One prompt, too many jobs
A prompt that asks the model to read, classify, summarize, and translate all at once spreads attention thin. Decomposition breaks that into separate focused subtasks, each with a clear single goal.
Ways to split
- By stage so each step does one transformation, like extract then format.
- By aspect so one call rates clarity and another rates correctness.
- By branch so the model first classifies the case, then routes to the matching handler prompt.
Why it helps
A focused prompt has a simpler objective, fewer ways to go wrong, and shorter context. You can test, cache, and tune each subtask alone, and swap one without disturbing the rest. Errors become easier to localize.
The tradeoff
More calls mean more latency and cost, and you must wire the pieces together so the output of one feeds the next. Decompose only as far as the gain in reliability justifies the added orchestration.
Key idea
Decomposition splits a broad request into focused subtasks by stage, aspect, or branch, making each easier to test, tune, and debug, at the cost of more calls and orchestration.