AI in the design workflow

The workflow

AI shows up differently at each stage of the work, and treating it as one tool undersells (or oversells) what it’s for:

• Research synthesis - the highest-leverage stage. Fast first-pass themes across a large interview set, freeing human attention for the judgement calls.
• Exploration - wide, cheap divergence. More directions in less time, so the team argues about the right things instead of running out of options early.
• Systemizing - turning a scattered set of decisions into a documented, reusable system (see Design systems).
• Coding - Figma-to-code and prototype-to-code, closing the gap between what’s designed and what ships.

Prompt patterns

The patterns that actually hold up are boring: ground every prompt in the team’s own principles and component inventory rather than general taste, and separate “generate options” prompts from “critique this” prompts - mixing them produces agreeable mush instead of useful friction.

Guardrails - what AI must not decide

• Strategic trade-offs. It doesn’t have the business context to weigh them.
• What “good” means for this product. That’s a judgement call, not a pattern match.
• The final call in a critique. AI can surface gaps; a human still owns the decision and carries it.

How this links to the Lab

Every claim above started as an experiment in the Lab - a specific question, a specific build, a documented outcome. This framework is the generalised version of what held up across more than one of those.