Back to work
Late-night AI experiment

Next Experiment

A one-line note on what I was poking at — the tool, the idea, and the itch I was trying to scratch.

Stack
Claude + Figma MCP
Time spent
5 hours
Date
Mar 2026
Result of the Next Experiment experiment
What I wanted to try

I kept hearing that these tools could take a rough idea to a working screen in minutes, and I didn't fully believe it. So one night I opened a blank file, picked an app idea I'd been sitting on, and decided to find out where it would actually break.

What came out of it

A half-finished prototype, a much better sense of what these tools are actually good at, and one screen I genuinely liked enough to keep. Also a new respect for writing things down before you prompt.

How it went

The actual process, mess included

  1. Placeholder screenshot of the first rough attempt
    step_01

    The first attempt was a mess

    I started by just dumping the whole brief into the prompt and hoping for the best. What came out looked fine at a glance, but the moment you clicked around, nothing made sense — buttons that went nowhere, states that didn't exist. Classic first-draft energy.

  2. Placeholder screenshot of component-by-component iteration
    step_02

    So I changed how I was asking

    Instead of asking for a whole screen, I started asking for one component at a time, with the states spelled out like a checklist. Slower, but the output suddenly started behaving like an actual interface instead of a poster of one.

  3. Placeholder screenshot of the pivot direction
    step_03

    The pivot that made it click

    Halfway through I realized the interesting part wasn't the generation at all — it was the editing loop. Once I treated the tool like a very fast junior collaborator instead of a vending machine, the whole thing got fun.

  4. Placeholder screenshot of the final result
    step_04

    Where it landed

    By hour five I had something that felt genuinely usable — rough around the edges, but real. It wasn't the thing I set out to build, which turned out to be the best part.

Screenshots & scraps

Stuff worth keeping

output_v1 — first pass, before the prompt rewrite
output_v1 — first pass, before the prompt rewrite
component_states — the checklist approach
component_states — the checklist approach
edit_loop — iterating on the recap screen
edit_loop — iterating on the recap screen
final — hour five, ship it
final — hour five, ship it
// notes to self

What surprised me

The tool was never the bottleneck — my own vagueness was. Every time the output drifted, it was because I hadn't decided something yet. Turns out prompting is mostly just designing out loud, and I want to do more of these on a regular cadence.