You are reading Chapter 15 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.
To make AI show its reasoning, stop asking it to think step by step and ask instead for the parts you can check: the assumptions it made about your situation, the options it weighed before it picked one, and the single item the recommendation rests on. Those are outputs you can read and argue with. "Think step by step" is a method instruction, and the assistants already do that part on their own.
This guide covers why the old phrase stopped paying for itself, the 3 asks that replaced it, and how to read the reasoning that comes back without mistaking length for rigour.
We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.
Why a confident answer you cannot check is a problem
You ask something that genuinely matters. A pricing call, a hiring choice, whether a contract clause is a problem. Back comes a confident recommendation, well written, with nothing in it you can test.
You cannot see what it assumed about your business, what else it weighed, or which single fact the whole thing rests on. You either take it on trust or you throw it away, and neither of those is using the tool properly.
What changes that is not a harder setting. It is requiring the answer to arrive in a shape that exposes its own logic: what it assumed, what it considered and rejected, and which step is load-bearing. That is an instruction, and it works identically on a free ChatGPT account, a Claude window, a Gemini tab and a Copilot pane.
This chapter assumes you can already write a clear instruction. If yours still come back thin, Chapter 14's worked examples fix that first. Bring one prompt you have already written, ideally one where the answer came back confident and you had no way to check it.
The setting is theirs. The instruction is yours
A setting buys depth: how long the assistant works before it speaks. The Think button, the effort level, the model picker. Those belong to the vendor, they differ by assistant, and they keep changing.
An instruction buys direction and shape: what the answer must consider, and what it must show you. It works on every assistant, on any setting, it survives a model change next month, and it costs you one extra sentence.
The distinction decides what to reach for. An assistant given twice as long to think about a question you asked badly produces a more thorough answer to the wrong question, and it sounds better while doing it.
Which switch to flip, and when to leave it alone, is Chapter 6's job. This chapter is about the words, which work the same whether the switch is on, off, or absent from your assistant entirely.
One boundary before you start. This is about the quality of thinking inside a single answer. When a job is genuinely too big for one answer, no instruction rescues it and the fix is sequencing.
The advice you have heard 3 times has quietly expired
Adding "think step by step" used to make answers measurably better. On a current assistant it mostly costs you time, and the vendors who publish guidance on this now say so themselves.
OpenAI's guidance for its reasoning models tells you to avoid chain-of-thought prompts, because the model performs that reasoning internally and asking for it is unnecessary. Google's Gemini guidance says it is generally not necessary to have the model outline, plan, or detail its reasoning steps in the response. Anthropic goes furthest and advises general instructions over prescriptive steps, because the model's own reasoning frequently exceeds what a person would have written out for it. Microsoft is the exception: Copilot's advice stays at be clear, give context, name your format.
The measurements agree. A Wharton research group ran the test on 198 PhD-level science questions, 25 trials per condition. Older models with no reasoning mode of their own gained a lot from the phrase, in one case over 13 points. Reasoning-capable models gained roughly 3 points, one of them got slightly worse, and every one of them took between 20% and 80% longer to answer. Their conclusion was that chain-of-thought prompting is not universally optimal.
Read that with the caveat it deserves. Multiple-choice science questions are not your board memo, and a benchmark is a proxy for your work rather than your work. The direction is consistent across independent measurement and 3 separate vendors, which is about as much agreement as this field ever produces.
The reason it broke tells you what to do instead. The phrase existed for a model that would otherwise commit to the first plausible answer and then defend it. Current assistants do that deliberation before they show you a word of the reply. Telling one to think step by step is telling it to do a thing it already did, and on some setups it repeats the reasoning out loud in the answer, where it costs you reading time and buys you nothing.
3 asks that still change the answer
The instruction that still works is not a phrase about thinking. It is naming a piece of work you want done, and never the method for doing it. Every clause below names an output you can read and argue with. Not one of them tells the assistant how to reason.
1. List what you are assuming about my situation
"Before you answer, list the assumptions you are making about my situation, and mark the ones you are least sure of." Your question was underspecified, because every real question is. It answered anyway, which means it filled the gaps silently and you never saw the filling.
This clause makes it print the list, and roughly half the time you read it and realise the problem was your question rather than its answer. For a senior reader it is the highest-yield sentence in the chapter.
2. Give me 3 options, with the trade-off of each
"Give me 3 options with the main trade-off of each, then recommend one and say what the other two would have been better at." The failure this fixes is specific and common: an assistant settles on the first workable answer and spends the rest of the reply building a case for it.
One recommendation is a conclusion. Three options and a pick is a decision you can audit. Requiring the field before the choice stops a reply from being an argument for a conclusion it reached in its opening sentence.
3. Name the one item you intend to check
This is where most people ask for too much. "Show your reasoning" gets you narration, a pleasant account of how a sensible person might have approached it. Name the item the answer rests on instead.
The numbers you used and where each one came from. The criterion you applied. The single assumption this recommendation would collapse without. Ask for what you would need in order to prove it wrong.
The fourth clause, when the answer contains something checkable
Google's guidance is explicit that you should include verification steps in the instruction, and Anthropic's is that appending a line asking it to verify the answer against your criteria catches errors reliably.
One honest caveat. The newest models increasingly do this unasked, and Anthropic now tells developers moving to its latest model to delete inherited verification instructions rather than rewrite them, because piling them on produces a model that checks, re-checks, and then reports at length on having checked. Add the clause when there is arithmetic, a date, a quoted line or a rule with conditions. Leave it off otherwise.
Read the working, do not admire it
What comes back when you ask for reasoning is a written account of an answer. It is not a recording of how that answer was produced, and the difference decides what you can safely do with it.
Anthropic's own researchers tested this directly. They planted a hint in a question, confirmed the model had used it, then checked whether the visible reasoning admitted to using it. Across the reasoning models they tried, it mentioned the hint that had actually changed its answer well under a fifth of the time.
Their concern was safety monitoring rather than your Tuesday, but the narrow lesson transfers cleanly: the trace tells you what the answer claims, not why the answer came out that way.
It is still worth having, for a narrower reason than most people assume. An assumption you disagree with is wrong whether or not it drove the conclusion. A number is right or wrong on its own terms. An option dismissed on a ground you know to be false is a bad dismissal, whatever else was going on. You are checking the claims in front of you, which is a normal editorial act. You are not auditing a mind, and you cannot.
The costs are real and worth naming. Asking for the working makes every answer longer, slower and more confident-sounding, and length reads as rigour when it is nothing of the kind. So the guide gives you a rule you read from the top and stop at the first line that is true. Can you check the answer on sight? Ask for nothing extra. Is it a judgement you would take on trust? Ask for assumptions. Is more than one answer defensible? Ask for options first. Does it contain numbers, dates or quotes? Ask it to verify those.
Leave it off for short factual questions, for drafting, for anything you can eyeball. Spend the extra sentence when the answer is a judgement you would otherwise accept on trust, which is exactly the category where an assistant is most useful to you and least checkable.
What actually changes
Before: you get a confident recommendation, well written, and you either take it on trust or throw it away, because there is nothing in it you can test.
After: the answer arrives with its assumptions listed, the options it weighed on the page, and one named item you can go and check. You are still the one deciding, and now you are deciding on something.
The guide closes with 3 exercises that take about 15 minutes, ending with one prompt rewritten to ask for 3 options and a recommendation instead of an answer.
Two things carry into Chapter 16: that rewritten prompt with the assumptions clause on the front, and one job you already tried to do in a single prompt and watched go thin. Chapter 16 takes on the jobs that are too big for a single answer at all, where the fix is sequencing rather than wording.
For the wider set of prompting moves this one sits inside, read the 26 principles of prompt engineering, and if a prompt keeps coming back wrong, the GPT-5 prompt optimizer rewrites it for you. The rest of the guides are in the AI Central Library.
Frequently Asked Questions
Does "think step by step" still work in 2026?
Mostly not, on a current assistant. OpenAI's reasoning guidance tells you to avoid chain-of-thought prompts because the model reasons internally, Google says it is generally not necessary to have the model outline, plan or detail its reasoning steps in the response, and Anthropic advises general instructions over prescriptive steps. A Wharton test across 198 PhD-level science questions found reasoning-capable models gained roughly 3 points from the phrase while taking 20% to 80% longer to answer.
What should I ask for instead of chain of thought prompting?
Name the output you want rather than the method. Ask it to list the assumptions it is making about your situation, to give 3 options with the main trade-off of each before it recommends one, and to name the single item the recommendation rests on. Add a verification line when the answer contains arithmetic, a date, a quoted line or a rule with conditions.
Is the reasoning an AI shows you accurate?
It is an account of the answer, not a record of how the answer was produced. Anthropic's researchers planted a hint in a question, confirmed the model had used it, and found the visible reasoning mentioned it well under a fifth of the time. Check the claims in front of you rather than treating the trace as a transcript.
What is the difference between a thinking mode and a reasoning prompt?
A setting buys depth: how long the assistant works before it speaks. An instruction buys direction and shape: what the answer must consider and what it must show you. Settings differ by assistant and keep changing, while the instruction works everywhere and survives a model change. Chapter 6 covers the settings.
When should I skip asking for reasoning?
When you can check the answer on sight. Short factual questions, drafting, anything you can eyeball. Every extra clause costs you length, latency and a longer thing to read, so spend it on the judgements you would otherwise accept on trust.

