Rapid exploration with AI in product design

Explore wide with AI, then choose narrow with a person. Set your criteria before you generate anything, let Figma Make or Claude Design produce many directions in hours, and cut them down to two or three that you test with users. Exploring is cheap now. Choosing is the skill.

This is how I run rapid exploration in step 03 of my process.

Why explore with AI at all?

A designer used to sketch a handful of directions in a day. With AI tools you can look at twenty. That changes what you can learn early, because a wrong direction costs minutes instead of days.

It also creates a new risk. Twenty directions that all look fine are harder to judge than three that were thought through. The method below is built to avoid that.

How do you run a fast exploration?

1. Start from the hypothesis, not from the screen

Exploration comes after you defined the problem. In my process, step 02 gives me a testable hypothesis and the riskiest assumption. Every direction has to help test that one. If it does not, it is decoration.

2. Benchmark before you generate

On an internal AI platform at a global insurance group, I first benchmarked how leading AI tools handle sources, projects, roles and long tasks. Only then did I explore directions for the chat, the projects and the sources view. The benchmark gave the exploration a vocabulary and a bar to clear.

3. Write the criteria first

Pick three to five criteria before you see any result. For example: does it make the risky assumption testable, is it clear without explanation, can it be built on the existing design system. Judging against a list protects you from falling for the first polished output.

4. Generate wide with the right tool

I use Figma Make and Claude Design for this. Give them real constraints: the design system, real content, the user’s task. Ask for directions that differ in structure, not only in colour. Ten variations of the same layout teach you nothing.

5. Cut hard

Cluster the results, throw out what fails a criterion and keep two or three. Write one line for each rejected direction. That note is useful later, when someone asks why you did not choose it.

6. Make it real

The survivors become working prototypes, built with Claude Code on the real design system. A direction that only exists as a picture has not been tested yet.

What goes wrong?

• Sameness. Tools return the average of what they have seen. Push for structural differences.
• Fake content. Placeholder text hides problems. Use real data, real names and real error states.
• Attachment. The first good output feels like the answer. Your criteria are the counterweight.
• No decision. Exploration ends with a choice and a reason, written down.

Key takeaways

• Define the hypothesis and the criteria before you generate.
• Benchmark first, then explore.
• Ask for different structures, not different colours.
• Keep two or three directions and make them real.

Next in the process: how coded prototypes replace the handover, and the full six steps.

Sources: Claude Design launch, VentureBeat