How to Test a Claude Skill Before You Install It

More skills don't make your agent better. Every skill you install rides on every brief you send from then on, and it changes what comes back. So before this one went on my machine, I tested it: one real job on a copy of my website, with it installed and without it. The skill is UI UX Pro Max, the job was a full design pass on my landing page, and the verdict is on camera. Spoiler: it stays. It earned the folder. The test is the point, and it's easy, because your agent does the heavy lifting. Two tests: 1. The Mirror Run. The same job twice, with the skill and without it, on a git-backed copy of your files, in a fresh session each time, so the second run can't see the first. 2. The Empirical Review. Did it change anything: the page, the diff (every file, including the ones the job never named), and what it said it changed next to what git says. Do this on the next skill you're about to install. Not on everything you already have. Just the next one. THE PROMPT (paste it into Claude Code, Codex, whatever you run: it copies your project and leaves your secret files behind, then walks your agent through both tests on one real job you name) https://github.com/marctalkstomachine... Every tool from my videos, free (leave your email and the workbench comes to you): https://marctalkstomachines.com Chapters 0:00 A hyped Claude skill on a copy of my site 0:43 What the skill changed 1:14 Why every skill you install costs you 1:41 Test 1: The Mirror Run 3:14 Installing the skill 4:31 Test 2: The Empirical Review 5:07 The diff: four files, two asked for 5:45 The wide-open run 6:27 The receipt 7:01 The verdict 7:35 The first job: the cube 8:39 Test the next one #Claude #ClaudeCode #AI #Codex #ContextEngineering