Code Review & Testing · IN-DEPTH PROFILE

skill-creator

Create, improve, and benchmark Agent Skills with trigger tests and comparative evaluation.

View listed source ↗Installation guide ↓
Best for

Authors maintaining reusable Agent Skills

What you get

Design the Skill + Build trigger evals

Main limitation

Meaningful benchmarks require representative prompts, repeat runs, and careful human judgment.

First risk

Untrusted evaluation prompts and artifacts. Test prompts, example files, generated scripts, and review artifacts can contain unsafe instructions or sensitive data. Isolate runs and inspect every artifact before execution or sharing.

EVIDENCE FRESHNESS

Three checks, kept separate

A recent source check is not a runtime test or security audit.

Upstream sourceChecked 2026-08-20

Pinned revision · 0a64e398

Open pinned commit ↗
SkillSignal profileEditorial metadata

Updated 2026-08-20

Runtime & securityNot independently verified

Source review does not certify behavior or safety.

30-SECOND BRIEF

What it does—and when it fits

Skill Creator is Anthropic's end-to-end workflow for creating, improving, and evaluating Agent Skills. It combines a concrete user interview, progressive-disclosure structure, trigger tests, baseline comparisons, grading, benchmark aggregation, and qualitative review.

INSTALL BY AGENT

Choose your Agent

Paths come from official Agent docs or the universal installer behind skills.sh. Compatibility still follows this Skill's record.

Native

This Skill's current record explicitly names this Agent. Still inspect scripts, permissions, and external dependencies first.

Project scope.claude/skills/skill-creator/
Personal scope~/.claude/skills/skill-creator/

Use project scope for team sharing and personal scope across repositories. The installer defaults to project scope; add -g for personal scope.

Install command (project scope)npx skills add anthropics/skills --skill skill-creator --agent claude-code
Official agent docs

Claude Code discovers custom Skill folders automatically at project or personal scope.

View path evidence ↗

TYPICAL WORKFLOW

A practical workflow

01

Design the Skill

Turn examples and edge cases into scoped instructions and resources.

02

Build trigger evals

Create should-trigger and should-not-trigger prompts for the description.

03

Benchmark iterations

Compare skill-enabled runs with baselines and review variance.

THE TRADEOFFS

Advantages and tradeoffs

Notable strengths

  1. Treats triggering and task performance as behaviors to evaluate, not assumptions.
  2. Encourages progressive disclosure instead of oversized SKILL.md files.

Limitations

  1. Meaningful benchmarks require representative prompts, repeat runs, and careful human judgment.
  2. Some optimization and blind-comparison paths depend on CLIs, subagents, or viewer tooling that may be unavailable.

BEST FIT

Who it is for

→

Authors maintaining reusable Agent Skills

→

Teams evaluating trigger quality and workflow impact

BEFORE YOU USE IT

Risks to review before use

High

Untrusted evaluation prompts and artifacts

Test prompts, example files, generated scripts, and review artifacts can contain unsafe instructions or sensitive data. Isolate runs and inspect every artifact before execution or sharing.

SECURITY

What the permission profile means

  • Run generated scripts and candidate Skills in disposable workspaces.
  • Do not place secrets or private production data in eval prompts or benchmark artifacts.

Not a security certification. External ratings are attributed references. SkillSignal has not independently executed or security-reviewed this Skill.

INSIDE THE PACKAGE

Indexed files

SKILL.mdUpstream package contentSource-linked
agents/Upstream package contentSource-linked
references/Upstream package contentSource-linked
scripts/Upstream package contentSource-linked
assets/Upstream package contentSource-linked

TAGS

skill-authoringevalsbenchmarking

Original SkillSignal editorial profile grounded in the current SKILL.md and linked files at pinned Anthropic commit 0a64e398, checked 2026-08-20; not independently executed or security-certified.