What prompting techniques help when using Claude to create and execute a research or project plan?

One approach is to turn the model's thinking/effort level up to at least "high" when generating a plan, then explicitly tell it to run deep research so it checks its own output against itself. It also helps to prompt it with things like "ask me something I haven't asked before" to surface gaps, and to keep talking to it iteratively rather than doing one-shot prompts. This is largely a matter of prompt engineering skill that improves quickly with practice. In one example, a plan created this way was partially executed directly and the remaining items were handed off to an engineering team.