Agent Skills
github.com/addyosmani/agent-skills- Category
- AI Agents
- Rank
- No. 252Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- addyosmani
- GitHub
- 94.7k stars
- Latest release
- 0.6.9
- Date
About
Agent Skills is a collection of 24 Markdown workflows that give AI coding agents a repeatable process for the software lifecycle — specifying, planning, building, verifying, reviewing, and shipping. Each skill defines concrete steps, exit criteria, and verification gates rather than open-ended advice, and several include tables of common excuses for skipping a step alongside rebuttals. Setup instructions cover more than ten different agent tools.
What it does
It is a set of Markdown playbooks that coding agents load when a task matches a trigger description. Each playbook supplies ordered steps, warnings against skipped work, and evidence-based completion checks. A routing skill selects workflows, while commands and specialist personas provide explicit entry points for planning, implementation, testing, review, performance work, and release decisions.
Why it's ranked here
The strongest case is disciplined coverage without forcing every instruction into every session. Skills load on demand, share checklists, and define verifiable exit criteria. The repository also documents structural, routing, and behavioral evaluation tiers, which address a common weakness of prompt collections: instructions may exist yet activate incorrectly or produce weak behavior. Platform fidelity varies, but the underlying Markdown remains broadly portable.
What's good
The skill format teaches useful prompt engineering. Descriptions separate capability from activation conditions, full instructions load only when relevant, and supporting material stays outside the main context until needed. Rationalization tables directly counter predictable shortcuts such as skipping tests or specifications. Specialist personas keep review perspectives separate, while the shipping workflow permits parallel independent audits followed by one combined decision.
Tradeoffs
Portability is uneven despite the plain Markdown core. Installing one skill can omit shared checklists stored at repository level. Codex consumes the skills but not the Claude-specific commands, personas, or lifecycle hook. OpenCode relies on model compliance instead of native routing, while Windsurf guidance recommends copying only a few skills because of context limits. The collection also offers less of one fixed autonomous execution pipeline than some alternatives described by the project.
How to use it well
Use it for teams that want repeatable agent behavior and visible checkpoints around software changes. New projects can adopt the specification, planning, incremental build, review, and shipping sequence immediately. Established codebases should start with context gathering, read-only review, debugging, and characterization tests before broader changes. Load workflows by phase rather than all at once. It does not provide durable cross-session memory, and platform-specific setup may require separate rules or manual copying.
Technical notes+
README.md defines 24 skills, eight lifecycle-facing commands, selective installation through the skills CLI, and native integrations for several agents. docs/skill-anatomy.md specifies YAML frontmatter, on-demand loading, recommended verification and anti-rationalization sections, supporting-file conventions, and the portability cost of repository-level shared references. docs/agents.md separates skills, personas, and commands, endorsing parallel persona fan-out only for independent investigations. docs/developer-onboarding.md states there is no build step or package manifest and describes three evaluation tiers plus command-parity checks. docs/codex-setup.md confirms that Codex reads the shared root skill collection while Claude-specific commands, personas, and hooks remain unavailable there.
Observed
- Primary format
- Markdown skill definitions with YAML frontmatter
- Catalog structure
- 24 skills: 23 lifecycle skills and one routing meta-skill
- Install surface
- Whole collection or individual skills through the skills CLI, plus several native plugin and copy-based integrations
- User interfaces
- On-demand skill activation, chat invocation, slash commands on supported hosts, and specialist personas
- Platform support
- Documented setup for Claude Code, Cursor, Antigravity CLI, Gemini CLI, Windsurf, OpenCode, GitHub Copilot, Kiro, Codex, and Command Code
- Repository packaging
- No build step and no package manifest; repository validators use plain Node scripts
- Evaluation structure
- Documented structural, trigger-routing, and behavioral evaluation tiers
Read from README.md, docs/agents.md, docs/comparison.md, docs/codex-setup.md, docs/cursor-setup.md, docs/copilot-setup.md, docs/skill-anatomy.md, docs/adoption-guide.md, docs/opencode-setup.md, docs/windsurf-setup.md, docs/getting-started.md, docs/gemini-cli-setup.md, docs/antigravity-setup.md, docs/commandcode-setup.md, docs/developer-onboarding.md.
Intel on Agent Skills
- Subagents vs Agent Skills: Executing Reusable Knowledge for Long-Horizon Agentic Tasks
- Who Maintains Agent Skills? A Longitudinal Study of Human-Governed, AI-Assisted Skill Maintenance
- Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents
- Rethinking Self-Evolving Agent Skills: Feedback Dynamics over Multiple Rounds
- Autoreview (Agent Skills)
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.