CORE GUIDE
Prompt decomposition
Break a complex request into explicit reasoning or production stages when one prompt mixes different responsibilities, evidence needs or verification methods.
Mental model
Prompt decomposition is responsibility separation at the interaction layer: each stage receives a clear input, produces a bounded artifact and hands that artifact to the next stage instead of asking one generation to research, decide, create and verify simultaneously.
Why it matters
Large prompts often hide several different tasks behind one instruction. The model may skip research, blur assumptions into conclusions or revise facts while polishing style. Decomposition makes intermediate artifacts inspectable and lets deterministic tools or independent verification enter at the right stage. It also makes retries cheaper because one failed stage can be repeated without regenerating everything.
01
Split by artifact and verification method
Identify distinct responsibilities such as clarify the question, gather evidence, extract claims, plan, draft and verify. Give each stage only the context required for that responsibility and define its output contract. Insert tools or external checks where the artifact can be verified directly. Avoid decomposition that creates ceremony without a different responsibility or verification boundary.
02
Example: research brief in three stages
Instead of prompting 'research this market and write a recommendation', the workflow first decomposes the question into evidence needs, then gathers and records sources in a claim–evidence matrix, then synthesizes a recommendation from that matrix. A disagreement in sources can be reviewed before persuasive prose hides it.
Common failure modes
- Breaking one simple task into many prompts that add latency but no clearer responsibility.
- Passing the entire growing transcript into every stage instead of bounded artifacts.
- Letting later drafting stages silently change facts established earlier.
Engineering heuristics
- Decompose where stages need different evidence or verification.
- Name intermediate outputs such as evidence matrix, plan or draft rather than vague 'thoughts'.
- Keep handoffs small enough that another stage can inspect them independently.
Takeaways
- 01Decomposition makes complex generation observable.
- 02Intermediate artifacts create natural verification boundaries.
- 03More stages are useful only when they localize responsibility or failure.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.