Robby 5 v2

Robby 5 v1 could build prototypes fast, but nothing made it follow the design system. v2 is the rebuild. A designer describes a screen and gets back a working prototype made from the real components — and the tool won’t build anything until it has read how the product works.

AI tools invent components you don’t have, reach for colours that aren’t yours, and describe your product in generic terms. I wanted to measure how often, not guess.

I set up a test version of Zendesk and planted three made-up words in its product description: a record type called Playbooks, a status called Waiting on vendor, and a screen called Frontline. No AI could know them without reading the description. Then I asked for eleven screens, each in a fresh session.

One built a screen from a different product entirely, complete with ServiceNow’s ticket numbers. One copied deliberately wrong vocabulary straight from the prompt. One built an invoice-escalation flow wired to NetSuite — the wrong product altogether.

All eleven still compiled, passed every automated check, and followed every layout rule.

One checkpoint before the AI writes anything, one after. The generation in between is the least interesting part — the checkpoints are what make the output trustworthy.

Every prototype is checked against ten design rules — empty states, loading states, colour never being the only signal — and against the product’s own vocabulary.

But only the checks that are almost never wrong are allowed to block. Some rules I care about, like confirming before deleting, stay as warnings, because a scan of the code can’t always tell whether they’re met.

Each test round sets up a fresh fake product, plants words no AI could guess, runs sessions blind, and searches the results for those words. The most revealing test asked for a word the product description forbids: the right answer is to use the correct word and say why.