I asked a client what she wanted to build. She said #usagi.

Every session I run starts the same way: what's a real thing in your life you'd actually want to build with this? I'm usually braced for "summarise my emails." This past weekend a client said: a Usagi. The unhinged little rabbit from Chiikawa. Living on her Windows desktop. Watching her work.

It took me a second. Not because it was a bad idea — because it was a better one than anything on my list, and I hadn't seen it coming. I'd quietly assumed "useful" had to mean "spreadsheet-shaped." I nearly steered her somewhere more practical. Glad I didn't.

The process is identical either way

Here's the bit that matters. The route from an idea in your head to working software doesn't change with the subject matter. You write a plan. You give the AI precise context — not a vague wish, but specifics: how big, how it behaves, what happens when you click it. You prompt with care, read what comes back, and iterate instead of accepting the first draft. Whimsy doesn't get its own methodology.

She did exactly that in Claude Code, running Anthropic's Fable 5 model, and got back considerably more than she'd asked for. That gap between what you typed and what you got is where people either fall in love with a tool or bounce off it — and prompt engineering is what decides which. Same gap I walk through with clients doing far less charming work: AI training for professionals is the rabbit-free version of this argument.

The process is exactly the same whether you're building a financial model or a cartoon rabbit.

What the prototype actually does

Watch the clip. Usagi loads the moment she logs in. She can drag her anywhere on screen, or let her wander around it on her own. She can resize the animation, and toggle the background between transparent and opaque depending on whether Usagi should float over the wallpaper or sit in her own frame. Not a mockup — software that runs.

And it's a prototype in the proper sense: obvious next steps already visible. Have Usagi flag the next calendar meeting. Have her twitch when an email lands. None of that needs a new skill — it's the same plan, context, prompt, iterate loop pointed at one more feature.

The real lesson on how to learn AI

If you want to learn AI in a way that sticks, point it at something you genuinely care about and would use in real life. Not a tutorial project. Not the exercise from a course. The thing you'd open again tomorrow. Curriculum teaches you the syntax; passion is what keeps you debugging a broken animation at 11pm because you want to see the rabbit move properly — and that's where the learning actually happens.

One weekend on a desktop pet taught my client more about prompting, context and iteration than a week of generic exercises would have, because she cared whether it worked. Find the thing you love, point Claude Code at it, and build.