# "think step by step" doesn't do what most people think on reasoning models anymore. openai's own guidance says it's not a reliable upgrade. here's what to ask for instead — retained record

> Retained reddit record from reddit:PromptEngineering.

Canonical: https://highsignal.app/data/records/b401140025207c34

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:PromptEngineering)  published 2026-10-03 04:58 ingested 2026-10-04 02:38

[original source ↗](https://www.reddit.com/r/PromptEngineering/comments/1wwem56/think_step_by_step_doesnt_do_what_most_people/)

## retained text



```
"Think step by step" was the most useful line in prompting for two years. On current reasoning models it mostly isn't. They already reason internally before answering, and OpenAI's guidance says asking them to think out loud isn't a reliable improvement. Same with "think harder": it doesn't change the reasoning effort setting. That's a model or effort selector, not a phrase.

What still moves the result is asking for an answer you can *check*:

    Work out [problem] carefully. Give me the answer with a
    short explanation of the key steps, the assumptions you
    made, and any calculations I'd need to verify it.

The difference is what you get back. "Think step by step" produces a long performance of reasoning. This produces the three things you actually need to catch a mistake: where it started, what it assumed, and the numbers.

Two more in the same family that hold up:

    For [my request], list the details you're missing and the
    assumptions you'd otherwise make. Identify the one missing
    detail most likely to change your answer, and ask me for it.

    Review [this answer]. Separate what's supported by what I
    gave you, what's assumption, and what's opinion. Don't
    label anything verified unless you checked a source.

The first one is the highest value. Most bad answers aren't reasoning failures, they're the model quietly filling a gap you didn't know you'd left.

If you want deeper reasoning on a hard problem, change the effort or model setting directly rather than asking for it in the prompt.

been building a library of these as ready-to-run workflows and agents, [here](https://www.promptwireai.com/templates) if you want to dig through.
```



document fetched 2026-10-04 00:34 · canonical https://reddit.com/r/PromptEngineering/comments/1wwem56/think_step_by_step_doesnt_do_what_most_people
