Why a workflow
Asking an agent for "a few options" produces three tints of the same idea. Asking it to build a throwaway picker with genuinely different directions produces a decision.
Steps
- Scope one thing. A toast, a page transition, the pricing card. Not "the dashboard".
- Recon. Stack, tokens, personality, where the piece renders. If it already exists, the current version is variant 1.
- Name the axis. Each variant gets a name and the dimension it diverges on: layout, density, motion, interaction model. Two variants on the same axis is one variant.
- Build the harness. Isolated route, one variant at a time at full size, a fixed picker pinned bottom-center. Any number you would hardcode goes on a slider.
- Verify before showing. Every variant renders, every interaction works, console clean, reduced-motion path exists.
- Pick. The user chooses. Present the set as a table: when each wins, what each costs.
- Write the decision. Direction, the numbers, and the rejected options with one line each on why.
- Promote and delete. Integrate the winner following project conventions, then remove the prototype route. Grep for the fence comments.
What we learned
Round two of prompted variants is the signal to stop prompting and build the picker. The rejected options are the part that saves time three weeks later.