I wrote a 900-line brief.
An AI built the product.
I directed everything in between.
The Idea
A deliberately silly project for a serious test.
I wanted to test something specific: can a designer write a complete enough product brief that an AI coding agent builds the whole thing without being micromanaged? The zombie subject matter was intentional — real scope, zero professional stakes. A proper multi-feature web application: database, detail pages, interactive tools, data charts, a design system.
If the methodology works here, it works anywhere.
The Spec
The product existed on paper before a single line of code was written.
Before anything was built, I designed the full specification — every feature, every data structure, every page, every design decision in one document. Not a mood board. Not a list of screens. A complete product brief precise enough that nothing needed to be invented during the build.
The hypothesis: the quality of what gets built is determined entirely by the quality of what gets specified. Vague brief → vague product. Complete brief → working software.
Directing the Build
One message to start. Then: use it, find what's wrong, describe it clearly.
I handed the spec to Claude Code with a single instruction: “start building.” The spec did the rest. No back-and-forth about what features to include or what pages should show — those decisions had already been made.
My role from that point was to be the user. Open every page. Try every interaction. Notice what looked wrong or felt off. Describe it specifically enough that it could be fixed.
What I Caught in Testing
Five things were wrong. I found all five by using the product.
None of these required technical knowledge to spot. They required paying attention.
Navigating to the Survival Guide showed a completely blank page. Something had broken silently — the page wasn't loading at all.
The Compare page showed a nonsense value instead of Universe B's score. The number was clearly wrong — not a design choice, a mistake.
The home page hero was a massive wall of pixel art. The title was pushed below the fold. The cityscape had taken over the entire viewport.
On desktop, the mobile hamburger menu was visible alongside the full navigation. Both were showing simultaneously.
The site deployed successfully locally but failed when pushed live. Something that worked in development didn't pass the stricter production check.
What It Proved
Writing the spec is the design work. Everything else is execution.
The part of this project that required skill wasn't the build — it was the brief. Deciding what data a zombie universe needs for survival analysis, what features make it a tool instead of a database, how 12 pages should relate to each other. That's product thinking. The spec is where it lived.
- Front-load the thinking. An AI coding agent doesn't fill gaps in requirements — it makes them up. The only way to control the output is to eliminate the gaps before the build starts.
- The designer's job doesn't end at handoff. Using the product seriously, with fresh eyes, is where real quality control happens. Five things were wrong that wouldn't have been caught any other way.
- Directing is a skill. “The hero is too tall and the title isn't visible” is more useful than “the layout doesn't look right.” Specific descriptions produce specific fixes.
