27,000-plus GitHub stars pushed Agent Skills into wider use as Addy Osmani framed it as scaffolding for AI coding agents that otherwise skip specs, tests, review and launch checks.
20 Markdown-based skills map to six SDLC phases—define, plan, build, verify, review and ship—with seven slash commands and a router that loads only the workflows relevant to a task.
Five design choices do most of the work: process over prose, anti-rationalization tables, mandatory verification evidence, progressive disclosure and strict scope discipline to keep agent-written PRs reviewable.
Osmani ties the project to published Google engineering practices, including the test pyramid, roughly 100-line PR sizing, trunk-based development and feature flags, arguing those habits are not applied by models by default.
The repo is MIT-licensed and portable across Claude Code, Cursor, Gemini CLI, Codex and other prompt-driven tools, positioning the skills as reusable harness components for longer-running AI coding sessions.
If AI needs a harness to code responsibly, what happens when these systems become fully autonomous?
Does enforcing senior discipline on AI agents stifle their creative problem-solving abilities?
Is managing an AI's 'comprehension debt' the new critical skill for human engineers?