We rebuilt PromptSewa — and your prompts were reset

Lesson 4 · 6 min

Breaking a big ask into steps

One prompt, one job. How decomposition turns an unreliable request into a sequence you can check.

Large requests fail quietly. Ask for "a launch plan with messaging, a timeline and a risk register" and you will get something that looks like all three, with the weakest part hidden in the middle. The model does not tell you which section it rushed. Decomposition is the fix: split the work into steps small enough that each result can be judged on its own, then run them in order.

Why one big prompt is risky

A single prompt asks the model to allocate attention across several jobs at once. Attention is finite. When one part is harder — usually the part with real constraints — the easier parts absorb the budget and the hard part gets a generic treatment. The output is not obviously broken, which is worse than an error. You ship a plan whose risk section says "consider security" and discover the gap a week later.

The three-step pattern

A workable decomposition has three moves. First, gather or restate the inputs: what facts, documents or decisions does this step depend on? Second, do one transformation: summarise, classify, extract, draft. Third, check the output against a stated shape before feeding it forward. That third move is what people skip. A step whose output is unchecked contaminates every step after it, and the final result inherits the error with no trace of where it began.

Example: instead of "write and schedule a launch campaign", run three prompts. One extracts the product's three differentiators from the brief. One drafts a message per audience using only those differentiators. One proposes a two-week schedule using only the drafted messages. Each output is small enough to correct before it is used.

Carry state explicitly

When you chain prompts, pass the previous result as input rather than assuming the model remembers. In a chat interface it may remember, but in an API call it will not. Even in chat, long conversations drift: earlier constraints get diluted by later turns. A short "here is the approved summary; use only this" line at the start of each step is cheap insurance and makes the pipeline reproducible outside the chat window.

When not to decompose

Not everything needs a pipeline. A single-sentence rewrite, a quick classification, a one-paragraph summary — those are already single jobs. Decomposition has a cost: more prompts, more places to make a mistake, more state to carry. Use it when the task has genuinely separable outputs, when one part needs a different kind of check, or when the result feeds a downstream system that needs a fixed shape.

Checklist

  • Can you name the single transformation in each step?
  • Does every step state its input explicitly?
  • Is each output checked before it is reused?
  • Would a smaller number of steps do the same job?

Finished this lesson?

Mark it complete to generate a certificate you can print or save as a PDF. We ask for an email only so you can be reached — no account is created.

All lessons