Give a Reasoning Model the Budget and the Success Criteria
OpenAI's reasoning guide gives short advice. Keep prompts simple and direct. Skip "think step by step," because these models reason internally. Be very specific about the end goal, and state the constraints you want the answer to respect.
Here is a planning task where that advice pays off. You have 64 GPU-hours for a sweep: 4 learning rates × 3 seeds × 2 model sizes = 24 runs. A small run takes 2 GPU-hours and a large run takes 4. Ask for a schedule.
The arithmetic says the full sweep does not fit:
12 × 2 + 12 × 4 = 24 + 48 = 72 GPU-hours, 8 over budget.
A good answer says so and proposes a cut. Dropping to two seeds for the large model gives 12 × 2 + 8 × 4 = 56, which fits and keeps every learning rate at both sizes.
Give the destination, not encouragement. Put the budget, the run costs and the success criteria in the prompt. The plan must total 64 GPU-hours or less, cover every learning rate for both sizes, and name any cut it makes and why. That is what the guide means by specific parameters for a successful response.
Check it with code, not by reading. Parse the plan, sum the hours, and confirm the coverage. Then run a small experiment: vary the budget and run costs across 30 cases, some feasible and some not, and count the plans that pass. Run the same 30 cases with a long prompt padded with "think carefully and double-check," and compare pass counts item by item. This course has not run it; the checker is about ten lines.
Chapter 10 derives how long work waits for a busy worker, the other half of planning compute. The reasoning model's hidden work is not what you grade. The plan is.
