How to write better AI prompts: the basics that actually work

The quality of what a language model gives back depends heavily on what you ask it, and the gap between a vague prompt and a good one is usually just a few plain habits — not a trick, a secret phrase, or special formatting. Give the model context, say exactly what output you want, show it an example where you can, and treat the first reply as a draft rather than a final answer. That’s most of what separates a frustrating exchange from a useful one, on any chatbot you use.

Give it context

A model only knows what’s in the conversation. It cannot see your screen, your inbox, your deadline, or the version of a task you’ve already tried and rejected. If you ask “write a project update” with nothing else, you’ll get a generic template, because a generic template is the only thing that fits a request with no context.

Tell it who the output is for (“a project update for a client who isn’t technical”), what you’re actually trying to achieve (“to reassure them we’re on schedule, not to list every task”), and any constraints that matter (“keep it under 150 words, no mention of the delayed supplier”). None of this takes long to type, and each detail removes a guess the model would otherwise have to make on your behalf. Vague input reliably produces vague output; this is the single biggest lever available to you.

Be specific about the output

Beyond context, say what you want the answer to look like. Length (“three sentences”, “a table”, “roughly 200 words”), format (a numbered list, plain prose, an email with a subject line), and tone (formal, casual, blunt) are all things the model will guess at if you don’t specify them, and its guess will not reliably match your intent.

“Summarise this” produces a different summary depending on whether you meant a headline, a paragraph, or five bullet points, and the model has no way to know which you had in mind unless you say so. The more precisely you describe the shape of the answer you want, the less back-and-forth it takes to get there.

Show an example

One example of the output you’re after is often worth more than a paragraph of description. If you want a particular tone, style, or structure, showing the model a short sample — even a single sentence — anchors it far more reliably than describing the tone in the abstract.

This matters especially for anything with a recognisable shape: a headline style, a report format, a way of phrasing a difficult message. Describing “professional but warm” in words leaves room for interpretation; a two-line example of what “professional but warm” sounds like to you leaves almost none. If you’re refining something the model has already written, pointing at the part you liked and saying “more like this” does the same job.

Break big asks into steps

A single sprawling request — “plan my week, draft the email, and check my numbers” — tends to produce a rushed, shallow pass at all three rather than a good one at any of them. Splitting a large task into smaller steps, or asking the model to work through a problem step by step, usually produces a more careful and more accurate result.

This is worth doing even within one message: ask it to first outline, then draft; or first list the assumptions it’s making, then proceed. For anything involving reasoning or numbers, asking the model to show its working makes errors easier to spot, both for you and, in a sense, for the process itself. Researchers at organisations including Stanford’s Human-Centered AI Institute have noted that structuring a task into explicit steps tends to improve reliability compared with asking for a complex output in one leap, though how much it helps varies by task.

Iterate

Treat the first answer as a draft, not a verdict. Chatbots respond well to follow-up instructions: “shorter”, “less formal”, “cut the second paragraph”, “make the opening stronger”. This is normal use, not a failure of the first attempt — few first drafts of anything, human or machine-written, are exactly right.

Telling the model specifically what to change is far more productive than starting over or vaguely asking it to “try again”. “Try again” gives it nothing new to work with; “this is too long and the third point is redundant” gives it something concrete to fix. A short back-and-forth of two or three rounds will usually get you closer to what you actually wanted than any single perfectly worded first attempt.

What NOT to expect

There’s a persistent folklore around “magic words” or so-called jailbreak phrases that supposedly unlock better or freer answers from a chatbot. For the most part, these don’t hold up: providers regularly adjust their systems specifically to close such gaps, and a phrase that seemed to work last month may do nothing today, or trigger the opposite of the intended effect. NIST’s guidance on AI risks notes that techniques for manipulating model outputs are inherently unstable, since they exploit particular quirks of a particular system rather than anything fundamental about how the model works.

What reliably helps is not a phrase but a habit: context, clear specifications, an example, and a willingness to iterate. Whether you’re using ChatGPT or another assistant such as Claude or Gemini, the underlying pattern is the same, because it’s about communication, not incantation. Researchers still debate how much of what looks like a model “reasoning” through a problem reflects genuine step-by-step inference versus a learned pattern that merely resembles it, and that’s a genuinely open question. But you don’t need to resolve it to get better answers. Clear instructions, given plainly, do more for the quality of a response than any special phrasing ever will.