How I build with AI.

I first needed a tool like this while designing Ultramarine, an ASMR paint-mixing game for oil and acrylic painters, built so the colors mix the way real paint does. It had to mix like paint, move like paint, and look beautiful enough for an audience of artists.

Paint is a non-Newtonian fluid. It moves with the resistance of a gel and mixes like pigment, not like light on a screen. So I built the game on a real fluid physics engine, and getting that to run on a phone meant writing the simulation from scratch.

A model like that lives or dies by its parameters. AI wrote the helper code for the harness. I dragged sliders until the paint read as real, down to that perfect ketchup-y ooze, then had it write the parameter set straight into the app. The tool above is the actual harness for the gel movement, not a mockup. I loved it, and I happily threw it away once the app launched.

Ovo is a women's health app I designed in Figma. It lets women journal the messy way they actually do, then structures those entries into insights about how we work, exercise, eat, and relate to ourselves. The goal is insights that fit the routine a woman already has.

As a PM, it feels pretty great to get closer to building real software, so for this app I intentionally went HAM on test-driven development: a design system, standard components, and a CI/CD pipeline, so anything I generated with Claude could plug right in. I designed the Journal input screen and the main flows myself. Then I used the pipeline to generate a new onboarding flow and custom screens for logging food and fasting protocols.

AI has removed the need for juniors to do the grunt work. That sounds like progress, but grunt work was how the apprenticeship model worked: you learned judgment by doing the small things badly first, next to someone who did them well. Take that away and the next generation of architects, engineers, and lawyers has no path to expertise, while the experts who already have it risk letting it go soft.

Most AI tutors make this worse. They optimize for answer fluency: you ask, it explains beautifully, and you move on feeling smarter than you are. I think of that as borrowed intelligence, fluent output that skips the thinking it was supposed to build. Learning sticks through experimentation and failure, so the fix is to put the human back in the driver's seat.

That is what the tutor does. It runs on the Claude API and is grounded in the National Research Council's How People Learn, Daniel Jackson's concept design, and Anthropic's own interpretability work. A classifier reads every turn and picks a mode: acquisition, practice, self-test, or reflect. Ask for a definition and it answers. Ask for a judgment call and it asks what you would do first, then holds its answer until you commit. The gate only arms in practice and self-test, which is the whole design: teaching mode shouldn't withhold, testing mode shouldn't hint. For the prototype I loaded it with widely taught PM concepts, but the knowledge base is swappable for any indexed body of knowledge.