Agent Skills, Safety-Verified

An agent skill is a reusable instruction file — typically a Markdown or YAML file — that gives an AI agent a specific, repeatable capability. Skills define how the agent should behave for a particular task, such as reviewing pull requests, writing commit messages, or debugging test failures. Unlike one-off prompts, skills are installed once and applied automatically.

Each skill in this directory is scanned for prompt injection attempts, dangerous shell commands, data exfiltration patterns, and instructions that could override the agent's safety guardrails. Skills receive a "Verified", "Caution", or "Untested" rating. The full criteria are explained at codaone.ai/skills/security.

Skills here are tested across Claude Code, OpenClaw, Cursor, Codex, Windsurf, and Gemini. Each skill's detail page lists exactly which platforms it works on. Some skills are platform-specific; others use a generic format that works anywhere CLAUDE.md-style instruction files are supported.

Open the skill's detail page and copy the install command or file content. For Claude Code, you typically drop the skill file into your project's .claude/ directory or reference it in your CLAUDE.md. Platform-specific instructions are shown on each skill page.

Yes. Submit your skill via the GitHub link in the footer. Include the skill file, a plain-language description of what it does, which platforms it targets, and any known limitations. It will go through the same safety review before appearing in the directory.

The directory is free and the skills themselves are open-source. You'll need the relevant AI coding tool (Claude Code, Cursor, etc.) to run them — most have free tiers. A small number of skills invoke paid APIs, which is noted on the skill's page.