Reasoning models
Ask for a defensible result, not a reasoning ritual.
Brief reasoning models with clear goals, evidence and acceptance criteria, and understand why step-by-step instructions are not a universal improvement.
Updated
Older prompting advice needs model context
A technique demonstrated on an earlier model is not automatically a requirement for a current reasoning model. OpenAI's reasoning guidance recommends direct briefs rather than requests to expose an internal chain of thought. Anthropic likewise maintains model-specific guidance rather than one timeless recipe.
You can still request an inspectable answer: supporting sources, calculations, assumptions, alternatives and a concise rationale. Those help you review the result without pretending that a long explanation proves the model's internal process was correct.
Evaluate [decision] using [sources and criteria]. Return the recommendation, evidence supporting it, material assumptions and the strongest alternative. Show calculations needed to verify numerical claims. Keep the rationale concise and identify unresolved evidence.
Define the hard part of the task
Name what makes the decision difficult. Is there conflicting evidence, an irreversible choice, a hidden dependency or an ambiguous definition? Give the model the material needed to confront that difficulty. A request to ‘think harder’ without a decision criterion leaves the target unclear.
Illustrative coding task: a form intermittently loses a saved value. The useful brief supplies the reproduction, expected state and relevant timing, not an instruction to produce pages of speculation.
The difficult part is [uncertainty]. Compare [options] under [constraints]. Identify the assumptions that could change the choice and the smallest check for each. Do not fill missing evidence with an invented estimate.
Specify checks without micromanaging every step
Sometimes a procedure is mandatory: a regulated process, a supplied calculation definition or a compatibility requirement. State it. Otherwise, define the result and its checks while allowing the assistant to choose an approach. Ask for an early plan when the approach requires your approval.
Produce [deliverable] for [audience]. Preserve [fixed requirements]. Before finishing, check [specific acceptance criteria]. Ask before [consequential action]. If two approaches would lead to materially different outcomes, present the trade-off before proceeding.
Review the conclusion against an alternative
A useful challenge changes one plausible assumption and asks whether the recommendation survives. This is more informative than repeatedly requesting greater confidence. Verify important claims against the underlying source even when the rationale sounds careful.
Assume [plausible alternative]. Does your recommendation change? Show which evidence and constraint drive the difference. Distinguish a genuine change in the decision from a change in wording, and list what we should verify next.
Common questions
Should I always say think step by step?
No. Current reasoning-model guidance is model-dependent. Ask for the evidence, calculations and concise rationale needed to review the answer, and check the official guidance for the model you use.
Does a detailed explanation prove an answer is correct?
No. Check the conclusion and supporting evidence. A convincing explanation can still contain a faulty assumption or calculation.
Keep exploring
Sources and editorial note
Guidance checked 9 October 2026. These AI-assisted guides and original templates are reviewed against the linked sources. Results depend on your inputs and available tools. Model recommendations are editorial starting points, not comparative test results.