October 11, 2026
agent-skills: 25 skills make the agent prove its code works before it says 'done'
The agent-skills repo has 25 skills, and each one ends with a requirement for evidence: passing tests, build output, runtime data. Installing into Claude Code takes two commands.
Addy Osmani
@addyosmani
Verification is one of the most important things for getting excellent results from an agent. Define clearly what “done” means before code review starts. A reviewer, human or agent, asked to find bugs without that will still find something to improve. Give Claude checks it can run (bash commands, tests, clicking through the app in a browser), and have it run them too before it says it's finished. When nothing blocks merging and shipping, then it's done.
Asking Claude to review your code is a slot machine. Every review finds brand new bugs. Review again: more. Review a tenth time: it still finds something. It never runs out of bugs to find... so how do you decide your code is actually done?
· 28.1K views
Passing the compiler and type checks is not enough: under the repo's Definition of Done, code is checked at runtime, and new behavior is covered by tests that fail without the change and pass with it.
Order. Addy Osmani advises writing down what “done” means first, and only then calling a reviewer, human or agent. Without that bar, a reviewer will always find something to improve.
Human at the end. Under the Definition of Done, a change must be approved by a human before merge or deploy, and dead code, debug output and commented-out blocks are removed from the code.
The sources contain no measurements of the effect: only rules and instructions, with no data on the result.