HavenSkills
Agent workflow

Skill Creator

The skill creator is an official Anthropic skill for building skills: it drafts a SKILL.md with you, writes test prompts, runs them with and without the skill, compares the results, and rewrites the skill until it behaves. It also optimizes the description, which is the part that decides whether a skill triggers at all.

By Anthropic

What the skill creator actually gets done

It does not write one SKILL.md and stop. It runs a build-and-test loop with you, and the loop is the product. Its own description puts it plainly: create new skills, modify and improve existing ones, measure performance, run evals, and benchmark with variance analysis.

In order, that means:

A capturing interview first. Before any file exists it asks what the skill should do, when it should fire, and what the output looks like — which is the part most people skip and then debug for a week.

A folder, not a file. You get SKILL.md plus the directories it is meant to use: references/ for material the model reads only when needed, scripts/ for deterministic helpers, assets/ for files the output needs.

Test prompts, written before the skill is trusted. The skill generates realistic requests, runs them with the skill loaded and without it, and shows both side by side. That comparison is what turns "seems better" into evidence.

A description rewritten for triggering. The description is the only text the client sees before deciding to load anything, so the loop treats it as the headline rather than a summary — third person, what it does, and the situations that should make it fire.

Iteration until it stops improving. Each round changes one thing — the description, a reference file, a script — and re-runs the same prompts, so variance does not get mistaken for improvement.

When it triggers — and when it doesn't

It fires when the work is about a skill as an artefact: writing a new one, fixing one that never triggers, trimming one that eats too much context, or turning a repeated workflow you keep retyping into a folder.

Reach for something else when:

  • You want a tool the model can call. That is mcp-builder. A skill is instructions in a file; an MCP server is a running service with a protocol. Different artefact, different lifecycle.
  • You want to know whether your product works. That is webapp-testing — it drives a browser. The skill creator's evals test a skill, not an application.
  • Your problem is a single prompt, not a process. A skill earns its keep when context or a procedure has to persist. For a one-off request, the prompt is enough and the folder is overhead.

How to check it's working

Ask it to build a small skill for something you already do by hand, and watch whether it interviews you first. If it starts writing SKILL.md from one sentence, the skill did not load — you are watching the base model's default behaviour.

The stronger check is the eval: ask it to show the same prompt with and without the skill and to state which output is better and why. A loaded skill produces that comparison without being asked twice.

Common failures

The new skill never triggers. The description is too abstract. "Helps with documents" matches nothing specific. Name the artefact, the verbs, and the situations — the client matches the user's request against that text, not against what the skill can actually do.

It triggers constantly and pollutes every conversation. The opposite error: a description broad enough to match ordinary work. Narrow it to the objects and tasks it is for, and move the long material into references/ so it is not loaded unless needed.

The eval says everything is fine. Two runs is not an eval. Ask for several runs and for the variance — a skill that wins once and loses twice is noise, not an improvement.

SKILL.md grows into a small book. Anything over a few hundred lines should live in a reference file. A long SKILL.md is loaded on every trigger and costs context every time.

Two skills fight over the same request. Names and descriptions overlap, so both match. Write the boundary into each description — say what it is for and, where it matters, what it is not for.

How it compares with the other workflow skills

skill-creator makes the artefact. mcp-builder makes a tool server behind a protocol, which is what you need when instructions are not enough. webapp-testing checks a running app rather than a skill. claude-api answers implementation questions about the API underneath — the reference you want before, not after, you build.

The order most people actually need: claude-api when you are unsure what is possible, skill-creator to package a procedure, webapp-testing to check what you shipped.

Try a sample task

This walkthrough gives you a prompt and an acceptance check to run in your own client.

Tune a skill’s trigger

Use a sandbox project with a draft SKILL.md and two prompts that should trigger it plus two close matches that should not.

Use skill-creator to review the draft, propose a clearer trigger description, and show how you would evaluate the positive and negative prompts before changing the original.

What to verify A revised description and a table of expected versus observed trigger decisions. Record the client, model and any unavailable automation.

Best for Reusable workflows, trigger tuning, and comparing a draft skill against example prompts.

How to use skill-creator

Collect representative prompts and decide what an acceptable result looks like before editing instructions.

Acceptance check Review results with and without the skill against the same examples. Check each action item for an owner and deadline when those were supplied.

Install for your agent

Run the command from your project directory. After installation, reload the client if it has already started. Tool access and execution features depend on the client.

Claude Code

npx skills add anthropics/skills --skill skill-creator -a claude-code

Codex

npx skills add anthropics/skills --skill skill-creator -a codex

Cursor

npx skills add anthropics/skills --skill skill-creator -a cursor

GitHub Copilot

npx skills add anthropics/skills --skill skill-creator -a github-copilot
Ask your agent to install the skill
Install skill-creator from anthropics/skills in this project. Show me the destination folder and report any dependencies the source lists.

For manual installation, copy the entire skill folder, including references and scripts, into a skill directory supported by your client. See the installation guide.

What the upstream guide covers

The upstream guide covers skill writing and evaluation in one workflow. It explains how to capture intent, run examples, compare results and improve how the skill triggers.

HavenSkills workflow check

No workflow check recorded yet. Use the sample task above to make your own check in your client.

Review methodology and recorded checks

Requirements, permissions and limits

Runtime
Python for evaluation helpers; available agent runner for automated comparisons
File and tool access
Evaluation helpers can create run artifacts and may invoke nested agents or client-specific tooling. Use a disposable project and review which files and model calls the runner can access.
Limits to check
Trigger results vary across clients and models. Use held-out prompts so a change that works on the training examples can still be challenged.

Frequently asked questions

Is the skill creator free?

Yes, under Apache 2.0 per the repository README. The SKILL.md itself carries no separate licence line.

What if my skill never triggers?

Then the description is the problem far more often than the body. The client matches your request against the description before anything else is read, so rewrite that first.

Is this the same as building an MCP server?

No. It handles instructions and procedure. A tool the model can call needs an MCP server, which is what mcp-builder is for.

What do I get at the end?

A folder with SKILL.md in it, plus whichever reference, script and asset files the skill actually needs — and test prompts you can re-run later.

Does it work in Cursor, Codex or Gemini CLI?

Yes. It runs against the client you have installed; the folder it produces is the same format the terminal clients read.