Skill Creator Pro

A meta-skill that creates, tests, and ships agent skills. Install it, describe what you need, get a production-ready .skill file back.

Install

# Claude Code
cp -r skill-creator-pro ~/.claude/skills/skill-creator-pro

# Cursor / Codex / Gemini — or all at once
./scripts/install.sh skill-creator-pro

Use

> "Make a skill that writes postmortem reports from Slack threads"
> "Turn this conversation into a skill"
> "Improve my skill" + paste SKILL.md
> "My skill doesn't trigger — fix it"

Claude handles the rest: interviews you about edge cases, drafts the SKILL.md, tests it, iterates, optimizes the description for triggering, packages it.

What It Produces

A complete skill folder:

your-skill/
├── SKILL.md              # Instructions with frontmatter, workflow, examples, edge cases
├── references/            # Templates, lookup tables, output examples
├── scripts/               # Helper scripts if needed
└── assets/                # Fonts, icons, templates

Ready to install on Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Copilot, and 10+ other tools.

What Makes Skills Built With This Better

• Edge Case Discovery — systematically probes for failure modes by skill type before writing
• Freedom-to-fragility — matches instruction rigidity to output criticality (exact templates where it matters, loose guidance where it doesn't)
• Reference file pattern — extracts real approved outputs into templates the model copies instead of improvising
• Common Mistakes with WHY — model follows reasoning better than rules
• Minimal input testing — tests with both rich and sparse input to catch hallucination on vague prompts
• Self-refinement — ## Learned section captures corrections so the skill improves with use
• Description optimization — automated loop (Claude Code) or guided manual process (claude.ai) to get triggering from ~20% to ~90%

What's Inside

Component	Purpose
references/gold-standard.md	Design principles, quality checklist, 5 anti-patterns
references/patterns.md	15 proven patterns (safety rails, lookup tables, tunable params, stale detection...)
references/example-skill.md	35-line complete skill to copy
references/example-medium-skill.md	120-line skill with lookup tables and references
references/cross-platform.md	Install on 10+ tools, convert to .mdc / AGENTS.md / .windsurfrules
scripts/install.sh	Auto-detect platforms and install
scripts/run_loop.py	Automated description optimization
scripts/package_skill.py	Validate and package as .skill
agents/grader.md	Grade skill output against assertions
agents/comparator.md	Blind A/B between skill versions
agents/analyzer.md	Root-cause why one version beat another

Platform Support

Works in claude.ai (conversational testing), Claude Code (subagents, eval viewer, automated optimization), and Cowork.

Produced skills install on: Claude Code, Cursor, Codex CLI, Gemini CLI, Windsurf, GitHub Copilot, Antigravity, and any tool supporting the SKILL.md spec.

License

Apache 2.0. Scripts derived from anthropics/skills (Apache 2.0).