Same brief, same agent, three people. Did it build the same screen?

One agent in Claude, one brief: a workspace access screen where an admin approves or rejects member requests, with a reason when rejecting. Same model, same brief, three prompts written from three roles.

I ran it in rounds. Round A had only the brief. Round B added the TAP design system (tokens and twenty components) and a short rules file. Round C added one pattern file. Then I ran the designer prompt a second time to see how much changes by chance.

Without a system, I got three different products: three fonts, three palettes, no shared components, and each screen covered different states.

With the design system, the visual drift went to zero: no hard-coded colours or sizes, the same type, 11 to 17 real components per screen, and empty, loading and error states in all three.

But the layouts still split. The PM version added a member view nobody asked for, in both rounds. And the button said “Reject” in one screen and “Decline” in another.

Round C added a single “Review queue” pattern: layout, list, detail, the five states with their exact copy, what is out of scope, and a short copy glossary.

All three screens converged: two panes from 1024px, no extra views, Approve and Reject, the same five states. The error message reads “Couldn’t load requests.” word for word in every screen.

1. Tokens and components: how it looks.
2. Patterns: how it is structured and what it says.
3. The brief: only what is new in this feature.

Running the designer prompt twice kept everything the pattern fixed. What changed were the details it didn’t: the selected-row style, whether the request sat in a card, the avatar colour.

Those were the same kind of differences I had seen between roles. So they came from chance, not from who was asking. I moved them into the pattern. Running a prompt more than once turned out to be a cheap way to find the decisions a system still leaves open.

Along the way, the agents also caught real bugs in the tokens: values with no units, a tablet body size of 64 and a blue foreground identical to its background.

One model, one screen, three prompts I wrote myself, one run each plus one repeat. It is a signal, not a study. A real team would write prompts I wouldn’t, and a harder screen would leave more decisions open.