[{"data":1,"prerenderedAt":294},["ShallowReactive",2],{"en-post-165":3},{"id":4,"title":5,"body":6,"description":12,"extension":287,"meta":288,"navigation":289,"path":290,"seo":291,"stem":292,"__hash__":293},"content/post/165.md","165",{"type":7,"value":8,"toc":275},"minimark",[9,13,18,21,25,80,89,93,142,146,211,214,218,221,225,245,248,252,255,259,262],[10,11,12],"p",{},"Getting an AI coding agent to write software is rarely the hard part. Getting it to follow the same discipline a senior engineer would is. Agents love the shortest path, so specs get skipped, tests get promised for later, and security review quietly disappears. addyosmani/agent-skills was built to close that gap. It packages the workflows and quality gates experienced engineers use into reusable skills, so tools like Claude Code, Cursor, Codex, and Copilot follow one consistent standard across the whole development cycle.",[14,15,17],"h2",{"id":16},"what-is-agent-skills","What Is agent-skills",[10,19,20],{},"agent-skills is a production-grade engineering skills library for AI coding agents, led by Addy Osmani, an engineering director at Google. It breaks the full lifecycle, from defining a requirement through planning, building, verification, review, and release, into skills and commands an agent can call directly, so the model stops improvising. The repository was created in February 2026 and is still updated. It has about 115,271 stars and 12,104 forks, with a homepage at skills.addy.ie and source under the MIT license. A portable core works across more than seventy hosts through adapters, including Claude Code, Cursor, Codex, Copilot, and Cline. What you install is a set of engineering habits, not another interface. The whole set drops into the editor you already use with a single command.",[14,22,24],{"id":23},"features","Features",[26,27,28,32,35,38,63,66,77],"ul",{},[29,30,31],"li",{},"Anti-rationalization tables. Each skill lists the excuses agents use to skip a step, like \"I will add tests later,\" and answers each to close the easy way out.",[29,33,34],{},"Verification is non-negotiable. Every skill ends with evidence, passing tests, a build log, or runtime data. Looking right does not count as done.",[29,36,37],{},"Lifecycle grouping. The repository holds 25 skills sorted into define, plan, build, verify, review, and ship, plus one meta skill that coordinates them.",[29,39,40,41,45,46,45,49,45,52,45,55,58,59,62],{},"Slash commands. Nine commands, including ",[42,43,44],"code",{},"/spec",", ",[42,47,48],{},"/plan",[42,50,51],{},"/build",[42,53,54],{},"/test",[42,56,57],{},"/review",", and ",[42,60,61],{},"/ship",", turn workflows into one instruction.",[29,64,65],{},"Expert roles. Four perspectives, code reviewer, test engineer, security auditor, and web performance auditor, switch on when needed.",[29,67,68,69,72,73,76],{},"Progressive disclosure. ",[42,70,71],{},"SKILL.md"," is only the entry point, and supporting ",[42,74,75],{},"references/"," load on demand to keep tokens low.",[29,78,79],{},"Host adapters. One portable core maps to many agent platforms, so changing tools does not mean rewriting the process.",[10,81,82,83,88],{},"It also encodes ideas from Software Engineering at Google, from the testing pyramid to Hyrum's Law and trunk-based development, and pairs with the spec-driven approach in ",[84,85,87],"a",{"href":86},"/post/76/","What Is github/spec-kit and How to Use It",". Every skill is plain Markdown, so a team can read, fork, and extend it like any source.",[14,90,92],{"id":91},"how-to-use","How to Use",[94,95,96,103,112,118,125,134],"ol",{},[29,97,98,99,102],{},"Run ",[42,100,101],{},"npx skills add addyosmani/agent-skills"," to install the skills into a supported agent. The repository docs list each host.",[29,104,105,106,108,109,111],{},"Trigger the workflow with slash commands, starting with ",[42,107,44],{}," to pin down the requirement, then ",[42,110,48],{}," to break it into tasks.",[29,113,114,115,117],{},"Move into ",[42,116,51],{}," to implement, keeping test-driven work and one commit per task.",[29,119,120,121,124],{},"For an approved plan you want carried through in one go, use ",[42,122,123],{},"/build auto"," autonomous mode. It still keeps tests and commits.",[29,126,127,128,130,131,133],{},"Before wrapping up, run ",[42,129,54],{}," and ",[42,132,57],{}," so verification and review leave real evidence.",[29,135,136,137,141],{},"To make security a dedicated pass, add the audit view from ",[84,138,140],{"href":139},"/post/104/","What Is cloudflare/security-audit-skill and How to Use It",".",[14,143,145],{"id":144},"supported-platforms","Supported Platforms",[147,148,149,162],"table",{},[150,151,152],"thead",{},[153,154,155,159],"tr",{},[156,157,158],"th",{},"Item",[156,160,161],{},"Detail",[163,164,165,176,184,195,203],"tbody",{},[153,166,167,171],{},[168,169,170],"td",{},"Skill format",[168,172,173,174],{},"Plain Markdown, entry point ",[42,175,71],{},[153,177,178,181],{},[168,179,180],{},"Tooling",[168,182,183],{},"JavaScript/TypeScript",[153,185,186,189],{},[168,187,188],{},"Config",[168,190,191,194],{},[42,192,193],{},"plugin.json",", TOML, Shell scripts",[153,196,197,200],{},[168,198,199],{},"Agents",[168,201,202],{},"Claude Code, Cursor, Codex, Copilot and 70+ more",[153,204,205,208],{},[168,206,207],{},"License",[168,209,210],{},"MIT",[10,212,213],{},"Code review and security audit roles work out of the box, with no prompts to write from scratch. Skills are plain Markdown, so reading or editing one means opening a file. The layered config keeps the portable core separate from host-specific adapters, and the whole set updates together from the repository. No account, subscription, or extra runtime is required.",[14,215,217],{"id":216},"pricing","Pricing",[10,219,220],{},"agent-skills is open source under the MIT license. Personal and commercial use are both free, and you may use, modify, and redistribute it as long as you keep the original license notice. There is no subscription, no paid tier, and no commercial support contract. Questions go to GitHub Issues and community discussion. To turn it into an in-house standard, fork the repository and maintain your own branch.",[14,222,224],{"id":223},"use-cases","Use Cases",[26,226,227,230,233,236,239,242],{},[29,228,229],{},"Team standards. Different editors and agents, but code that follows one structure, so review stays on substance, not style.",[29,231,232],{},"Onboarding. Put a senior engineer's checklist into skills instead of repeating reminders in every review, which shortens the ramp for new hires.",[29,234,235],{},"Test gaps. Give a test-skipping agent a hard constraint it cannot talk its way around.",[29,237,238],{},"Long tasks. Split a large request into verifiable steps with plan and build commands, recording evidence at each step.",[29,240,241],{},"Security passes. Have the agent review its own work through an auditor's eyes before anything is marked done.",[29,243,244],{},"Process encoding. Turn a company workflow into team-owned skills that every new agent inherits.",[10,246,247],{},"It suits teams that care about engineering quality, while smaller groups doing throwaway work can leave the ceremony off until a project grows. Skills can be adopted one at a time, so nothing changes all at once.",[14,249,251],{"id":250},"how-it-works","How It Works",[10,253,254],{},"agent-skills is led by Addy Osmani and backed by Google engineering practice, with the community maintaining the skills and checklists. The repository separates the skill core from host adapters, so the center stays stable while adapters follow each agent's changing interface. That layering is how one set serves seventy plus tools without bloating. Contributions follow the usual open source route, with improvements gathered through issues and pull requests. It sits in an ecosystem alongside Superpowers and Matt Pocock's skills, with comparison docs still growing. For users, the result is a portable, forkable engineering base that can evolve with a team. Adoption is gradual by design, since a single skill can be added without replacing the whole workflow, and the anti-rationalization tables keep old habits from quietly returning.",[14,256,258],{"id":257},"conclusion","Conclusion",[10,260,261],{},"The ability of AI coding agents is rarely the bottleneck. Discipline is. agent-skills turns the old but effective habits, write a spec, add tests, run review, show evidence, into skills a model can actually follow. Install it once and let the process carry the work. If you are tired of half-finished deliveries from your agent, this library is worth a try. It asks the model to slow down just enough to finish what it started and to show the work behind each claim of done.",[10,263,264,265,270,271],{},"Project ",[84,266,267],{"href":267,"rel":268},"https://github.com/addyosmani/agent-skills",[269],"nofollow","\nHomepage ",[84,272,273],{"href":273,"rel":274},"https://skills.addy.ie",[269],{"title":276,"searchDepth":277,"depth":277,"links":278},"",2,[279,280,281,282,283,284,285,286],{"id":16,"depth":277,"text":17},{"id":23,"depth":277,"text":24},{"id":91,"depth":277,"text":92},{"id":144,"depth":277,"text":145},{"id":216,"depth":277,"text":217},{"id":223,"depth":277,"text":224},{"id":250,"depth":277,"text":251},{"id":257,"depth":277,"text":258},"md",{},true,"/post/165",{"description":12},"post/165","tGx5fNL10IUgIPbDNZQLEKd6DbIkVZRNDgbXTA4a_Xs",1794378305115]