Key answer
State the finished deliverables, their sources and which results must precede others; AI can arrange implementation steps. “Write a report, make slides and list actions” names three outputs without saying whether they share conclusions or require separate analyses.
Full explanation
Distinguish independent tasks, dependent tasks and shared constraints. Data checking and layout exploration can run together; final charts depend on cleaned data; reports and slides must share statistical definitions. Content dependency does not automatically require user approval at every stage.
Some experiments support decomposition. Zhou et al.'s least-to-most prompting first decomposed problems, then used earlier answers to solve later ones, substantially improving GPT-3 code-davinci-002 on SCAN compositional generalization. This concerns particular reasoning tasks and demonstration design, not a universal rule for office work. Paper
Anthropic's engineering guide distinguishes chaining, independent parallel work and dynamically delegated orchestration. Choose according to real dependencies. Official guide
For example:
Deliver an analysis report, slide deck and action list from the attachment. Verify data and establish conclusions first. Use one statistical basis; slides distill the report and actions address its findings. Material organization and layout preparation may run in parallel. Continue clear work; report conflicts and missing information only when conclusions cannot be determined.
This defines output relationships without scheduling every search/edit. Check completeness and correspondence of shared figures, conclusions and actions; three generated files establish quantity only.