Two shapes for the same task
You can ask the model to execute a process ("first do X, then do Y, then do Z") or to achieve an outcome ("return Z, given that you have access to X and Y"). Both can work; they fail differently.
When step-by-step wins
- The task has a known, brittle order (parse → validate → transform → emit).
- You need traceability — each step's output is logged.
- The model has been overshooting (skipping verification, jumping to a guess).
- You're using tool calling — explicit steps line up with explicit tool calls.
When open-ended wins
- The task has many valid paths (creative writing, planning).
- Steps are obvious to the model and writing them out is just noise.
- You're using a reasoning model that already plans internally.
The pivot
If a step-by-step prompt is producing rote, mechanical outputs, soften it. If an open-ended prompt is skipping required checks, harden it. Move along this dial deliberately, not by accident.