Claude skill creator: build a valid SKILL.md in the browser

Write the file

Fill in the two required fields and the body. The preview on the right is the file you will download, byte for byte, and the checks under it are the constraints the specification and Anthropic's authoring guidance state.

---
name: reviewing-pull-requests
description: Reviews a pull request against the team's engineering conventions and separates blocking problems from suggestions. Use when the user asks for a code review, opens a pull request, or asks whether a change is ready to merge.
license: Apache-2.0
---

# Reviewing pull requests

## When this applies
The user asks for a review of a diff, a branch, or an open pull request.

## Steps
1. Read the whole diff before commenting on any part of it.
2. Sort every finding into Blocking, Worth fixing, or Optional.
3. Quote the exact line for each finding and say what to do instead.
4. Say so explicitly when a change is correct and needs nothing.

## Conventions this team holds
- Database access goes through the repository layer, never inline SQL.
- A new public function ships with its test in the same commit.
- Generated files are read for intent rather than line by line.

## Output
A Markdown review with the three sections above, in that order.

reviewing-pull-requests/
  SKILL.md

01 / Format

What makes a SKILL.md valid

Six fields, two of them required. The specification states each limit below; the last column is why the field changes what an agent does with your skill.

For the authoring decisions behind the fields, including the body budget and how to test whether a skill fires, read how to write a SKILL.md file.

02 / Limits

What this tool checks, and what it cannot

The checks are worth exactly what they cover. Stating the boundary is more useful than implying the file is certified because a form went green.

It checks the frontmatter, not the skill

The name rules, the two character limits, the shape of the metadata map, and the YAML quoting are all mechanical, so they can be checked while you type. Whether the instructions are any good is not, and nothing on this page claims to know.

It cannot tell you whether the skill will trigger

Triggering is measured rather than inspected. The published method is a set of around twenty realistic prompts, half that should fire the skill and half near-misses that share its vocabulary, run several times, with part of the set held back so the description is tuned rather than memorized. This page can warn that a description never says when to use the skill. It cannot tell you how often it fires.

It is not the reference validator

The specification points at one validator, the skills-ref library, and that project states in its own README that it is for demonstration purposes and not meant for production use. There is no official validator that certifies a skill. Running skills-ref validate on the downloaded folder is still worth doing, and it checks the same frontmatter rules this page does.

It quotes YAML only when it has to

A description with a colon and a space in it, or one that reads as a number or as the word no, is written as a double-quoted scalar so it parses back as the string you typed. Everything else is left plain, the way the published examples write it. This matters because malformed frontmatter does not error loudly: the body loads with empty metadata and the skill quietly stops triggering on its own.

Nothing leaves the browser

There is no API route behind this page and no model call. The draft lives in the tab until you close it.

For eight real SKILL.md files read line by line, including where the published examples diverge from the specification, see skill examples.

03 / The official one

Anthropic ships a skill creator of its own

The term belongs to a skill before it belongs to any website. anthropics/skills publishes skill-creator, and it does things a static page cannot. Here is what it actually does, read from its own SKILL.md, so you can tell which of the two you want.

It interviews you first

The skill opens by capturing intent: what the skill should let Claude do, which user phrases should trigger it, the expected output format, and whether test cases make sense for this kind of skill. It is told to pull answers out of the conversation you already had, when the request is to turn that work into a skill.

It runs the draft against test prompts

For each test case it launches two runs in the same turn, one with the skill and one without, saves the outputs, grades them against assertions, and aggregates pass rate, time, and token counts into a benchmark. You review the outputs and the numbers in a viewer it generates, and your written feedback goes back into the next revision.

It optimizes the description as a measured loop

A separate script builds twenty trigger queries split between should-trigger and deliberate near-misses, splits them into a training set and a held-out test set, runs each query several times to get a trigger rate, proposes new descriptions, and picks the winner by the held-out score rather than the training score.

It packages the folder

A packaging script turns the directory into an installable file, and that path accepts only the six specification fields, which is the rule the unexpected-key error enforces.

It needs a Claude session to do any of that

It is a skill, so it runs inside Claude Code, claude.ai, or Cowork, and its own file records which parts are unavailable in each: no subagents on claude.ai means no baseline runs and no benchmark, and the description optimizer needs the Claude Code command line.

For what Anthropic publishes in total, across the pre-built document skills, the folders in anthropics/skills, and the ones bundled with Claude Code, see Anthropic skills: the first-party catalog.

04 / Questions

Frequently asked questions

What is the skill creator skill?

skill-creator is a skill Anthropic publishes in the anthropics/skills repository. Loading it turns a Claude session into an authoring workflow: it asks what the skill should do and when it should trigger, writes a draft, runs test prompts with the skill and without it, grades the outputs, and can run a loop that tunes the description for triggering. Because it is a skill rather than a website, it runs wherever Claude reads skills, and it needs a session each time.

Is this the same as the Claude skills creator from Anthropic?

No. Two different things with similar names. Anthropic's skill-creator is interactive and runs inside Claude, and its value is the interview and the evaluation loop. The tool on this page is a generator: it writes the frontmatter, checks it against the published limits while you type, and hands you the folder. It never calls a model, so it costs nothing and answers instantly, and it cannot do the parts that need a running agent.

Does this write the instructions for me?

No, and that is on purpose. It scaffolds the structure and validates the fields that decide whether the skill loads. The body has to come from a task you have already completed. Anthropic's own authoring guidance is direct about it: a skill generated from general knowledge produces vague procedure such as handle errors appropriately rather than the API patterns, edge cases, and project conventions that make a skill worth loading at all.

Will the generated SKILL.md work outside Claude Code?

Yes. The generator writes only the six frontmatter fields the Agent Skills specification defines, which is exactly the set claude.ai uploads, the Skills API, and packaging with package_skill.py accept. Any other key fails those three paths with an unexpected-key error naming allowed-tools, compatibility, description, license, metadata, and name. Claude Code accepts all six as well, so the same folder loads there without changes.

What should the description actually say?

Both what the skill does and when to use it, in the words a person would type. The two primary sources disagree on voice: Anthropic's best practices require third person and name second person as an anti-pattern, while the specification site recommends imperative phrasing such as use this skill when. The published examples reconcile them by stating the capability in third person and then adding a Use when clause, which is the shape the example in the tool follows.

Where does the downloaded folder go?

The download gives you the directory the specification requires, with SKILL.md inside it and the directory name matching the name field. Where that directory goes depends on the agent, and each one documents its own locations. The install guide lists the documented paths for Claude Code, from personal and project skills to plugin and managed ones, and how to confirm a skill actually loaded.

Is my draft uploaded anywhere?

No. The generator is a script that runs in this page, there is no API route behind it, and no model is called. The draft is not posted to a server, stored in an account, or used for anything else, and reloading the page clears it.

How does my team keep the skill after I write it?

Commit the folder to a repository first, then save it to your team's library on Skills Board. Skills Board is the web app where a team keeps and shares its AI skills: the saved entry records the repository and path it came from, and a teammate can open the source, copy a compatible install command, or download the latest files as a ZIP. An MCP-compatible agent can search the same library. It is free, MIT licensed, and it does not write or edit SKILL.md for you.

Sources

Editorial method: every constraint the tool enforces and every claim on this page comes from the first-party documentation below, fetched on the date beside each entry. Where two sources disagree, both are named. Where nothing is documented, the page says so instead of filling it in.

• Agent Skills specification

  The directory layout, the six frontmatter fields with their required status and limits, the full name constraints including the match with the parent directory and the ban on consecutive hyphens, the 1024-character description cap, the 500-character compatibility cap, the experimental marking on allowed-tools, the progressive disclosure tiers, and the skills-ref validate command. Checked 25 August 2026.

• Claude Code: skills documentation

  The twenty frontmatter fields Claude Code accepts, the fourteen it adds on top of the specification, the table restricting claude.ai uploads, the Skills API, and package_skill.py packaging to the six specification fields, and the verbatim unexpected-key error those paths raise. Checked 25 August 2026.

• anthropics/skills: the skill-creator skill

  Read from its own SKILL.md: the intent capture questions, the paired with-skill and baseline runs, the grading and benchmark aggregation, the generated review viewer, the twenty-query trigger eval set with its train and held-out test split, the packaging script, and the per-surface notes on what is unavailable in claude.ai and Cowork. Checked 25 August 2026.

• Anthropic: skill authoring best practices

  The third-person rule for descriptions and the second-person anti-pattern, gerund naming, the reserved words anthropic and claude in a name, the statement that neither name nor description may contain XML tags, the 500-line body guidance, and the instruction to build evaluations before writing more documentation. Checked 25 August 2026.

• skills-ref reference library

  The validate command the specification points at, and the project's own statement that the library is intended for demonstration purposes and is not meant to be used in production. Checked 25 August 2026.