Prompt for reasoning

Amara Deep and the reasoning setting give models room to think before they answer. Prompting for them is a little different from prompting fast models.

Give the whole problem up front

State the goal, constraints and what a good answer looks like in the first message. Reasoning models plan; feeding requirements one turn at a time wastes that planning.

Prefer goals over procedures

“Find the cheapest schedule that satisfies every constraint and explain any trade-off” works better than a numbered algorithm. Prescribe steps only when the process itself is required (for example, a regulated checklist).

Do not ask it to think step by step

That instruction helps fast models. Reasoning models already do it; repeating it can make the visible answer verbose. Ask instead for the format of the final answer.

Ask it to check its work

“Before answering, verify the result against each constraint and say which ones you checked” catches more errors than asking for confidence scores.

Budget the effort

Use reasoning.effort: "low" for routine questions and "high" only where errors are costly; the difference in tokens, and therefore price and latency, is large. Measure on your own evaluation set.

Keep reasoning out of your outputs

Reasoning is returned separately from the answer when you ask for it. Do not paste it back into later turns as part of the conversation; it costs tokens and can anchor the model on earlier mistakes.