Addy Osmani agent-skills review: 24 production skills, from a web-performance engineer

This is the fourth and final repo in the run of trending Claude Code skill repos. The first three were superpowers, karpathy skills, and mattpocock skills. This one is addyosmani/agent-skills, 68k stars, by Addy Osmani, a well-known web-performance engineer and writer at Google. Its description is short: "Production-grade engineering skills for AI coding agents."

Opening it up, the difference from the first three is immediate. It is not small or minimal, it is the most comprehensive of the set, 24 skills covering from spec all the way to ship, and it has something none of the other three do: genuine web work, both browser-based testing and performance tuning, which happens to match exactly what its author is known for.

This post goes in order: what it is and why the stars, then the skills that genuinely stand out, then how the four repos I reviewed each sit at a different pole and which to reach for, and finally how to install it and how I use it myself.

Part 1What addyosmani agent-skills is, and why the stars

This is the most comprehensive set of the four. Inside are 24 skills, ordered along real work: spec, plan, build, verify, review, ship. There are specialized personas, slash commands, and reference checklists. It is packaged as a plugin that installs to many hosts, including Claude Code, Cursor, and Copilot.

What is interesting is that, reading inside, each skill carries an excuse-preemption table, the same pattern superpowers uses, naming the reasons an agent tends to give for skipping a step, with rebuttals. This is the third time we have seen this excuse-naming pattern across unrelated repos, which only confirms it is a principle the people serious about this all arrive at.

Why 68k stars? Part of it is the author's name, Addy Osmani is well known in the web world for performance. But more importantly, it is the one set that turns real web knowledge into actual skills, not just a generic process.

Part 2The skills that genuinely stand out

Four worth calling out, distinctive and rarely found elsewhere.

1. browser-testing-with-devtools gives the agent eyes into a real browser through Chrome DevTools instead of guessing from code, used to chase UI bugs, inspect the network, and measure speed, while treating browser content as untrusted data for safety.
2. performance-optimization follows measure before you optimize, using both synthetic tools like Lighthouse and real-user data, hunting common issues like duplicate fetches, unoptimized images, and oversized bundles.
3. source-driven-development the rule is that every framework-specific decision must be backed by official docs, not the model's memory, with an authority order of official docs, official blogs, web standards, not random forum answers.
4. doubt-driven-development after a long session where assumptions have set in, spawn a fresh reviewer that starts from zero and is biased to disprove, not to approve, capped at three cycles, then let the human decide if it was worth it.

What they share is an obsession with confirming things for real: measuring performance for real, seeing the page for real, citing real docs, actively looking for where you are wrong. source-driven and doubt-driven line up exactly with what this blog keeps saying: do not trust what the AI says until you have confirmed it against a real source.

Part 3Four poles: which to pick

Having reviewed all four, the picture is clear: each answers a different need.

• superpowers a skill set with enforcement gates, for letting an AI run long on its own.
• karpathy skills 4 abstract principles in one file, for the leanest set of principles.
• mattpocock skills a practitioner's directory, for tools in pieces to remix.
• addyosmani agent-skills the most comprehensive set, for web people who want both the process and performance knowledge.

If you work in frontend or care about page speed, Addy's is the only one of the four that speaks to it directly. All four do not clash. You could take karpathy's 4 principles as a base, add Addy's web skills, borrow superpowers' excuse-naming to make the rules hold, and pick skills piece by piece the way mattpocock encourages. They are all examples for assembling your own set.

Part 4Should you install it, and how to use it

Should you try addyosmani agent-skills?

Very much so if you do web work or want a process set that runs end to end. It installs as a plugin and supports many hosts. The upside is completeness; the caution is that it is large, so do not enable every skill at once. Pick the phases you do most, especially the browser and performance skills that are hard to find elsewhere.

The idea you can use right now

Even without installing the repo, two principles transfer immediately: source-driven and doubt-driven. Do not let the AI write framework-specific code from memory, have it pull the official docs first and cite them, and after a long piece of work, set up a review that deliberately looks for what is wrong rather than confirming it is right. Those two prevent a lot of mistakes with nothing to install.

For my part, I compared the source-driven approach with the rules I already use, since it lines up with the verify-before-you-trust principle I have held all along. Many of the frontend skills do not match my daily work, so I took only some. How I blend them into my own stack is a detail I am leaving out of this post.

The one rule to remember

If you remember one thing from this four-repo series, let it be this: no repo is meant to be swallowed whole. superpowers, karpathy, mattpocock, and addyosmani are four corners of the same thing. Take the part that fits your work from each, and assemble your own set. That is how you get the most out of them.

This post is one layer in the 7-layer architecture of a production AI agent.