What it is
Asking for reasoning first means putting the working before the conclusion in the answer: the calculation before the total, the quotes before the rating. Order matters. When a model states the answer first, what follows tends to justify it. When the steps come first, the conclusion is built from them, and you can see each one.
This used to be advice for getting better answers: the “think step by step” line. That has changed. Anthropic’s newest Claude models have thinking always on and decide how much each request needs, OpenAI’s reasoning models reason internally, and Google says Gemini’s 2.5 and 3 series models do too. For those models, telling them to think adds little. What still helps is asking for the specific working you need to check, as part of the answer.
That is a different request from asking to see the model’s thinking. “Show each step of the calculation with the figures used” asks for a deliverable. “Write out your full reasoning” asks for the model’s internal process, and on some current models that request can be declined.
When it helps
Use it for calculations whose result goes somewhere that matters: a budget request, a price, a forecast, a pay change. Each figure on its own line lets you spot the step where a monthly price was multiplied by the wrong number of seats.
Use it for judgments against criteria, such as screening applications, scoring proposals or classifying risk. Ask for the evidence for each criterion first, quoted from the source, then the verdict. A verdict that rests on a quote you can read is easier to trust, and easier to overturn when the quote does not say what the model thought.
Use it when you have to defend the answer to someone else, who will ask how you got there. And use it with faster models, or with thinking turned off, where Anthropic still suggests asking the model to work through a problem before it answers.
When it does not
For a lookup, a rewrite or a short classification, working adds length and nothing else. OpenAI says that with its reasoning models, prompting them to “think step by step” or “explain your reasoning” is unnecessary, and that some techniques of this kind can sometimes hinder performance.
It costs money and time. Visible working is output tokens, the more expensive side of a call, and on a model that already thinks internally you may be paying for the reasoning twice.
Written-out steps are not proof. A model can show tidy steps with an arithmetic slip in the middle. Check the figures, or have a tool do the arithmetic: Google recommends Gemini’s code execution tool whenever a task needs arithmetic, counting or calculation.
And do not ask for the model’s private reasoning. Anthropic’s documentation says that on its newest models, a request to fill in a thinking or scratchpad section, add a reasoning field to JSON, or give the reasoning in full can be declined. Ask for a short explanation or the evidence behind the result instead.
Before and after
Two requests where the working is worth reading. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.
Is the annual plan cheaper than paying monthly for our team?
Our team of 14 uses a design tool. Monthly billing is $18 per user per month. Annual billing is $180 per user per year, paid up front. We expect to grow to 20 users in month 7. Work out which option costs less over the next 12 months. This goes in my budget request, and the finance manager will read it and check every figure. Show the working before the answer, as a numbered list: each step of the calculation on its own line, with the figures used. Then give the answer in one sentence. Do not round until the final total. Assume annual seats added mid-year are charged the full annual price, and say so; if that assumption changes the answer, show both results.
The first prompt gets a general answer about annual plans usually being cheaper, with no numbers for your team. The second gets the two totals built step by step, so the finance manager can check the six extra seats from month 7 line by line, and an answer that says which assumption it depends on.
Which of these candidates should we interview? """ [12 applications] """
Applications for our payroll specialist role: """ [12 applications] """ Review the applications above against our three must-haves: two years of payroll experience, experience with multi-state payroll, and availability to start by June 1. The hiring manager will use your notes to pick five people to interview. For each applicant, first quote the lines from the application that bear on each must-have, then mark each must-have met, not met or unclear, and only then give a yes or no for interview. If an application says nothing about a must-have, mark it unclear rather than guessing. Present the results as a table, one row per applicant.
A bare request returns a shortlist with reasons that sound right but may rest on a skim. With evidence first, every yes points to the lines that earned it, and an “unclear” tells the hiring manager what to ask on the phone instead of hiding it inside a no.
The line to add
For a calculation, ask for the steps you will check:
Show the working before the answer: each step of the {{calculation}} on its own line, with the figures used. Then give the answer in one sentence, and name any assumption it depends on.
For a judgment, ask for the evidence before the verdict:
For each {{item}}, first quote the lines that bear on each criterion, then mark each criterion met, not met or unclear, and only then give your verdict. If the source says nothing about a criterion, mark it unclear.
Leave out “think step by step”, “write your reasoning in thinking tags” and “explain your full reasoning”. They add length on current models, and the last two can be declined.
In ChatGPT, Claude, Gemini and Copilot
OpenAI’s guidance for its reasoning models says to avoid chain-of-thought prompts, because the models reason internally and asking them to think step by step or explain their reasoning is unnecessary (Reasoning best practices).
Anthropic’s guidance says thinking is always on for Claude Opus 5.5 and the Fable models, and treats manual step-by-step prompting as a fallback for when thinking is off (Prompting best practices). Its refusals page lists the requests that can be declined, such as a thinking section or a reasoning field, and says you can still ask Claude for a short explanation or the evidence behind a result (Keep reasoning in thinking blocks).
Google’s guidance says it is generally not necessary to have Gemini outline or detail its reasoning steps in the response, and that for hard problems a simple “Think very hard before answering” can help, at the cost of extra thinking tokens (Prompt design strategies). For arithmetic, it recommends turning on code execution (grounding and code execution).
In Copilot, as in the others, ask for the working when the result goes into a budget or a decision someone else will check, and check the figures before you send it.
Key takeaways
- Ask for the working when you will check it: figures for calculations, quotes for judgments.
- Put the working before the answer, so the conclusion follows from it.
- Many current models reason internally; “think step by step” adds little there.
- Ask for the evidence, not the model’s private reasoning, which some models decline to show.
- Check the arithmetic yourself, or let a tool do it.
Questions
Should I tell ChatGPT to think step by step?
With the reasoning models now behind most assistants, it is unnecessary: OpenAI says these models reason internally. It can still help on a fast model without built-in reasoning. Either way, if you need to check the result, ask for the specific working, such as each step of a calculation with its figures, as part of the answer.
Why does AI get math wrong?
A language model writes text one piece at a time, and a long calculation gives it many chances to slip, even when every step reads well. Ask for each step with the figures used so you can check it, put the numbers in a spreadsheet, or use a tool that runs code. Google, for example, recommends Gemini’s code execution tool for arithmetic.
What is chain-of-thought prompting?
It means asking a model to write out intermediate steps before its answer. It helped older models on multistep problems. Current reasoning models do this internally, so the request matters less, and some Claude models decline requests to reproduce their internal reasoning. Asking for the working you need to check, such as figures or quotes, still works.