Design task decomposition strategies for complex workflows

Put every file in one prompt so the model can see the whole change at once.

Attention spreads thin across a long context, so per-file defects get shallower treatment the more you add — and the review reads as complete.

Analyse each unit in its own focused pass, then run one pass whose only job is what crosses between them.

Write a fixed pipeline for an open-ended investigation so it is reproducible.

The plan is fixed before anything is known, so a discovery in step two cannot change step three. Everything unanticipated is skipped.

Predictable multi-aspect work takes prompt chaining; open-ended work takes decomposition that regenerates from what it finds.

Use dynamic decomposition everywhere — it is strictly more capable.

For a review whose aspects are known in advance it buys nothing and costs reproducibility, latency and the ability to diff two runs.

Match the pattern to the work. Chaining where the aspects are known; adaptive where they are not.

Run per-file passes and concatenate the findings.

Every defect that exists between files — a changed contract, a broken invariant — is invisible to every pass and therefore to the report.

The integration pass is a separate step with its own question, not a summary of the others.

Work a legacy codebase in file order until coverage is reached.

Effort lands wherever the directory listing points, so the risky code may never be reached and the count looks fine.

Map the structure first, identify what carries the risk, then work a prioritised plan that adapts as dependencies surface.