A Reliable AI Design Generation Workflow
Generate stronger interface directions by separating product reasoning, visual direction, specification, and critique into distinct passes.
Updated 2026-08-17 · Published by Claude Designer
Key takeaways
- Separate divergent concepts from convergent specification.
- Ask for rationale and tradeoffs, not only visual output.
- Use explicit critique gates before implementation.
Do not ask one prompt to do four jobs
Concept generation, product reasoning, visual styling, and implementation specification require different kinds of attention. Combining them in one request often produces polished but shallow output. Split the work into passes and preserve the decisions from each pass as constraints for the next.
Use a four-pass sequence
- Frame: define the user job, conversion goal, constraints, and missing information.
- Diverge: request two or three meaningfully different interaction approaches.
- Converge: select one direction and specify hierarchy, components, states, and content.
- Critique: test accessibility, responsiveness, edge cases, trust, and business alignment.
Control variation deliberately
When exploring, tell the model which dimensions may vary—navigation model, information density, interaction pattern, or visual tone. Keep product facts fixed. When converging, lock the chosen structure and allow only small refinements. This prevents every iteration from becoming an unrelated redesign.
Define a handoff contract
End with an artifact engineering can evaluate: route and component map, responsive behavior, data dependencies, error handling, copy, analytics events, and acceptance criteria. A screenshot can communicate appearance; a specification communicates how the product should work.