Skip to main content
plugins/skill-creator/ ships Anthropic’s official skill-creator skill along with its supporting agents, scripts, eval viewer, and tests.

​
Manifest

{
  "name": "skill-creator",
  "version": "0.1.0",
  "description": "Create new Claude Code skills, improve existing skills, and measure skill quality through evals and benchmarks.",
  "author": { "name": "Anthropic", "email": "support@anthropic.com" }
}

​
What’s bundled

plugins/skill-creator/
├── README.md
├── .claude-plugin/plugin.json
└── skills/skill-creator/
    ├── SKILL.md
    ├── LICENSE.txt              # Apache 2.0
    ├── agents/                  # analyzer, comparator, grader subagents
    ├── scripts/                 # 9 Python scripts
    ├── eval-viewer/             # generate_review.py + viewer.html
    ├── references/              # claude-ai.md, cowork.md, schemas.md
    ├── assets/eval_review.html
    └── tests/                   # pytest suite

​
Skill

​
skill-creator

Creates new Claude Code skills and iteratively improves existing ones through eval loops and variance analysis. Trigger phrases: “create a skill”, “turn this into a skill”, “make a skill for X”, “write a SKILL.md”, “run evals on my skill”, “benchmark my skill”, “optimize my skill description”, “improve triggering”, “test my skill”. Workflow (high level):
1

Draft

Decide what the skill should do, then write a draft SKILL.md.
2

Evaluate

Create test prompts, run them with the skill enabled, and review the results both qualitatively and quantitatively.
3

Iterate

Adjust the description and content based on routing failures and output quality. Re-run evals.
4

Package

Use scripts/package_skill.py to bundle the skill for distribution.

​
Subagents

AgentPurpose
analyzerInspect skill structure and surface routing issues
comparatorCompare two skill variants on the same eval set
graderScore eval outputs against rubrics

​
Scripts

ScriptPurpose
init_skill.pyScaffold a new skill directory
package_skill.pyBundle a skill for distribution
quick_validate.pyValidate SKILL.md frontmatter and structure
run_eval.pyRun a single eval against a skill
run_loop.pyRun an iterative improvement loop
improve_description.pySuggest description rewrites for better routing
aggregate_benchmark.pyAggregate benchmark results across runs
generate_report.pyRender an eval report
utils.pyShared helpers

​
Local cleanup

__pycache__/ and .pytest_cache/ are excluded by .gitignore. Do not commit them back.

​
Sources

The Apache 2.0 license is bundled at plugins/skill-creator/skills/skill-creator/LICENSE.txt.
Last modified on April 8, 2026