Skip to content
SkillMixed7,253 tokens in use

skill-creator

Anthropic's skill for making skills: Claude drafts a SKILL.md, runs test prompts with and without it, shows the results in a review viewer and tunes the description that triggers it.

197,764 stars on the repositoryanthropics/skills · skills/skill-creator
/plugin install example-skills@anthropic-agent-skills
View source

Install command as written in the repository's README

Skill label

measured from its files

What installing skill-creator costs

64tokens idle
7,253tokens in use
18,660tokens, up to

Idle is what sits in context while the skill is installed. In use is its instructions, loaded when it runs. Up to adds the files its instructions link. Counts are approximate, made offline with cl100k_base.

Model dependence Mixed

Script-ledMixedModel-led

Scripts do part of the work and the model's judgement does the rest.

This describes the skill's structure, how much of it is scripts. It does not say how well it works. Scripts: eval-viewer/generate_review.py, scripts/__init__.py, scripts/aggregate_benchmark.py, scripts/generate_report.py, scripts/improve_description.py, scripts/package_skill.py, scripts/quick_validate.py, scripts/run_eval.py, scripts/run_loop.py, scripts/utils.py.

Touches

shelldetectedwrites filesdetected

Declared tags come from the skill's frontmatter; detected ones from its scripts and code blocks. This is detected, not audited.

How it scores

Quality

How well the skill is written, counted from its own files: its frontmatter, its steps, examples, stated output and constraints, its size and whether every file it ships is one it names.How we score it →
Excellent
  • Frontmatter12/15
  • Structure25/25
  • Examples15/15
  • Output15/15
  • Size13/15
  • Resources14/15

Security

Starts at 100 and loses points for each thing in the skill's files that a reader should look at first.How we score it →
Excellent
  • Runs shell commands−5

    does this on the machine, and the frontmatter does not say so

    SKILL.md, line 229
  • Writes files−8

    does this on the machine, and the frontmatter does not say so

    SKILL.md, line 229

Repository health

Pre-computed score based on development activity, maintenance, community, maturity, and trend momentum.How we score it →
Good
  • Development activity72
  • Maintenance32
  • Community68
  • Maturity36
  • Momentum40

skill-creator is the agent skill Anthropic publishes for creating new skills and improving existing ones. It works out where the user is in the process and joins there: capturing what the skill should do and when it should trigger, interviewing for edge cases and success criteria, writing the SKILL.md with its frontmatter and bundled scripts, references and assets, and then testing it. It also covers updating an installed skill without renaming it, and packaging the finished folder as a .skill file.

Testing is the centre of the skill. Claude writes two or three realistic prompts, runs each one in a subagent with the skill and in another without it (or with the previous version), grades the outputs against assertions, aggregates pass rate, time and tokens into a benchmark, and opens a browser viewer where the user reads each output and leaves feedback. The feedback drives the next iteration. A separate loop rewrites the skill’s description against a set of should-trigger and should-not-trigger queries, choosing the best one by its held-out score.

What You Get

  • A workflow for writing a skill: intent capture, an interview on edge cases and dependencies, and guidance on the SKILL.md frontmatter, the scripts/references/assets layout and keeping the body under about 500 lines
  • An evaluation loop that runs each test prompt in parallel subagents with the skill and without it (or against a snapshot of the old version), organized into iteration and eval directories in a workspace
  • Grader, blind comparator and post-hoc analyzer agent instructions, plus a schemas reference for evals.json, grading.json, timing.json, benchmark.json and the other result files
  • Scripts that aggregate runs into benchmark.json and benchmark.md and generate a review viewer with an Outputs tab for per-case feedback and a Benchmark tab for the numbers
  • A description optimizer that tests trigger queries through claude -p, proposes improved descriptions and picks the best by its score on a held-out test split
  • A validator and packager that check the frontmatter and zip the skill folder into a distributable .skill file

Common Use Cases

  • Turning a workflow already carried out in the conversation into a reusable skill
  • Writing a new skill from scratch and checking with test prompts that it actually improves on Claude without it
  • Improving an existing skill and comparing the new version against a snapshot of the old one
  • Fixing a skill that does not trigger when it should, or triggers when it should not, by optimizing its description
  • Packaging a finished skill as a .skill file for someone else to install

What Makes It Unique

The other Anthropic skills listed here teach Claude a task, such as writing a doc, building a deck or calling the Claude API; this one teaches Claude to make and measure skills themselves. Where doc-coauthoring checks a document by handing it to a fresh Claude reader, skill-creator turns that idea into an experiment: every test prompt runs with and without the skill, assertions are graded, and the difference is reported as a benchmark, with a human review step in a browser viewer before Claude revises anything. It also treats the frontmatter description as something to tune empirically, with a train and held-out split, and adapts its method for Claude Code, claude.ai and Cowork, where subagents, a browser or the claude CLI may be missing.

Under The Hood

SKILL.md is a long set of instructions, with agent prompts for grading, blind comparison and analysis in agents/ and the result file formats in references/schemas.md. Claude creates a <skill-name>-workspace folder beside the skill and spawns subagents for the with-skill and baseline runs, saving their outputs, timing and grading files there. The bundled Python scripts do the mechanical parts. aggregate_benchmark.py reads those files and writes benchmark.json and benchmark.md. eval-viewer/generate_review.py starts a local web server on 127.0.0.1 (port 3117 by default), first stopping whatever process is already listening on that port through lsof, opens it in the browser and writes the user’s feedback to feedback.json; with —static it writes a standalone HTML file instead. For description optimization, run_eval.py writes a temporary command file into the project’s .claude/commands folder and runs the claude CLI with claude -p on each query to see whether the skill is invoked, removing the file afterwards; improve_description.py also calls claude -p to propose new descriptions, and run_loop.py writes and opens an HTML report. Both run the CLI with the user’s own Claude login, with the CLAUDECODE variable removed so it can run inside a Claude Code session, so these runs use the user’s Claude usage. quick_validate.py checks the frontmatter with PyYAML, and package_skill.py zips the skill folder into a .skill file. The trigger eval review page is a bundled HTML template that Claude fills in, writes to a temporary file and opens.

Join founders building with open source

Opinionated takes, migration guides, cost-saving tips, and insights from the open source ecosystem.

Subscribe on SubstackJoin 750+ subscribers