The instinct: "I'll put it all in one prompt."

Orbit works the other way around. Before touching the model she sketches the task on the whiteboard and cuts it into steps. She'll make you do the same the whole way through this track.

The first reaction is always to make one monster prompt. And for small tasks it works. For real problems it breaks fast, for four reasons:

1. The model gets confused when many rules are active at once.
2. You can't iterate partially. to fix one part you have to touch everything.
3. You can't cache. every call starts from scratch.
4. You can't evaluate. a system-wide failure doesn't tell you which piece broke.

The alternative: split

When you, reading the task, already identify conceptually distinct stages (classify, then extract, then summarize), your system should have those stages as separate steps. Each step is a focal prompt: one input, one typed output, one decision.

The criterion: if your task is naturally described as "first X, then Y, then Z", you probably want 3 prompts. If it's described as "take this and give me that", probably one.

Look at the two designs on the right. One is a monster prompt. The other is a chain. Decide which one scales. and why.