Agent Skills Framework Released

Most artificial intelligence coding agents can write a feature in one pass, but they often skip the habits that keep real software from breaking in production, like writing a spec first, proving behavior with tests, and checking for security holes before merge. Addy Osmani’s new open-source project tries to turn those habits into reusable instructions an agent can follow every time. (github.com) (addyosmani.com) An artificial intelligence agent is not just autocomplete; it is a software system that can plan, call tools, run code, and revise its own work across multiple steps. The problem is that giving an agent a terminal and a codebase does not automatically give it the judgment of a senior engineer. (addyosmani.com) (github.com) That missing judgment is what “skills” are meant to supply. Osmani describes a skill as a modular package of instructions, decision rules, and reference material, like handing a mechanic a repair manual instead of just a wrench. (addyosmani.com) The framework is called Agent Skills, and the GitHub repository says it packages production-grade engineering workflows for artificial intelligence coding agents. On April 9, 2026, the repository showed about 9,600 stars and roughly 1,000 forks, which is unusually fast pickup for a developer workflow project. (github.com) The core idea is simple: move the agent through the same stages a disciplined team uses, from defining the idea to planning, building, testing, reviewing, and shipping. The repository maps that flow to seven slash commands, including `/spec`, `/plan`, `/build`, `/test`, `/review`, `/code-simplify`, and `/ship`. (github.com) The first guardrail is “spec before code.” Osmani’s own writing on agent workflows argues that large, vague prompts fail because the model’s attention gets spread too thin, so the better pattern is to start with a concise brief, expand it into a detailed specification, and treat that file as the shared source of truth. (addyosmani.com) (github.com) The second guardrail is test-first proof instead of “looks good to me.” Agent Skills puts testing in its own phase, and the repository describes each skill as a structured workflow with verification gates and anti-rationalization checks, which is a way of stopping the model from talking itself into shipping unproven code. (github.com) The third guardrail is loading the right instructions only when they are needed. Osmani’s lesson on skills says this on-demand approach can shrink startup instructions from around 150,000 tokens to about 2,000 tokens, which matters because every extra token costs money and competes for the model’s attention. (addyosmani.com) The framework is also meant to plug into the tools developers already use. The repository includes setup paths for Claude Code, Cursor, Windsurf, GitHub Copilot, Codex, and other agents that can read Markdown instruction files. (github.com) You can see where this goes in Osmani’s separate Web Quality Skills project, which applies the same skill format to performance, accessibility, search engine optimization, and Core Web Vitals checks based on Google Lighthouse guidance. That turns “make this page faster” from a fuzzy request into a repeatable checklist with thresholds, scripts, and references. (github.com) What Agent Skills really changes is the unit of reuse. Instead of copying one giant prompt from chat to chat, teams can reuse named workflows, update them independently, and share them across projects the way they already share code libraries and style guides. (addyosmani.com) (github.com)