Back to lab

Choose, Don't Generate

  • Built with Claude Code
  • React, TypeScript, Vercel
  • 6 working days
  • 3 versions
  • 74 commits
  • 8 merged pull requests
3 min read

This is a practice project for designers, not an AI design tool. The site makes no AI calls: the four practice cases are written in advance.

As the tool says on its first screen: “Directions are pre-written: drafted with AI, edited by me.”

AI can draft design options in seconds. Choosing one, and saying why, is still the designer’s job. This tool trains that part.

How a round works

  1. Pick a case. For example, a fitness app: get a new user to their first booking in under 2 minutes.
  2. Get a situation. Drawn at random, like “You have 2 weeks and one engineer.”
  3. Choose and change. Pick one of three directions, each with a sketch and its trade-offs. Change it to fit the situation.
  4. Say why. Write why, and what you give up. You keep it as a decision record.
One case, one situation, three directions. The situation changes which one is right.

What it trains

  • Choosing for a situation. The same case has a different best answer with 2 weeks or 2 months. You learn to argue from the constraint, not from taste.
  • Saying why in plain words. Answers that are too short are blocked with a hint. A reason has to point at something real.
  • Naming the cost. Every choice gives something up. Writing it down is what makes a decision you can defend later.

Why it matters to a product team

  • Fewer decisions argued twice. The record holds the goal, the situation, what was chosen and rejected, why, and the cost. That is a format a team can reuse in reviews and handoffs.
  • Sharper design reviews. A 90-second round is a warm-up for critique: people practise reasoning from a constraint before they discuss real work.
  • Using AI without giving up judgment. AI makes the options; a person chooses and signs the reason. Every screen says who drafted what.
What you keep: a decision record. Copy it as Markdown, download it, or share a read-only link.

These are the uses I designed it for. It hasn’t been tested with teams yet.

How I built it

I directed an AI coding agent, Claude Code, with a short rulebook and five review gates where it had to stop and wait for me. I chose the four cases, edited all 12 direction cards, approved 12 situations and 24 changes line by line, and set the look. The agent wrote the code and the first drafts.

What real AIs did with it

The morning after “Your own case” was built, I ran three test goals through ChatGPT, Claude and Gemini.

  • All three answers read into three directions, with sketches drawn from their text, and go through the normal flow.
  • ChatGPT’s own copy button squashes each block into one long line. Most ChatGPT users would have hit this. The reader now splits the text at each label.
  • Claude came closest to my edited cards, with real numbers and honest costs. Gemini used hype words and wrote a situation that gave the answer away. That is why those screens say “Drafted by your AI”.
  • Seeing real cards side by side, I moved “Best when”, the line that decides, above the gains and costs on every card, so people read it before the lists.

Limits and next steps

There are no usage numbers: the site has no analytics, by design, and I haven’t tested it with real users yet. Next, I want to watch a few designers use it. Do the situations change their choices? Do their reasons get more specific?

Contact