What it is
Reasoning prompts ask the model to write out its intermediate steps before it gives an answer. Instead of jumping straight to a number or a decision, it works through the parts in order: what is known, what follows from that, and what the conclusion should be. The written steps are not a guarantee of correctness. They are a scaffold that keeps the model from skipping the hard part.
The same idea covers several related habits. You can ask for a step-by-step breakdown, ask the model to plan before it writes, or ask it to check its answer against the question. All of them trade a longer response for a better chance of getting the logic right.
When to use it
Use it for problems with more than one moving part: arithmetic with several terms, comparisons across a few documents, planning a sequence of tasks, or anything where the wrong answer usually comes from dropping a constraint.
Skip it for simple lookups, short rewrites, or anything where you mostly care about tone. On easy tasks the extra steps can introduce invented details that were not in your input, so the answer gets worse rather than better. If you are unsure, try the task both ways and compare.
How to use it
- State the problem precisely, including every constraint and the units or format you expect.
- Ask for the reasoning first and the final answer last, in a clearly marked line.
- Tell the model to list what it knows before it calculates, so gaps show up early.
- Add a check step: ask it to confirm the answer against each constraint you gave.
- Read the steps, not just the answer. If a step is wrong, point at that step and ask for a correction rather than starting over.
For example, suppose you are pricing a small print job. You tell the model: paper costs 0.04 per sheet, ink costs 0.02 per sheet, there is a 15 setup fee, and the client wants 250 copies. Ask it to show the per-copy cost, the run cost, then the total. A good response writes 0.06 per copy, multiplies by 250 to get 15, adds the setup fee, and ends with 30.00 on its own line. You can verify that in your head, which is the point.
What a good result looks like
The steps should be short and checkable. Each one should follow from the last, with no leap that you cannot trace. Numbers should carry their units. The final answer should sit alone so you can copy it without hunting.
A weak result hides the reasoning in one dense paragraph, or produces steps that quietly restate the question. If that happens, tighten the instructions and ask for one step per line. You are not looking for a longer answer. You are looking for one you can audit.
A practical lesson on asking a model to reason in visible steps, when that helps, and when it just adds noise. For anyone writing prompts that involve multi-step problems.
Read this page, then check whether AI assistants can read yours: one scan, six pillars, a score.