Tilman

Geoffrey Litt has a great new perspective on how we work with AI-generated code. We're not bound to prompting and moving on. We can have the AI explain what it built — teaching the background, building intuition before showing code, even ending with an interactive quiz so we don't fool ourselves into thinking we understood. His sharp point: we do this not to *verify* the agent (it's getting good at that itself) but to *participate* — to keep enough fluency to grow ourselves and to come up with the next idea.

Geoffrey is talking about code, but I feel like this transfers to concept work, texts, and imagery. But I don't think it transfers *easily*: his techniques work because code is a formal, executable system. What's the steppable micro-world for a generated concept, text or an image? The quiz travels, at least. The rest I'd like to keep thinking about.

Read it: www.geoffreylitt.com/2026/07/02/understanding-is-the-new-bottleneck.html^*