Task Decomposition Prompting
Split a complex job into chained simpler prompts for better control, cheaper models on easy steps, and inspectable intermediate outputs.
Multi-step work often performs better as a chain of simpler prompts than as one giant prompt. Example: customer support as intent classification JSON, then a per-intent response generator. Benefits go beyond raw accuracy: intent classification for routing enables monitoring and debugging of intermediate outputs, parallelization of independent steps, and easier prompt writing.
Costs: higher perceived latency to the final answer if intermediates are hidden, and more model queries — though each prompt is often shorter and weak models can handle easy subtasks. Bloated monolith prompts (thousands of tokens after iteration) are a signal to decompose. In agent systems, decomposition aligns with plan-validate-execute and hierarchical planning.
One recurring split: separate a chain-of-thought step from the value that gets retained downstream — steps that must think from steps whose output is retained. A "generate a plug-in concept" task, for example, can decompose into a brainstorming step (produce several options, pick the best) and a separate elaboration step (write up the chosen idea in detail): emit only the chosen concept for downstream tasks under task I/O schema design, keeping the reasoning trace from a templated-prompt task out of the output those tasks actually consume, rather than relying on prefix/suffix extraction to strip it back out of one combined completion. Keep local retries inside the task when possible rather than adding a cyclic workflow topology. When subproblem answers must feed the next prompt for harder-than-exemplar cases, use least-to-most prompting.
Sources
- AI Engineering (Chip Huyen), ch. 5
- Prompt Engineering for LLMs (John Berryman), ch. 9
- Least-to-Most Prompting (Zhou et al.), pp. 1–15