skill-creator: Anthropic's skill for writing, editing and evaluating skills

When to use

• You want to write a skill from scratch or improve an existing one, with a loop of draft, run test cases, review results, revise.[1]
• After a skill is written, you want to check whether its description triggers when it should and tune it with the official script (needs the claude CLI).[1]
• You want the SKILL.md frontmatter and folder layout checked against the Claude Code docs.[6]
• Also usable for a simple skill without evals: the skill says the user can skip a bunch of evaluations, so the full eval loop can be left out (a step limit, not a whole-skill non-use case).[1]

When not to use

• You only want to use an existing skill for an everyday task, not to create, modify or evaluate a skill: the description covers creating new skills, modifying and improving existing ones, and measuring performance. The official text does not list non-use cases; this item is an inference from the description scope and the task boundary.[1]

What it does

The official description: create new skills, modify and improve existing skills, and measure skill performance; use it to create from scratch, edit or optimize a skill, run evals, benchmark with variance analysis, or optimize a description for triggering accuracy.[1]

• Decide what the skill should do and write a draft.
• Prepare a few test prompts and run Claude with access to the skill on them; for a new skill also run each without the skill as a baseline, and when improving an existing skill use the old version (snapshot taken before editing) as the baseline.
• Use eval-viewer/generate_review.py to show results to the user, alongside quantitative metrics.
• Rewrite the skill from feedback, repeat until satisfied, then expand the test set.
• After the skill is done, optionally run the description-improvement script; if the environment has the present_files tool, package the skill into a .skill file with package_skill.

[1]

Step limits (not whole-skill non-use): the user may skip evals; the description-tuning step needs the claude CLI (claude -p) and is skipped without it; Claude.ai has no subagents, so cases run one at a time with baselines and quantitative benchmarking skipped. These only shrink the workflow; the rest of the skill stays usable.[1]

Skill file format, per the Claude Code docs

[6]

my-skill/
├── SKILL.md (required - overview and navigation)
├── reference.md (detailed API docs - loaded when needed)
├── examples.md (usage examples - loaded when needed)
└── scripts/
    └── helper.py (utility script - executed, not loaded)

[6]

The skill-creator SKILL.md says "YAML frontmatter (name, description required)", its quick_validate.py returns "Missing 'name' in frontmatter" when name is absent, and it advises keeping SKILL.md under 500 lines with bulk material in scripts/, references/ and assets/. The Claude Code docs say "All fields are optional" and name defaults to the directory name. These are two sources each making its own statement: the former is skill-creator's writing and validation rule, the latter the Claude Code documentation; how the two relate is not explained officially: to be confirmed.[1][8][6]

How to call and install it

[1][6][4][5][2][3]

/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

[4][5]

Dependencies and gaps

• Python: aggregation, description tuning and packaging run as python -m scripts.<name>. SKILL.md says the packaging script "works anywhere with Python and a filesystem"; checking the pinned scripts adds: package_skill.py first calls quick_validate.py, which does import yaml (the third-party PyYAML package, not in the Python standard library), so packaging also needs yaml to be importable; the official text does not say how to install PyYAML.
• claude CLI: description tuning (run_loop) needs claude -p. The Claude.ai section says it is "only available in Claude Code"; the Cowork section says it "should work in Cowork just fine" via a claude -p subprocess. The two statements differ, so it is conditioned on having the claude CLI: to be confirmed. run_eval.py and improve_description.py both call claude -p, and a comment in improve_description.py says it uses the session's Claude Code auth with no separate ANTHROPIC_API_KEY.
• Temporary file: run_eval.py writes a temporary command file into .claude/commands/ under the project root (found by walking up to a .claude directory) and removes it in a finally block; the account running description tuning needs write access to that folder (inference).
• Browser: the eval viewer opens in a browser by default; without a display, use --static to write a standalone HTML file.
• Subagents and baselines: each test case spawns one run with the skill and one baseline. For a new skill the baseline is no skill; when improving an existing skill it is the old version snapshotted before editing. Claude.ai has no subagents; the skill gives a sequential fallback and skips baselines.
• Account, permission, cost: account type and permission needs are not stated officially (description tuning reuses the session auth, see above); no cost amount is given officially. Several subagent runs and claude -p calls use up quota, which is an inference from the described workflow.

[1][7][8][9][10]