claude/for
Search
Subscribe
Claude CodeUpdated

Skill of the day: `skill-creator`

Radar: anthropics/skills skill-creator inside example-skills — draft, eval, iterate, package Agent Skills. Install example-skills@anthropic-agent-skills; not a solo plugin.

skill-creator is an official Anthropic marketplace skill shipped inside the example-skills plugin. It helps you create new Agent Skills, improve existing ones, run with-skill vs baseline evals, optimize the SKILL.md description for triggering, and package a .skill for install/share. Register anthropics/skills, install example-skills@anthropic-agent-skills, then ask to draft a skill, run evals, or tune a description. Marketplace metadata version: 1.0.0 (catalog field — not per-skill semver). Checked 2026-09-27.

TL;DR: Meta-skill for authoring + measuring skills — not a one-off task helper. Install the example-skills plugin (path ./skills/skill-creator). Use for reusable workflow capture and eval loops. Skip Office binaries (document-skills), API/docs primary (claude-api radar), or structured doc co-writing (doc-coauthoring). [GAP]: there is no solo /plugin install skill-creator@….

What it does

SKILL.md describes an end-to-end loop:

  1. Capture intent — what the skill should do, when it should trigger, expected output.

  2. Draft SKILL.md (progressive disclosure: metadata always loaded; body on trigger; scripts/ / references/ / assets/ as needed).

  3. Write test prompts → run with-skill and baseline (parallel subagents in Claude Code).

  4. Review — qualitative viewer (eval-viewer/generate_review.py) + quantitative benchmarks (aggregate_benchmark, grader/comparator/analyzer agents).

  5. Iterate on feedback → optional description optimization (run_loop / improve_description) → package (.skill).

Bundled tooling is richer than a guidance-only sibling: scripts/ (run_loop, run_eval, aggregate_benchmark, package_skill, …), eval-viewer/, agents/ (grader, comparator, analyzer), references/schemas.md, assets/eval_review.html. Description field = primary trigger; SKILL.md nudges “pushy” descriptions to fight under-triggering.

Triggers: “make a skill”, “improve this skill”, “run evals”, “optimize skill description.”

When to use / when not

Use when you want a reusable skill from a workflow you just ran, eval-driven improvement (with-skill vs baseline), or description-trigger tuning.

Do not:

  • One-off task with no reusable skill intent

  • Primary deliverable is Office binaries → document-skills

  • Claude API / SDK docs primary → live claude-api radar — do not re-Intake

  • Long structured doc co-writing → live doc-coauthoring — do not re-Intake

  • Pure UI aesthetic direction → sibling frontend-design in the same plugin (ship-ready; live radar still 404 — prose only, no href)

  • MCP server authoring → sibling mcp-builder (draft-ready; live radar still 404 — prose only, no href)

Install (Claude Code)

/plugin marketplace add anthropics/skills

Then Browse → anthropic-agent-skills → example-skills → Install, or direct:

/plugin install example-skills@anthropic-agent-skills

Fact (marketplace.json, 2026-09-27): plugin name is example-skills; skill path is ./skills/skill-creator. Sibling skills in the same plugin include algorithmic-art, brand-guidelines, canvas-design, doc-coauthoring, frontend-design, mcp-builder, skill-creator, theme-factory, web-artifacts-builder, webapp-testing, and others — installing pulls the whole example set.

Walkthroughs: add the marketplace · sync Claude.ai skills / plugins into Claude Code · skill-doctor if load looks wrong.

Also available context (one line): Claude.ai paid plans already include many example skills — this radar stays Code/marketplace-first.

Limits

  • README disclaimer: skills are demo/reference; product behavior may differ

  • Installs the whole example-skills bundle — not a solo skill plugin

  • Full eval/baseline loop prefers Claude Code (subagents); Claude.ai path is adapted/simplified; description optimizer needs claude -p / scripts

  • Does not replace your own review judgment or a security review of skill contents

Pair with prior radars / guides

Guide / radarMoment
doc-coauthoringSibling example-skills — structured doc co-writing
document-skillsDeliverable is xlsx / docx / pptx / pdf
add anthropics/skills marketplaceFirst /plugin setup
sync skills / pluginsClaude.ai ↔ Code
skill-doctorSkill not loading
claude-apiOptional prior C

Sources (checked 2026-09-27)

One Claude move in your inbox, every Sunday

Four minutes, tested on a real job, then back to your weekend. Free.