Lesson 01 in reverse

In lesson 01 you saw when one prompt isn't enough. Now, before learning to build chains, Orbit stops you cold: most tasks aren't multi-step. If you split everything, you'll pile on cost, latency, and bugs without gaining anything.

Orbit's rule: one good single prompt beats five mediocre chained prompts. Multi-step pays when the problem really branches. Not when you imagine it could.

The four "don't split" signals

1. Output = direct transformation of the input. Format strings, parsing a known JSON, literal translation. One model call or one pure function. Not three.
2. Validatable with regex/parser. If success is checkable in code, you don't need a judge or a chain. A well-written prompt plus a deterministic test is everything.
3. Short, focusable prompt. If all the rules fit in 10-15 lines without contradicting each other, one call handles them.
4. No external data. No retrieval, no tools, no conditional on the world. If everything needed is in the input, there's nothing to orchestrate.

Real costs of over-splitting

Every step you add brings:

• Latency. 200-2000ms per LLM call. Three steps = possible 5-second response.
• Cost. Each step charges tokens in + out. Three steps repeat the same context three times.
• Failure points. Each step can fail, hallucinate, or return an unexpected format.
• Maintenance. Three prompts to maintain, three rubrics to write, three versions to track.

If the task fits in one prompt and you can validate it with regex, keep it in one prompt. Complexity is earned, not gifted.

The decision trick

Before you split, write the task in one line. If the line has one real decision, it stays in one prompt. If it has several genuinely independent decisions (classify AND extract AND summarize), then split.

Look at the two designs on the right. One is honest: one decision, one prompt. The other invents steps to look like engineering. Pick the one you'd ship.