Taste Lab

So what does it actually hand back? A complete design brief. The pipeline runs four agents in sequence, each reading the same page through a sharper lens until what's left is a file an AI tool can build from. You don't configure anything beyond the URL, which keeps the workflow short. That simplicity is the point.

The first pass pulls precise, objective values from the page across a broad set of categories. Colors come as hex codes, spacing and sizes come as pixel values, and type comes with weight and tracking numbers. Nothing is approximated. That matters. AI tools tend to drift toward round numbers when left to guess, and the drift adds up across a whole layout.

Instead of dumping raw numbers, Taste Lab detects the repeating rules behind them. A dozen similar spacing values collapse into one base unit with a scale. Each rule comes with an evidence line and a short design goal explaining why it exists. You see the logic, not just the data.

The most useful part is the reasoning layer. Taste Lab writes a handful of principles, each covering a design decision, what triggered it, why it was made, and what it gave up. Including at least one restraint principle is intentional. Good design is often about what a designer refused to add, and that gets captured here. It's the part most tools skip.

Every run writes two files to your machine: a Markdown brief for humans and a JSON token set for tools. The Markdown version reads like a design memo you can share or edit. The JSON version plugs straight into an AI coding workflow. Your agent gets machine-readable values to reference.

Before writing the output, Taste Lab runs an anti-slop pass that greps for filler language and validates the JSON structure. That's a small detail, but it matters. The brief you get is less likely to be padded with the vague, confident-sounding phrasing that AI tools produce by default.