When building is cheap, build to learn.

Everyone's talking about how easy it is to build with AI right now.

Spin up an agent, write a prompt, ship. Done.

But if you've actually deployed something in the wild built mostly by AI, you know it's more complicated than that. Especially when you care about quality.

We're currently building a Mac app for Pretty Prompt.

Before we started, we made a call that felt like a detour: build a second Mac app first. Something we wanted for ourselves. A totally different problem.

We'd never shipped a native Mac app before, so instead of going straight into the main thing, we used this as a way to learn everything you only learn by actually building.

The biggest takeaway: when code is cheap, build to learn.

Read all the docs you want, watch as many youtube videos about LLMs... But the real lessons come from shipping.

A few things I learnt by doing this:

↳ Rebuilding is a strategy. When code is cheap, you can rebuild often. Don't like how it came out? Try again. Go further just to see what happens. "What happens" is where the learning lives.

↳ Writing is still the basis of the product. I still hear Charlie in my head: "this is a sh*t ticket" 🫠. Lazy writing means burning tokens and wrong outputs. If you didn't write it, don't get upset when the agent didn't build it. Gave too much? Then be worried about the agent going on a tangent. (An advantage for PMs).

↳ Describe the why, not just the what. Don't tell the agent what to build. Tell it how it should feel. The intention. The user context. I've found that when agents understand the why, they “go further”.

↳ Taste is now the expensive thing. When code is cheap, taste is expensive. Spend more time with your users. Learn their real problems. Get feedback and actually absorb it. Don't outsource your taste. (Well, taste was always the USP)

What's been your biggest learning while building with agents?