Addy Osmani introduces Agent Skills, a GitHub project (26K stars) that adds senior-engineer scaffolding to AI coding agents. The core problem: agents default to the shortest path to 'done,' skipping specs, tests, reviews, and scope discipline. The project encodes these practices as markdown workflow files (skills) that get progressively injected into agent context. Key design principles include process over prose, anti-rationalization tables (pre-written rebuttals to excuses agents use to skip steps), mandatory verification exit criteria, progressive disclosure of skills, and strict scope discipline. The skills map to a full SDLC with slash commands (/spec, /plan, /build, /test, /review, /ship) and draw heavily from Google engineering practices like Hyrum's Law, the test pyramid, DAMP over DRY, and ~100-line PR sizing. Works with Claude Code, Cursor, Gemini CLI, Codex, Aider, and any tool accepting system prompts.
Table of contents
What a “skill” actually isThe SDLC the skills encodeFive principles that are doing the workThe Google DNAHow to actually use itWhat to steal even if you never installWhere this fits in the harnessClosing6 Comments
Sort: