Apply decomposition techniques for complex problem solving
Break complex problems into units by data boundary, by step or by expertise so each call is focused, fits in context and can be validated, and aggregate results deterministically in code.
Key points
- 1
Decompose by data boundary when the input is too large or consists of independent units: chapters of a rulebook, patient records, accounts, source systems. Process each unit independently (often in parallel), then merge.
- 2
Decompose by step (prompt chaining) when one call is juggling several judgements and errors between them compound: extract as structured data, then classify, then draft, with programmatic checks between steps.
- 3
Decompose by expertise when a problem needs independent specialist perspectives (medical, financial, legal): run specialist prompts in isolation with only the evidence each needs, then reconcile in a synthesis step. Sequential hand-offs anchor later assessments on earlier ones.
- 4
Keep aggregation, counting and arithmetic in code. Model-side aggregation over thousands of units is unreliable and unauditable; per-unit structured extraction plus deterministic aggregation is exact and cheap.
- 5
Do not mix many independent units (patients, customers) into one prompt to "spot trends"; it degrades per-unit accuracy and blurs boundaries, including privacy boundaries.
- 6
Very large evidence sets (hundreds of thousands of tokens) are summarised per source into structured findings with references, then synthesised in one step for a consistent narrative, then verified so every claim maps to evidence.
- 7
Over-decomposition is a real failure mode: 14 micro-steps add latency, cost, compounding error and maintenance. Merge steps that share an input and need no validation between them; keep a gate only where a programmatic check adds value.
- 8
Lossy shortcuts are traps: a one-paragraph summary of a 400-page rulebook discards the obligations a gap analysis needs.
- 9
Letting an agent choose which sections to read is not decomposition; when the input has a natural structure, partition it deterministically so coverage is complete and auditable.
- 10
Chain-of-thought inside one call, more few-shot examples or majority voting reduce some errors but do not fix inter-step consistency failures; explicit intermediate outputs with checks do.
Read the source
Test yourself on Apply decomposition techniques for complex problem solving
Ten questions, with the answer and explanation after each one.