When I started redesigning this site, I didn’t want it to look like every other site an AI builds: a cream background, a big serif font and one coloured word in the headline. A coding agent writes code that works, but the look keeps drifting back to the same habits. When I saw the first design, I could tell some things looked wrong, but I couldn’t say what.
Impeccable was one of the Skills I installed before we started, and it named those things one by one.
What Impeccable is
A free, open-source Skill by Paul Bakaus. You install it into the coding agent you already use, and it works with Claude Code, Codex, Cursor, GitHub Copilot, Antigravity and others.
It gives the agent specific design words instead of vague descriptions. Instead of telling it “make this stand out”, you run bolder on one part, quieter on a busy page, or polish before you ship. There are 24 commands, including:
critique: a review that scores the page out of 40 against ten well-known usability rules, with a ranked list of fixes from P0 (blocks the work) to P3 (final polish).audit: a technical check of accessibility, page speed, light and dark themes, and how the page looks on a phone.
It also comes with a detector of 62 rules that looks for the habits that give away an AI-made design, such as purple gradients, Inter everywhere and cards inside cards. The detector uses no AI model and needs no API key, so the same page gets the same result every time.
How I used it on this site
After each design round, I ran critique on the tools page, the home page and the Pick page. I split each review in two, each part in its own agent: one reviewed the design, the other ran the detector and measured in the browser, so the agent that made the design never graded its own work.
The tools page’s score rose over five rounds: 22, then 25, 28, 30 and finally 35 out of 40.
The score comes from Impeccable itself, so it isn’t a final verdict on the design. What helped was the list that comes with it, because it pins down exactly what’s wrong. In the second round, for example, it said the top of the home page was too close to another site I’d been inspired by. In the fourth round it found that clicking a tool’s name sometimes opened the kind explanation instead of the tool.
What I liked
It saves a lot of time when designing sites and apps. The things that feel wrong but that you can’t name, it names for you and ranks by importance, and you can hand that list to any agent to fix.
What I didn’t like
It goes back and forth a lot. Every review reads files, opens the browser, takes screenshots and measures, then comes a round of fixes, then another review. All of that burns tokens, which is why I don’t recommend it on a $20-a-month plan.
Its docs also say you get better results by focusing on one part at a time instead of the whole page.
How to start
From your project folder:
npx impeccable installThe installer looks for the coding tools you have and asks whether you want it for this project only or for all projects. Then close the tool, open it again and tell the agent (in Codex, type $impeccable instead of /impeccable):
/impeccable init
It reads your code and asks you about the product: who it’s for, what it has to do and what its limits are. Then it writes a PRODUCT.md file that every command reads from then on. After that, pick one part and start:
/impeccable critique the pricing cards
You can run the detector without any agent at all:
npx impeccable detect src/If you’re on a $20-a-month plan, start with the detector on its own, since it uses no model, and save critique and the long rounds for when your plan is bigger.
Versions at writing
- Impeccable · 4.8.2