One participant in a 2023 University of California, Berkeley study wanted a cooking-coach chatbot to stop making an odd remark about the flavor you get from using your hands, so they added a line telling it not to. Nothing changed. What finally worked was deleting the “don’t” line and keeping only an instruction that described what the chatbot should do.¹

Learning how to write a good prompt is mostly learning the difference between briefing a person and briefing a language model. The model has nothing but the words you give it, so a good prompt names the goal, the reader, the material and the shape of the answer, in plain sentences, and a weak first answer is a cue to rephrase rather than proof the tool can’t help. In five steps:

1. Start with the goal and who the answer is for
2. Give it the background and material it can’t guess
3. Describe the answer you want back
4. Say what to do, not only what to avoid
5. Rephrase and retry before you give up

Here is the difference in a single request.

The second version names the goal, the material, the shape of the answer and a limit, and each of those is a guess the model no longer has to make.

Prompting skill still matters as models get better

Better models do not make the prompt irrelevant. In a preregistered online experiment with US adults, reported in a preprintpreprint: A research paper posted publicly before it has been peer reviewed or published in a journal; in economics and finance the same stage is usually called a working paper. Its findings may change, or fail to hold up, by the time it is published.Full entry in the glossary last revised in 2026, people given a newer image generator did better partly because they learned to write longer, more descriptive prompts for it, and those same prompts gave no benefit on the older model.²

The study

Limited evidence

Prompt adaptation: an image-copying experiment (Jahani, Holtz, Suri and colleagues, preprint 2026)

Participants were randomly given DALL-E 2 or the newer DALL-E 3 and asked to recreate a target image as closely as possible in at least 10 attempts, with a bonus for the best matches. DALL-E 3 users wrote prompts about 24 percent longer, and the extra words were as descriptive as the rest. When the researchers replayed the DALL-E 2 users’ prompts on DALL-E 3, those prompts captured only about half of the improvement; the rest came from how DALL-E 3 users had rewritten their prompts.²

Part of what a better tool can do shows up only when you ask for more. The caveats are real: the task was copying pictures, each prompt was sent without memory of the last, and in a second experiment by the same team, on open-ended logo design, changes in how people prompted accounted for much less of the gain. Microsoft funded the research, and two authors work there.²

So spell out requirements when you know exactly what you want, such as a summary that must cover three named risks, and let the model contribute more when you are exploring, such as brainstorming names. When your tool changes, revisit the prompts you rely on rather than assuming they carry over. If you are new to chatbots altogether, the beginner’s guide to AI assistants and their limits starts with the basics.

Why a good prompt reads like a brief for a capable stranger

People new to prompting tend to write as if the model shared their background. In a 2023 Berkeley study, non-experts wrote prompts the way they would instruct a colleague, expected to be understood, and read one or two failures as proof the model could not do the task.¹

The study was small and qualitative: 10 participants, mostly professionals and graduate students in technical fields, building a recipe-coaching chatbot on GPT-3. Very few tried rephrasing a request the chatbot ignored, and none made much use of examples, even after seeing them work. The authors’ explanation is that participants expected to be understood, when the words of a prompt simply steer what the model produces, which is how naming an unwanted phrase could make it appear.¹

A colleague shares your background and can ask what you mean; the model can do neither. Anthropic’s own guide compares its assistant to a very capable new hire who doesn’t yet know how your team works, and offers a test: if a colleague with no background on the task would be confused by your prompt, the model will be too.³ Ask a new temp to “tidy up the Q3 report” and they would ask which report, for whom and by when. The model will guess.

What the AI makers’ own guides agree on

The prompting guides from Microsoft, Google, OpenAI and Anthropic, all checked on 27 September 2026, agree on the core: say clearly what you want and give the context the model cannot guess. Microsoft’s Copilot guide names four parts, goal, context, expectations and source, and says only a clear goal is required, though you will often need more.⁴ Google’s Gemini for Workspace guide lists persona, task, context and format, in natural, complete sentences without jargon.⁵ OpenAI’s ChatGPT help page asks for clear, specific prompts with enough context, refined after each response.⁶ Anthropic’s guide asks for clear, explicit instructions, context and a stated output format.³

1. The goal, and who the answer is for
2. The background the model cannot guess, and why the task matters
3. Your own material: the document, notes or data, pasted in with names and confidential details removed
4. Your request after the material, with the length, format and tone you want, and an example if you have one
5. If the answer misses, change one thing and ask again

These are guides from the companies that sell the tools, based on their own testing rather than independent trials, and outside research does not always back them. Microsoft’s tips, for example, say polite language improves Copilot’s response, a claim that independent tests of other models’ accuracy, described below, did not bear out.⁷ Treat the shared core as a sound default and the details as each company’s advice for its own product.

In practice the core fits in two or three sentences. A volunteer coordinator might write: “Draft a reminder to our volunteers about Saturday’s food drive, which now starts at 9 a.m. instead of 10. Keep it under 100 words and friendly, and use only the details in my note below.” That is the goal, the context, the expectations and the source in one short request, and it fits all four guides.

How to write a good prompt in five steps

Every step here turns part of that shared core into a habit and adds what outside research found about it. They are meant for any chatbot, though nearly all the evidence behind them comes from English-language tests.

1. Start with the goal and who the answer is for

Lead with what you want done and who will read the result. “Summarize this report” leaves the model guessing; “Summarize this report for my manager, who has five minutes before a budget meeting and needs to know whether we are over budget” gives it a target. The exception is a long document: paste it first and put your request after it, as the next step explains.

2. Give it the background and material it can’t guess

Paste the document, figures or notes the answer should rest on, with names and confidential details taken out, and say why the task matters. Anthropic’s guide advises explaining the reason behind an instruction, because the model can generalize from it, and putting long documents near the top, above your question and instructions.³

Order matters with long material. In 2023 tests of several language models, they answered best when the relevant passage sat near the start or end of a long input and worst when it was buried in the middle; in the worst case, GPT-3.5 Turbo did worse than with no documents at all.⁸ Newer models may cope better, but the habit costs nothing: paste the document, write your request underneath, and if one section matters most, say which.

3. Describe the answer you want back

Say how long, in what format, for which reader and in what tone: “five bullet points, plain language, no jargon.” In a 2025 test by the Wharton School’s Generative AI Labs in the US, removing the line that told two GPT-4o models how to format their answers lowered their accuracy on a hard science quiz, though the authors expect the effect to vary by model and task.⁹

Examples do more than most people expect: Anthropic calls a few good ones among the most reliable ways to steer format, tone and structure.³ Paste a past email you liked or the column headings you need. The AI email prompts you can copy and adapt show the pattern applied to email.

4. Say what to do, not only what to avoid

Describe the behavior you want. In the Berkeley study, researchers saw “don’t do X” instructions work much worse than “do Y” on several occasions, yet participants wrote the “don’t” form far more often.¹ Instead of “Don’t be so formal,” write “Use the relaxed tone of a message to a teammate you know well.”

5. Rephrase and retry before you give up

Small changes to a prompt can change the answer a lot, in either direction. A peer-reviewed 2024 study found that changes which kept the meaning identical, such as dropping a colon or switching a label to capital letters, moved open models’ accuracy on the same task by dozens of points at the extreme, and a format that suited one model was only weakly linked to what suited another. The tests used older models and short tasks with worked examples, so today’s chatbots may swing less.¹⁰

Up to 76 pointsswing in accuracy from meaning-preserving format changes, for one open model (LLaMA-2-13B) on one task, in 2023 testsSource: Sclar et al., ICLR 2024

When an answer misses, change one thing: restate the goal, add the missing context, move your question after the material or ask for a different format.

A weak first answer is a cue to change the prompt, not proof that the tool can’t do the job.

A weak first answer is a cue to change the prompt, not proof that the tool can’t do the job. Compare the two answers, and for anything important, check the facts yourself, since fluent answers can still be invented, as explained in why chatbots make things up.

WiserHours

Ready to send the prompt?

• The goal and the reader are clear from the start
• The material the answer should use is pasted in, with names and confidential details removed
• Long material comes first and my request comes after it
• Length, format and tone are stated, with an example if I have one
• Instructions say what to do rather than only what to avoid
• If the answer misses, I will change one thing and try again

Find this checklist online:
https://wiserhours.com/ai-productivity/how-to-write-a-good-prompt/#checklist-how-to-write-a-good-prompt-steps

Politeness, threats and expert personas barely move accuracy

Politeness, pressure and expert roles did not reliably change how many hard test questions a model answered correctly in a series of 2025 tests by the Wharton School’s Generative AI Labs, although small wording changes could swing individual questions either way. “Please answer” and “I order you to answer” made little overall difference for two GPT-4o models.⁹ The third report, run with computing credits from OpenAI, tested tips of up to a trillion dollars and threats such as kicking a puppy across five models and found no meaningful gain in overall accuracy.¹¹ The fourth, also run with OpenAI credits, found that expert personas such as “you are a physics expert” gave no consistent benefit across six models, and that personas of people with little knowledge often lowered accuracy.¹²

The picture is not unanimous. A 2023 technical report by researchers at Microsoft, the Chinese Academy of Sciences and several universities found that adding emotional lines such as “This is very important to my career” improved results across a range of tasks on the models of the time.¹³ The Wharton tests used harder questions and newer models, and they measured accuracy on quiz questions, not the quality of an email or a plan.

Myth

Telling the AI it is a world-class expert makes its answers more accurate.

Fact

In 2025 Wharton tests on six models, expert personas gave no consistent accuracy gain on hard multiple-choice questions, and some roles made answers worse.

Roles still have a use: the persona report’s authors note that a persona shapes tone and perspective, so “read this as a skeptical finance director would” is a fair way to get a particular angle.¹² Put your effort into the goal, the context and the format rather than into flattery, bribes or pressure. Whether asking a model to reason step by step still helps with newer models is a separate question, with evidence of its own.

Tested on benchmarks, rarely on office work

The strongest evidence here comes from controlled tests of models on benchmark questions, several of them preprints, plus one small study of how people prompt and the vendors’ own guidance. Almost none of it measured everyday office tasks, so treat the steps as well-grounded defaults, not guarantees.

The bottom line

The words you type are all the model has to go on, so put the goal, the reader, the material and the shape of the answer into them, in plain sentences, with your request placed after any long document. When the answer misses, change the prompt before you decide the tool can’t do the job, and skip the magic words, which the best tests so far found make little difference to accuracy.