command / Skill	What it does	When to use it
`Capture Intent`basic	Understand what the skill should do, when it should trigger, and what output format it produces	Start here every time, even if you have a draft, confirm intent before evaluating
`Interview & Research`basic	Probes edge cases, input/output formats, example files, success criteria, and dependencies	Before writing any test prompts, the more you answer here, the less you iterate later
`Write Draft`basic	Produces the initial SKILL.md following the Superpowers skill format	After intent and research are solid, agent writes, you review
`Test & Evaluate`basic	Runs the skill across test prompts in background Claude sessions; surfaces quantitative metrics and qualitative output	After every draft change, never ship a skill that hasn't been tested against real prompts
`Iterate`basic	Rewrites the skill based on evaluation results, targets specific failure modes, not general improvement	When test results show gaps, be specific about what failed and why
`Optimize Description`advanced	Runs the description improver script to maximize triggering accuracy, makes the skill fire when it should and not when it shouldn't	Final step before shipping any skill, poor descriptions cause skills to silently miss their trigger
`eval-viewer`advanced	HTML viewer for browsing quantitative eval results, shows pass/fail per test case with qualitative output	After any eval run, faster than reading raw output, helps identify which specific cases failed
`/plugin marketplace add anthropics/skills`	Install the Anthropic skills library as a plugin marketplace source in Claude Code	To access skill-creator and all official Anthropic skills via plugin install