The AI Mise en Place
TL;DR: The model should not have to do the grocery shopping before it cooks dinner. Good AI products prepare the station first (i.e., context, data, tools, and deterministic work) then use the model where judgment matters.
Building an AI product is a lot like running a professional kitchen. The interesting part is cooking. A great deal of the work happens before anyone starts.
Chefs call this *mise en place*: ingredients portioned, sauces prepared, knives where they belong, station organized. When an order arrives, the chef should be cooking—not looking for parsley.
AI products benefit from the same discipline.
The model is the visible part, but much of product quality is determined by what happens before the prompt reaches it. What context has already been assembled? What data retrieved? What calculations handled deterministically? What tools are available? What decisions can the system safely make in advance?
You can ask a model to find the data, interpret it, do the math, decide what matters, and write the answer. Sometimes this works. It is also roughly equivalent to handing a chef a grocery bag and telling them table six is waiting.
A good system tries to make the model’s job smaller before it ever calls the model. Fetch the data. Resolve the easy choices. Put the right tools within reach. By the time the model gets involved, ideally the remaining problem is the one you actually wanted a model for.
Mise en place does not tell a chef how to cook the fish. It makes sure that, when the fish hits the pan, cooking the fish is the only problem left.
The model matters. But the product is often everything you arrange around it before the real work begins.