agent-skills is a production-grade engineering skills package for AI coding agents. Its core idea is to encode the workflows, quality gates, and best practices that senior engineers use when building software into structured skills, and package them so that AI agents can follow them consistently at every stage of development.
Purpose AI coding agents by default tend to take the shortest path, often skipping specifications, tests, security reviews, and other practices that make software reliable. agent-skills provides agents with structured workflows that force them to follow the same strict discipline as senior engineers when writing production code. Each skill encodes engineering judgment: when to write a spec, what to test, how to review, and when to ship. It is not a generic prompt, but an opinionated, process-driven workflow that helps produce production-grade quality rather than prototype-grade quality.
Main capabilities It includes 9 slash commands mapped to the development lifecycle, and each command automatically activates the correct skill. For example, /build automatically generates a plan after the spec is finalized and implements all tasks after a single approval: you only need to approve the plan once, and then it runs autonomously; what it removes is the manual step between tasks, not verification - each task is still test-driven and committed separately, and it pauses when encountering failures or risky steps. In Claude Code, these are installed as plugin commands; if a built-in command has the same name (such as /review), use the namespaced form /agent-skills:review. Skills also activate automatically based on what you are doing, for example designing an API triggers api-and-interface-design, building UI triggers frontend-ui-engineering, and so on.
Skill composition There are 25 skills in total: 24 lifecycle skills plus 1 using-agent-skills meta-skill. Each skill is a structured workflow containing steps, verification gates, and an anti-rationalization table. Skills are grouped by purpose: Meta (discover applicable skills), Define (clarify what to build), Plan (break down), Build (write code), Verify (prove it works), Review (pre-merge quality gate), Ship (deploy with confidence). There are also preconfigured expert roles for targeted review, as well as reference checklists that skills bring in when needed.
How to use The fastest path is to use the open skills CLI, which installs to 70+ agents with a single command (Claude Code, Cursor, Codex, Copilot, Cline etc.). You can also obtain individual skills; however, note that when installing by skill, npx only copies skills/<name>/ and does not include the repository-level references/ directory. The skill will still work, but the shared manifest paths will be unavailable. It is recommended to integrate the whole repository, clone the repository, or copy the required manifests into the references/ directory inside the installed skill. Each host tool also has native integration methods: Claude Code can be installed via a marketplace or through local development installation; Cursor places workflow skills in .cursor/skills/ and brief policies in .cursor/rules/*.mdc; Antigravity can install skills and subagents as native plugins; Gemini CLI can be installed as native skills with automatic discovery, or added to GEMINI.md; Windsurf adds skill content to its rules configuration; OpenCode copies skills to .opencode/skills/ and adds a project-level AGENTS.md; Copilot uses definitions in agents/ as roles, with skill content placed in .github/copilot-instructions.md; its lifecycle slash commands are limited to Claude Code, and skills must be invoked by name; Codex can be installed as a native plugin (requires Codex CLI v0.122+), and after installation, @ is used to invoke it in conversation; Command Code uses the built-in cmd skills command. Skills are pure Markdown, and any agent that accepts a system prompt or instruction file can use them.
Design Features Each skill follows a consistent structure, and key design choices include: process rather than prose, where a skill is a workflow that the agent follows, with steps, checkpoints, and exit criteria; anti-rationalization, where each skill includes common excuses (such as "I'll add tests later") and rebuttals; verification is non-negotiable, where each skill ends with evidence requirements (passing tests, build output, runtime data), and "looks correct" is never enough; progressive disclosure, where SKILL.md is the entry point, and supporting references are loaded only when needed to save tokens. Skills also incorporate best practices from Google engineering culture, such as Hyrum's Law in API design, the Beyonce Rule and the test pyramid in testing, change size and review speed norms in code review, Chesterton's Fence in simplification, trunk-based development in Git workflows, shift-left and feature flags in CI/CD,and a dedicated deprecation skill.
Use Cases and Limitations Suitable for developers, teams, and tools that want AI coding agents to follow production-grade engineering processes. The project is licensed under MIT. For existing codebases, adoption guidance is provided: new projects can follow the full lifecycle from day one, while existing codebases can adopt an incremental, verification-first rollout approach. The repository also includes comparison documents with Superpowers and Matt Pocock's skills. Note: for specific installation commands, detailed configuration for each host, and version requirements, please refer to the official repository documentation; when installing skills individually, there is a known limitation regarding the availability of shared manifest paths (see #361).
