Addy Osmani Open-Sources 19 Google Engineering Practices as AI Agent Slash Commands

Google's Addy Osmani dropped Agent Skills this week, and it hit 2,500+ likes and 340+ retweets before I finished reading the README. Seven slash commands: /spec, /plan, /build, /test, /review, /code-simplify, /ship. Each one encodes quality gates from Google's internal engineering practices, packaged as Markdown skills that work with Claude Code, Cursor, Antigravity, and any agent that accepts Markdown input.

This solves a problem I've been banging my head against. AI agents skip specs. They skip security reviews. They optimize for "done" over "correct." You can tell them not to in your system prompt, but they drift. Osmani's approach is different. Instead of telling the agent what not to do, you give it a structured workflow that forces the right sequence. /spec before /plan. /plan before /build. /test and /review before /ship.

The 19 practices span the full development lifecycle, and they're opinionated. This isn't a generic framework. It's one senior engineer's distilled experience from years at Google, turned into executable agent instructions. The /code-simplify command alone, which focuses on reducing complexity rather than adding features, addresses one of the biggest failure modes I see in AI-generated code: unnecessary abstraction.

What makes this land for me is that it connects directly to Karpathy's agentic engineering thesis. Vibe coding is accepting what the agent gives you. Agentic engineering is directing the agent through quality gates. Osmani's skills are the quality gates, made portable and open source.

I installed the /spec and /review commands into my Claude Code setup yesterday. First impression: the spec command forces a level of upfront thinking that I was skipping, and the output is better for it. The review command catches things I'd normally catch on my third read-through, which saves me the first two.

For builders: grab the repo, install the slash commands that match your workflow, and customize them. The real value isn't the specific commands. It's the pattern of encoding your team's engineering practices as executable agent skills instead of hoping the agent follows written instructions.