I’m comparing DeepSeek R1 on multi-part questions and have noticed that a long prompt can produce a different emphasis than asking the parts one at a time. For example, the model may address a constraint early in a combined prompt but overlook it when several requirements are packed together.

Have you found a useful rule for deciding when to split a question into smaller prompts? I’m interested in whether the difference is mostly about prompt clarity or whether breaking it up can cause the model to lose important connections between the parts.