product-design

When one agent becomes many agents, the bottleneck moves from model capability to task, context, permission, state, and acceptance management.

A product note on why useful AI often removes the friction that makes thinking happen, and what a slower, less obedient AI interaction might look like.

When you cannot describe the website you want, that is not always a prompt failure. Some design judgment is tacit and only becomes visible through iteration.

A practical note on AI writing: voice is not copied from a sample, it is found in the delta between the model's default draft and the version you actually publish.

Adding an agent to a product is not the same as making an agent-first product. The hard parts are permissions, tool boundaries, context, task contracts, and treating AI as a real user.

A product-builder reading of model interpretability work: if internal concepts can be found and steered, the UI should stop pretending behavior is only prompt text.

My practical take after reading prompt engineering research: prompt templates are useful, but treating them as universal recipes is the wrong level of abstraction.