Breaking a goal into steps before starting. Models are noticeably better at multi-stage tasks when they write a plan first, for the same reason people are. The difficulty is knowing when to stick to the plan and when the world has changed enough to make a new one.
Technical view
Approaches include upfront plan generation followed by execution, and interleaved planning where the plan is revised after each observation. Upfront planning is cheaper and brittle when reality diverges; interleaved planning adapts and costs more. Hierarchical decomposition into sub-tasks with their own contexts is the standard way of handling tasks too large for one context window. Explicit plans are also valuable for oversight, since a human can approve a plan more easily than a trajectory.
Related
Part of AI Terminology Index, section: Agents and Tools.