Task decomposition strategies

• Fixed sequential pipelines (prompt chaining) fit predictable work; adaptive decomposition fits work that depends on intermediate findings.
• Prompt chaining for reviews: analyse each file individually, then run a cross-file integration pass.
• Adaptive plans generate subtasks from what is discovered at each step.
• Reviewing many files in one pass dilutes attention, producing uneven depth and contradictory findings.
• Open-ended tasks start by mapping structure and identifying high-impact areas before planning.

• Pick prompt chaining for predictable multi-aspect reviews and dynamic decomposition for open-ended investigation.
• Split large code reviews into per-file local passes plus a separate cross-file integration pass.
• Decompose tasks like "add comprehensive tests to a legacy codebase": map structure, find high-impact areas, build a prioritised plan that adapts as dependencies appear.

• "A bigger context window fixes attention dilution." It does not; focus comes from splitting the work.
• "Adaptive is always better because it is flexible." It adds variability to work that should be consistent.
• "Per-file passes alone are enough." Cross-file issues need the integration pass.

Single-pass reviews of many files, fixed pipelines applied to investigations, and test plans that skip the structure-mapping step.

• 4.6Multi-instance and multi-pass review
• 1.2Coordinator-subagent orchestration
• 3.4Plan mode vs direct execution