Design guidance for AI coding agents
Every model trained
on the same
SaaS templates.
Skip the guidance and you get the same handful of tells on every project. Impeccable answers with one skill, 24 commands, live browser iteration, and 61 deterministic detector rules — and the deterministic half runs with no LLM and no API key.
From your project root
$ npx impeccable install
> /impeccable init # writes PRODUCT.md
The tells it names out loud
project’s own anti-patterns- ×Overused fonts — Arial, Inter, system defaults
- ×Purple-to-blue gradients
- ×Cards nested inside cards
- ×Gray text on colored backgrounds
- ×Pure black or gray — always tint
- ×The rounded-square icon tile above every heading
- ×Bounce and elastic easing — feels dated
A named tell is a checkable tell. This list is why the detector rules can be deterministic at all — you cannot grep for “bad taste”, but you can grep for a nested card.
Two kinds of check
only one of them needs a modelDeterministic
Run by the CLI and the browser extension. The project states they execute with no LLM and no API key — same input, same verdict, every time.
LLM-only critique
Hierarchy, clarity, emotional resonance — the judgments a rule cannot encode, kept in a separate tier rather than mixed into the rule count.
Keeping these apart is the honest move: it means the number on the left is a real number.
A shared vocabulary, in opposing pairs
Shipping bolder and quieter as two commands is a quiet admission that there is no single correct amount of design — which is a more useful thing for a design tool to believe than a house style. 24 commands in total; the two highlighted are the pair that makes the point.