Do I need to tell ChatGPT to think step by step?

Published 11 September 2026 by PrecisionPrompts. The library figures on this page are direct counts over the 455 prompts we publish across 15 packs, measured on that date. The vendor quotations were checked against the live documentation the same day.

No, not on a current model, and OpenAI says so in its own documentation: "Since these models perform reasoning internally, prompting them to 'think step by step' or 'explain your reasoning' is unnecessary." Our own library agrees by accident. Across all 455 prompts we publish, written for 15 different professions and sold to people who kept using them, the phrase appears zero times, and so does every other instruction about how the model should think.

That zero is the part worth sitting with. These are not prompts written to prove a point about chain of thought. They were written to produce a lease abstract, a lesson plan, a product listing, a fund comparison. Nobody removed the incantation on principle. It simply never earned a place, because the prompts were built around naming what the answer must contain, and once you have done that there is nothing left for "think step by step" to do.

What follows is the census, the two things working prompts do instead, and the two situations where the old advice is still correct.

Did "let's think step by step" ever actually work?

Yes, and the numbers are worth knowing, because they explain both why the phrase spread and why it stopped mattering.

It comes from Kojima et al., Large Language Models are Zero-Shot Reasoners (arXiv 2205.11916, May 2022). Adding the single line "Let's think step by step" before the answer took text-davinci-002 from 17.7 percent to 78.7 percent on the MultiArith benchmark, and from 10.4 percent to 40.7 percent on GSM8K. No examples, no fine tuning, one sentence. A result that large travels, and it did.

But read what it was fixing. A 2022 model asked for the answer to a word problem would emit an answer immediately, with no intermediate steps, and get it wrong. The phrase made it write the steps first. Current reasoning models write the steps first by default, whether or not you ask, which is what OpenAI's guidance is pointing at. The instruction has not become wrong. Its job has been absorbed.

What does a library of 455 working prompts say about it?

Nothing, which is the finding. We looked for eight separate ways of instructing the model's process and found none of them.

The word "think" does appear, on 32 lines across the library, and not one of those lines is addressed to the model's process either. They split four ways: five are banned phrases the output must avoid ("I think we could work together", "Still thinking about it?"), four describe what the reader should think ("make the reader think 'yes, that's exactly my situation'"), four are user supplied inputs ([WHAT YOU THINK THEY CARE ABOUT]), and the rest are subject matter, like a classroom think-pair-share or a tone reference to Architectural Digest.

The nearest miss is one line in the Business Strategy pack asking for "one thing everyone in the industry believes that you think is wrong, with reasoning". That does address the model, but it asks for its opinion as content, and it still says where the reasoning goes. Which is the whole pattern.

What do the prompts do instead?

Two things, and both produce reasoning as a side effect of asking for something you can check.

They name the parts of the answer. 276 of the 455 prompts enumerate the deliverable as a labelled numbered list, at a median of 5 parts. A prompt that demands eight named comparison dimensions has routed the model through eight pieces of reasoning without using the word. We measured that pattern in detail in how to get ChatGPT to follow the format you asked for.

They make the reasoning itself a line item. 139 of 455 prompts, 31 percent, require justification in some form, and the form is almost always an attachment to a named output part rather than a free floating instruction.

The structural detail is the load bearing one. Of the 199 individual lines carrying a justification requirement, 156 sit inside a numbered or bulleted deliverable item (62 numbered, 94 bulleted) and only 43 stand alone in prose. The requirement is attached to a thing, and the thing has a name, and you can look at the answer and see whether that name came back with a reason under it.

Why does naming the parts work better than naming the process?

Because one of them is checkable and the other is not.

This is the same argument, arriving from a different direction, as the one we made about personas. "Act as a financial analyst" and "think step by step" are both attempts to reach the model's process through a phrase, and both were measured against something specific instead. That test is written up in does telling ChatGPT to act as an expert actually work, including the one result that went against our prediction.

What a hard reasoning task looks like with no reasoning instruction

Here is a prompt from our Personal Finance pack, reproduced complete. Choosing between five index funds across eight dimensions is a genuinely multi step judgement. The prompt contains no instruction about thinking at all.

I'm choosing between these index funds/ETFs for my [ACCOUNT TYPE — 401k/IRA/taxable]:

[LIST 2-5 FUNDS — e.g., VTI vs. FSKAX vs. SWTSX, or VOO vs. SPY vs. IVV]

Compare them across:
1. Expense ratio and the dollar cost difference over [TIME HORIZON] on a
   [INVESTMENT AMOUNT] investment
2. Tracking error vs. the benchmark index
3. Tax efficiency (capital gains distributions history, qualified dividends %)
4. Liquidity and bid-ask spread (for ETFs)
5. Minimum investment requirements
6. Securities lending revenue (does this offset some fees?)
7. Dividend yield and distribution frequency
8. Which is best for my specific account type and investment size

Give me a clear winner with reasoning, not a "they're all fine" answer. If the
difference is negligible, say so and tell me to pick based on [specific
tiebreaker criterion].

Three moves are doing the work, and none of them is an incantation.

1. The eight axes are the reasoning, written down. Skipping tracking error is now a visible omission rather than an invisible shortcut.
2. "A clear winner with reasoning" attaches the justification to a named deliverable. Not "explain your thinking", which could be satisfied by anything. A winner, and the reason it won.
3. The escape route is closed in advance. "Not a 'they're all fine' answer" names the hedge the model would otherwise reach for. That is the prohibition pattern counted in how to tell ChatGPT what not to do, and it sits in the last quarter of the prompt, where 281 of our 412 prohibitions sit.

The library's single use of the literal phrase "explain your reasoning" follows the same rule. In the Educator pack, converting a lecture plan into an inquiry lesson, the line reads: "Mark changes with [CHANGED] so I can see exactly what was modified vs. kept. Explain your reasoning for each change in a 'Teacher Notes' section at the end." Even when it does ask for reasoning in those words, it says which changes, and it says which section, and it says where that section goes.

Which kinds of task actually need a justification requirement?

The ones where the reason is part of what you are buying. The rate across our packs runs from 77 percent down to nothing at all, which tells you it is not a house habit.

Business Strategy tops it at 77 percent because its deliverable is an argument. A market sizing without the assumptions behind each number is not a shorter market sizing, it is a number you cannot use. AI Image Prompts sits at exactly zero, and that is the row that proves the rule rather than breaking it: a picture has no reasoning to show, because you check it by looking at it. The same pack was the outlier in our prohibitions census for the same underlying reason.

So the test is not "is this task hard". It is: would I read the reasoning if it were there? If yes, ask for it as a named field. If no, asking for it produces a paragraph of throat clearing above the thing you actually wanted.

What do the model vendors say now?

Both of the two largest have published guidance that points the same way, and both were checked live on 11 September 2026.

OpenAI, in its reasoning best practices guide, lists "Avoid chain-of-thought prompts" as a rule: "Since these models perform reasoning internally, prompting them to 'think step by step' or 'explain your reasoning' is unnecessary." The same page tells you what to do instead, and it is the output side: "Be very specific about your end goal: In your instructions, try to give very specific parameters for a successful response."

Anthropic, in its prompting best practices for current Claude models, goes slightly further on the prescriptive version: "Prefer general instructions over prescriptive steps. A prompt like 'think thoroughly' often produces better reasoning than a hand-written step-by-step plan. Claude's reasoning frequently exceeds what a human would prescribe."

Note the difference between the two, because it is the honest nuance. OpenAI calls the instruction unnecessary. Anthropic warns that a hand written plan can actively constrain a model that would have found a better route. Neither says it will break your answer. The realistic cost of typing it out of habit is tokens, latency, and a false sense that you have done something.

When is "think step by step" still the right advice?

Two cases, and they are narrower than the habit suggests.

When the model has no reasoning mode, or it is switched off. Anthropic documents manual chain of thought as exactly this fallback, paired with structured tags so the reasoning and the answer come out separable. Smaller and older models, local models, and cost tuned modes are all real situations where the intermediate steps will not appear unless you ask.

When you want an auditable trace. If you intend to read the working rather than the result, asking for it visibly is correct. But notice that this case immediately converts into the output pattern anyway, because a trace you plan to read is a deliverable, and you will want to say what it should contain and where it should sit.

Neither case is what is happening when the phrase gets pasted at the end of a prompt in a chat window because it is what people do.

One thing our own library does not do, and probably should

Reporting the zero cuts both ways, so here is the other side. Anthropic's guidance also recommends asking the model to check its own answer: "Append something like 'Before you finish, verify your answer against [test criteria].'" That is not a process instruction, it is a second deliverable with a criterion attached, which fits everything argued above.

Across all 455 prompts we publish, zero ask the model to check or verify its own output. Three tell it not to guess, one tells it to flag a gap, and that is the whole of it. On tasks with a checkable answer, particularly the arithmetic in the Personal Finance pack, that looks like a genuine omission in our library rather than a deliberate choice, and we would rather say so here than quietly count it as another point in our favour.

A prompt that converts a thinking instruction into a checkable one

Paste in a prompt that currently ends with some version of "think step by step". This rewrites the request as named parts, which is the form that survives a model upgrade.

Here is a prompt I use, and it currently tells you how to think rather than
what to produce:
[PASTE YOUR PROMPT]

Rewrite it so that every instruction about your process becomes a requirement
about the output. Produce:

1. **Process instructions found**: quote every phrase in my prompt that tells
   you how to think, reason, or approach the task rather than what to deliver.
2. **The hidden steps**: the reasoning steps a good answer to this actually
   requires, as a numbered list. Name them the way they would appear as
   headings in the answer.
3. **The rewritten prompt**: my prompt with the process instructions removed
   and the steps from item 2 written in as named, numbered parts of the
   deliverable. Keep my inputs exactly as they are.
4. **Which parts need a reason attached**: of the parts in item 3, say which
   ones are worth reading a justification for, and write the requirement as a
   field on that part (for example "with a one sentence rationale"). Say which
   ones are not worth it and why.
5. **The check**: one sentence I can append that names a criterion you should
   verify the answer against before finishing.

Do not add new requirements I did not ask for. If removing a process
instruction loses something real, say what and keep it.

Item 2 is the one to read first. The list it produces is usually the outline you would have written yourself if you had spent five more minutes on the prompt, which is the actual reason the technique works.

How this was measured, and what it does not show

Library figures are direct counts over the 455 prompts published in our 15 packs, taken on 11 September 2026 by parsing every fenced prompt body out of the pack files. Matching for Table 1 and Table 2 is case insensitive regular expression search over the full prompt text. The Table 1 zeros are the strongest kind of count available here, because a false negative would require a prompt to instruct the model's process without using any of eight common phrasings, and we read all 32 lines containing the word "think" by hand to check.

The limit is the same one that applies to every census we publish, and it matters more than usual on this topic. This is a count of prompts we believe work. It is not a controlled test of whether adding "think step by step" changes an answer. We did not run that test and nothing here should be read as though we did. What the library can support is narrower and still useful: 455 prompts built for paying users across 15 professions never needed the phrase, which is evidence that it is not a requirement, not evidence that it is harmful. The one controlled comparison we have run on a prompt phrasing, with the markers fixed before either output existed, is the persona test in does telling ChatGPT to act as an expert actually work.

The Kojima figures are quoted from the paper's abstract (arXiv 2205.11916). The OpenAI and Anthropic quotations are from their published prompting guidance and were read on 11 September 2026. Vendor documentation changes, so if you are reading this much later, check the source.

Questions people ask about chain of thought prompting