The Art of Asking: Clear Thinking Before Linguistic Tricks

In 2022 researchers added a single phrase to an arithmetic problem: "Let's think step by step." The accuracy of the model they tested rose from 17.7% to 78.7% on the MultiArith test, and from 10.4% to 40.7% on the GSM8K word-problem set, as Takeshi Kojima and colleagues reported. Same model, same problems; only the wording changed.

Why does wording change the answer?

A language model generates text one word after another, depending on what came before (see the article "How Do Large Language Models (LLMs) Like ChatGPT Work?"), so everything we write in the request steers what follows. A vague request calls up a generic answer that resembles the familiar; precise instructions, attached material and an example bring the model closer to what we mean. The literature calls this "prompt engineering". We call it here "the art of asking", because its essence is asking well, not issuing commands.

What has lost its effect?

"Let's think step by step" is a simplified form of a technique called "chain of thought": the model writes out its reasoning steps before announcing the answer. But models have changed. In a 2025 research report from a team at the Wharton School, an explicit "think step by step" request gave a small average improvement in non-reasoning models and could make answers less consistent; its gain in reasoning models was marginal, if present at all, at a higher cost in time and tokens. Yesterday's advice is not today's law.

Do roles have magic?

It became common to open a request with "You are an expert in…". Yet a systematic study tested 162 roles on 2,410 factual questions across four families of models. It did not find that adding a role raised accuracy compared with having none, and concluded that the effect of each role can be largely random. The study measures accuracy on knowledge questions; adjusting tone and audience is a different matter, which it did not measure.

Anthropic's guidance starts from an earlier step: a clear success criterion, a way to test the prompt empirically, and a first draft to improve. It cautions that a prompt can only address what prompts control, and that cost and latency can sometimes be improved by choosing a different model. Among the techniques it lists are clarity and examples, thinking, and prompt chaining.

Habits that survive changing models

This is a practical summary consistent with the above, not a recipe proven in itself:

• State the purpose and the audience: who is it for, and what decision will the reader base on the answer?
• Attach the material: the document, the figures, the constraints. Do not leave the model to guess.
• Specify the format: length, number of points, language, tone.
• Add one or two examples when what you want is hard to describe.
• Break a large task into steps and check each one.
• Say what you do not want, then adjust the request and measure the difference against a criterion you set before the trial.

And in Arabic

A benchmark published in 2026 for tool calling in Arabic reports that accuracy drops by 5% to 10% on average when the user interacts in Arabic, and notes that many models are pretrained mainly on English. That concerns one specific task and should not be generalised to every use. But it makes it plausible that clarity, examples and checking the answer help an Arabic speaker more than most (see the article "Arabic Artificial Intelligence").

A skill or a profession?

In April 2025 Korn Ferry reported, citing a Microsoft survey of 31,000 workers in 31 countries, that prompt engineering ranked second from last among the new roles companies were considering over the next 12 to 18 months. It attributed this to newer models answering most questions well, and to companies now expecting every employee to use AI. It is more likely a general competence, like clear writing, than a job in its own right.

Its root does not change: whoever has not decided what they want will not be saved by phrasing. And there is a point of honesty. A model answers fluently even when it is wrong, so read its answer as a critic would, and do not use your skill to produce what misleads people. Taking the means in wording and verification does not dispense with reliance on God, nor does reliance on God dispense with them.