AI Skills Catalog
AI Workers

Skill Creator

Anthropic’s Skill Creator helps draft SKILL.md workflows, compare trial outputs with baselines, collect human feedback and refine when an agent should use a skill.
SKILL SCORE88How it is scored
VerifiedFreeApache-2.0

Works with

Packaged from the source skill for Claude, Codex and 4 more.

About this skill

Skill Creator is Anthropic's workflow for authoring and improving reusable agent skills. It covers more than writing a SKILL.md file: the process connects the intended task, representative test prompts, comparison outputs and human feedback so changes can be judged against concrete examples.

The package includes instructions for graders, analysts and blind comparisons; JSON schemas; Python evaluation and packaging tools; and a browser review viewer. An author can start with a new workflow or revise an existing skill while retaining its identity.

Use this skill when

  • A repeated task needs explicit inputs, steps, output requirements and activation criteria.
  • An existing skill produces inconsistent results and needs a controlled revision.
  • You want to compare results with and without a skill, or compare a revised version against the previous version.
  • A skill activates too broadly or fails to activate for relevant requests.

What it can do

  • Capture the task's intent and organize instructions, reference material, scripts and assets into a skill directory.
  • Prepare realistic test prompts and assertions that can be checked against generated outputs.
  • Guide paired trial runs and aggregate pass rates, timing and token measurements when the host supplies those measurements.
  • Build a local or static review page so a person can inspect outputs and leave feedback.
  • Support optional blind comparisons and iterative description testing with positive and near-miss queries.
  • Validate basic frontmatter and package a finished skill as a .skill archive.

Typical workflows

  1. Describe the repeated task, expected inputs, required output and situations where the skill should activate.
  2. Draft or revise the skill and choose a small set of representative test cases.
  3. Run comparable trials, retain outputs and grade only criteria that can be supported by evidence.
  4. Inspect the review viewer, collect feedback and revise the instructions where the examples reveal a broader problem.
  5. Repeat with a separate iteration directory, then validate and package the version you choose.

Inputs and output

Useful inputs include the current workflow or skill directory, example files, representative user requests and clear success criteria. Outputs can include a skill directory, evaluation JSON, generated artifacts, benchmark summaries, a review page and a distributable archive.

Limitations

The evaluation workflow needs an agent host capable of running the requested tasks. Its description-optimization scripts invoke an installed, authenticated claude command, and the frontmatter validator requires PyYAML. Those dependencies are not bundled in the catalog download. Trial runs can use model capacity and create local files; review the scope before running them. The interactive viewer uses localhost port 3117 and attempts to stop a process already listening on its selected port; static review output avoids starting that server. The upstream instructions adapt parts of the workflow for hosts without subagents or a browser, so every feature is not available in every client. A small benchmark does not establish general reliability, and the catalog has not executed the bundled evaluation scripts.

What is inside the skill

Section headings from the original SKILL.md (about 4,848 words of instructions). Read the full text in the source repository before installing.

  1. Communicating with the user
  2. Creating a skill
  3. Capture Intent
  4. Interview and Research
  5. Write the SKILL.md
  6. Skill Writing Guide
  7. Writing Style
  8. Test Cases
  9. Running and evaluating test cases
  10. Step 1: Spawn all runs (with-skill AND baseline) in the same turn
  11. Step 2: While runs are in progress, draft assertions
  12. Step 3: As runs complete, capture timing data
  13. Step 4: Grade, aggregate, and launch the viewer
  14. What the user sees in the viewer
  15. Step 5: Read the feedback
  16. Improving the skill
  17. How to think about improvements
  18. The iteration loop
  19. Advanced: Blind comparison
  20. Description Optimization
  21. Step 1: Generate trigger eval queries
  22. Step 2: Review with user
  23. Step 3: Run the optimization loop
  24. How skill triggering works

Bundled files

  • LICENSE.txt
  • agents/analyzer.md
  • agents/comparator.md
  • agents/grader.md
  • assets/eval_review.html
  • eval-viewer/generate_review.py
  • eval-viewer/viewer.html
  • references/schemas.md
  • 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

Example prompts

  • Turn our weekly CSV reconciliation process into a SKILL.md skill. Specify the required inputs, matching rules and exception report, then propose three realistic test cases including a missing customer ID and a duplicate transaction.
  • Improve this existing release-notes skill without changing its name. Compare the current and revised versions on the same three change logs, record factual omissions and formatting errors, and show the outputs side by side before choosing a version.
  • Our accessibility-review skill triggers on unrelated backend tasks. Draft a balanced set of positive queries and near-miss negatives, explain which errors each case detects, and prepare the description-evaluation workflow without running it until I approve the evaluation scope.

How to install Skill Creator

Claude Packaged
  1. Download the Claude package from the skill page. The ZIP contains one folder named after the skill, with SKILL.md and any bundled reference files.
  2. For Claude Code, unpack the folder into ~/.claude/skills/ to use it in every project, or into .claude/skills/ inside a repository to share it with your team.
  3. For Claude apps, upload the ZIP in the Skills section of the capabilities settings.
  4. Start a new session and describe the task in your own words. Claude picks the skill when the request matches its description.
More about skills for Claude
Codex Packaged
  1. Download the Codex package. The ZIP contains one folder named after the skill, with SKILL.md and bundled resources.
  2. Unpack the complete folder into .agents/skills/ inside your repository, or into ~/.agents/skills/ for personal use. Preserve its relative reference-file paths.
  3. Start a new Codex session and invoke the skill by name or with $skill-name, or describe a matching task.
  4. AGENTS.md is separate project guidance. The ZIP does not generate or modify that file.
More about skills for Codex
GPT Packaged
  1. Download the GPT package from the skill page.
  2. Create or edit a custom GPT (or a ChatGPT project) and paste the contents of SKILL.md into the instructions field.
  3. Upload the bundled reference files as knowledge so the GPT can consult them.
  4. Test with one of the example prompts from the skill page and adjust the instructions if your use case differs.
More about skills for GPT
Gemini Packaged
  1. Download the Gemini package from the skill page.
  2. For a project, unpack the complete folder into .gemini/skills/ or .agents/skills/. For personal use, use ~/.gemini/skills/ or ~/.agents/skills/.
  3. Run /skills reload in Gemini CLI, then /skills list to check discovery. There is no need to paste the skill into GEMINI.md.
  4. Test a matching task and verify the referenced tools and files are available.
More about skills for Gemini
Cursor Packaged
  1. Download the Cursor package from the skill page.
  2. Unpack the complete skill folder into .cursor/skills/ or .agents/skills/ in your project. For personal local use, ~/.cursor/skills/ and ~/.agents/skills/ are also supported.
  3. Start a new Cursor session. Type / in Agent chat and select the skill name, or describe a matching task.
  4. Commit the skill folder and resources for team use. Personal local skills are not automatically available to remote agents or automation runs.
More about skills for Cursor
OpenCode Packaged
  1. Download the OpenCode package from the skill page.
  2. Unpack the complete folder into .opencode/skills/ in your project, or ~/.config/opencode/skills/ for personal use. OpenCode also supports .agents/skills/ and .claude/skills/ compatibility locations.
  3. Keep supporting files at their original relative paths. AGENTS.md is separate project-wide guidance and is not generated by this ZIP.
  4. Start OpenCode in the project and test a matching task with the model and permissions you use.
More about skills for OpenCode

Permissions & security

Low risk

Inspect the source, license and declared permissions before use. Read the scripts before running them. Review method

Read project files
Write generated skill files

Skill Creator FAQ

What does the Skill Creator skill do?

Anthropic’s Skill Creator helps draft SKILL.md workflows, compare trial outputs with baselines, collect human feedback and refine when an agent should use a skill.

Which AI agents can use Skill Creator?

Packages for Claude, Codex and 4 more are built from the same source instructions.

How do I install Skill Creator in Claude?

Download the Claude package from the skill page. The ZIP contains one folder named after the skill, with SKILL.md and any bundled reference files. For Claude Code, unpack the folder into ~/.claude/skills/ to use it in every project, or into .claude/skills/ inside a repository to share it with your team.

Who made Skill Creator and what is the license?

Skill Creator is published by Anthropic under the Apache-2.0 license. The source is https://github.com/anthropics/skills/tree/main/skills/skill-creator. The catalog keeps the original attribution.

What permissions does Skill Creator need?

Declared permissions: read project files, write generated skill files. Risk level: low. Check the scripts in the source before running them.

Skill Creator Skill for Claude & Codex | AI Skills Catalog