Prompt decomposition: split stages at verifiable handoffs

Last updated: 5 Oct 20268 min read

tutorial

Task decomposition separates extraction, calculation, decision, and presentation when those steps require different evidence or checks. It does not mean asking for long hidden reasoning. Each stage should emit a compact artifact that the next stage can validate: normalized facts, a computed total, a policy decision, then a reader-facing explanation. Use deterministic code for arithmetic and parsing when possible. Stop the chain if a required field or validation fails; otherwise one early mistake can be polished into a convincing final report.

Decision in practice

A shipping analyst needs a late-delivery credit recommendation from 86 tracking events. Stage one extracts event times and route IDs. Code computes transit duration and compares it with a 39-hour service threshold. Stage two asks the model to explain policy exceptions from the relevant contract section. Stage three drafts a customer-facing note only for cases that passed the calculation and policy gates. If one event has two incompatible arrival times, the chain marks it unresolved and skips the credit note. The team stores each handoff artifact with the prompt version for replay.

Stage 1 -> normalized tracking events with source IDs.
Gate 1 -> reject incompatible timestamps.
Code -> transit hours and threshold comparison.
Stage 2 -> exception decision with policy section.
Gate 2 -> approved or review; only approved reaches customer note.

Performance and operating cost

Multiple calls increase latency and cost, but they also create places to catch a failure before it becomes a customer action. For K sequential model stages, minimum request latency is approximately the sum of stage latencies unless independent stages run in parallel. Measure stage-specific error rates and end-to-end completion. Do not add a stage that merely paraphrases another stage without a check; it spends tokens and can alter the meaning. A compact handoff avoids carrying the entire conversation into every call.

Common Mistakes

Do not pass an unvalidated stage output forward.
Do not use a model for arithmetic that code can perform exactly.
Do not add stages without a measurable quality or safety gain.

Connected lessons

Continue with: Tool plans: separate independent reads from dependent actions.

Continue with: Least-to-most prompting: solve smaller dependencies first.

Continue with: Specialist agents: prove that delegation earns its cost.

prompt engineering

tutorial