skill-creator

Author and audit Agent Skills (SKILL.md files) for OpenCode and any agent that supports the agentskills.io v1 spec — Claude Code, Cursor, Codex, Copilot, Gemini CLI, Goose, and 30+ others.

This is the source repo. The skill itself lives in SKILL.md at the root.

What's in the skill

• SKILL.md — workflow for creating and auditing skills (lean, dense, ~200 lines / ~2.4k tokens)
• references/frontmatter.md — full field reference + platform-specific extensions
• references/design-patterns.md — problem-first vs tool-first, 5 common patterns, foundation skills
• references/pitfalls.md — top 13 community pitfalls with fixes
• references/validation.md — 4 validators, CI integration, pre-commit hooks
• references/checklist.md — A/B/C tiered pre-publish checklist

Install

With skillshare

skillshare install dewdad/skill-creator

Tracked install (preserves .git for updates):

skillshare install dewdad/skill-creator --track

Project-level install (into the current repo's skills directory):

skillshare install dewdad/skill-creator -p

Update later:

skillshare update skill-creator

Manual

Clone into any of the locations your agent discovers:

# OpenCode global
git clone https://github.com/dewdad/skill-creator ~/.config/opencode/skills/skill-creator

# Cross-platform (Codex, Gemini CLI, OpenCode, Antigravity)
git clone https://github.com/dewdad/skill-creator ~/.agents/skills/skill-creator

# Claude-compatible
git clone https://github.com/dewdad/skill-creator ~/.claude/skills/skill-creator

Or install into a project's local skills directory:

git clone https://github.com/dewdad/skill-creator .opencode/skills/skill-creator

When this skill triggers

Phrases like:

• "create a skill for this"
• "new SKILL.md"
• "audit this skill"
• "review my SKILL.md"
• "skill not triggering"
• "make this reusable across sessions"
• "improve skill description"

The skill teaches the agent to:

1. Wait for a real workflow before generalizing it into a skill (eval-driven dev).
2. Pick the right discovery path (.opencode/, .agents/, .claude/, etc.).
3. Write spec-compliant frontmatter with trigger keywords.
4. Use dense imperative voice, not polite prose.
5. Split into references/ via progressive disclosure.
6. Validate with skill-validator, skills-ref, or skillcheck before publishing.
7. Test via the Claude-A-builds / Claude-B-tests loop.

Validation

This skill passes:

• skills-ref validate ./SKILL.md → Valid skill
• skillshare audit skill-creator → LOW risk (1/100), 100% auditable, 0 critical/high/medium/low findings

Compatibility

Designed for OpenCode and any agent compatible with the agentskills.io v1 spec.

Tested with:

• OpenCode (primary)
• Claude Code
• Cursor
• VS Code Copilot

Frontmatter uses only open-spec fields (name, description, license, compatibility, metadata) for maximum portability. See references/frontmatter.md for platform-specific extensions if you need them.

Provenance

Built from synthesized 2026 best practices across:

• agentskills.io specification
• Anthropic Skills authoring best practices
• Anthropic engineering blog: Equipping agents for the real world
• OpenCode Skills documentation
• agent-ecosystem/skill-validator
• Serghei Iakovlev — Agent Skills 101 (Feb 2026)
• Strapi — What Are Agent Skills and How To Use Them (Feb 2026)

License

Apache-2.0. See LICENSE.