How to Write Better Prompts: A Practical Guide
Most bad AI answers are bad briefs. The model isn't being dumb — it's answering the question you actually asked, which wasn't the question you meant. Good prompting isn't a technical discipline or a bag of magic words. It's just clear communication: telling a very capable, very literal assistant exactly what you want, with the context it needs and an example of what "good" looks like. Here's what reliably works, and what's superstition.
Say exactly what you want — format, length, audience, tone. Give the background the model can't see. Show one example of the output you're after. Then treat the first answer as a draft and iterate with follow-ups. That's the whole method. Skip the magic words; specificity beats incantation every time.
Say what you actually want
Vague in, vague out. "Write something about our product launch" gives you generic filler because you asked for generic filler. The fix is embarrassingly simple: specify the things you'd specify to a human colleague.
Format: email, bullet list, table, three paragraphs? Length: a sentence, 200 words, one page? Audience: experts, beginners, your boss? Tone: formal, casual, blunt? Each of these is one clause in your prompt, and each one visibly improves the output. "Write a 150-word casual email to customers announcing our launch, focusing on what's new for them" will beat "write something about our launch" every single time — not because of clever technique, but because you finally said what you meant.
A quick before-and-after shows how much this matters. Before: "Help me write a follow-up email." After: "Write a 100-word follow-up email to a client who went quiet after a proposal. Friendly but direct, one clear call to action: a 15-minute call next week. Here's the tone I use:" — followed by a past email. The second version takes thirty seconds longer to write and produces something you can nearly send as-is. That gap, repeated across every prompt you write, is the entire difference between people who find AI useful and people who find it disappointing.
Honestly: about half of all prompting advice reduces to this one habit. Before hunting for advanced techniques, check whether you actually stated the obvious parameters. Most people haven't.
Give the context the model can't see
The model doesn't know your situation, your constraints, or what you've already tried. It will happily answer the generic version of your question — which is rarely the version you need. Two or three sentences of background transform the answer.
Compare: "How should I price my freelance work?" versus "I'm a freelance designer with three years of experience, mostly working with early-stage startups, currently charging $75/hour but booked solid for two months. How should I think about raising rates?" The second prompt gets advice calibrated to a real situation instead of a pricing 101 lecture. The pattern works everywhere: the relevant facts are the ones that would change a human advisor's answer. Include those.
This also means telling the model what doesn't apply. "We're a team of four, so enterprise solutions are out" saves you from wading through recommendations built for companies with compliance departments.
Show an example
One example of the output you want beats three paragraphs describing it. If you have a previous email, report, or message in the style you're after, paste it in and say "write it like this." The model is excellent at pattern-matching tone, structure, and formatting from an example — far better than it is at interpreting abstract descriptions like "professional but approachable."
This is the closest thing to a power technique in this whole guide, and it's just... showing your work. Writers have always done this with style guides and reference pieces. You're doing the same thing, conversationally.
Iterate — the first answer is a draft
The most underused prompting skill is the follow-up. People write one giant prompt, get a mediocre answer, and conclude the model can't do the task. But conversation is the actual interface — the thing these tools are built for.
Treat the first response as a starting point and direct the revision like an editor: "shorter," "more formal," "keep the structure but rewrite the second section," "that missed the point — the real issue is X." Each round converges. In practice, three quick iterations beat one painstaking mega-prompt, because you can react to what's actually on the page instead of trying to predict everything up front.
This also lowers the stakes of the first prompt. It doesn't need to be perfect; it needs to be clear enough to get a useful draft. Perfectionism at the prompt stage is just procrastination with extra steps.
Break big tasks into steps
Don't ask for the whole report in one prompt. Long, complex requests produce long, complex answers where the weak sections hide among the decent ones — and regenerating means redoing everything. Instead, build it in pieces: outline first, then draft section by section, then assemble.
This has a second benefit: you stay in control of the structure. "Write a report on X" hands the model the steering wheel for the argument itself. "Here's my outline — draft section two" keeps you as the author and the model as the draftsperson, which is the arrangement that produces work you're willing to sign.
What to avoid
Magic-word mysticism. The internet is full of incantations — "take a deep breath," "you are a world-renowned expert," elaborate roleplay setups. Honestly: most of this is superstition. Telling the model to "act as an expert copywriter" can nudge tone and vocabulary in a useful direction, but it's a light steer, not a superpower — and no magic phrase compensates for a vague request. If a technique sounds like a spell, it probably works like one.
Mega-prompts that bury the request. Twenty paragraphs of context with the actual question hidden in the middle produce worse answers than a tight brief. Front-load what you want, then support it. If the prompt is longer than the answer you expect, something's off.
Assuming memory. Don't assume the model remembers earlier sessions or knows things you haven't said in this conversation. (Some products have memory features — check yours — but the default assumption should be a fresh, amnesiac assistant every time.) If it matters, restate it.
Pasting things you shouldn't. Passwords, API keys, customer data, unreleased company material — once it's in the prompt, it's left your machine. The convenience of pasting a sensitive document for summarization is real; so is the data trail. Redact first, or don't paste at all.
A note on framing: this is practical guidance, not lab research. These techniques reflect widely reported best practice and how these systems are documented to behave — not controlled experiments. The throughline is unglamorous on purpose: clear briefing beats clever trickery, and the people getting the best results are usually just the most specific.
Frequently asked questions
Do I need to learn "prompt engineering"?
No — and you should be suspicious of anyone selling it as a discipline you need to master. The skill underneath is clear writing: stating what you want, providing context, giving examples. If you can brief a colleague well, you can prompt well. Everything else is marginalia.
Does saying please and thank you help?
Honestly: there's no solid evidence it meaningfully changes output quality. Politeness doesn't hurt — and it keeps your requests clearer, since polite phrasing tends to be more explicit — but the model isn't grading your manners. Clarity is what counts; courtesy is just good habit.
Should I start prompts with "Act as a…"?
It can help as a tone-setter — "act as a skeptical editor" does steer the response's stance — but treat it as a nudge, not a transformation. It won't make the model an actual expert, and it can't fix a vague request. Useful seasoning, not the meal.
Why do I get different answers to the same prompt?
That's normal behavior, not a bug. These models generate responses probabilistically — the same prompt can produce different wordings, emphases, and occasionally different substance on each run. If you need consistency, be more specific about format and constraints; if an answer matters, verify it rather than assuming the next run would agree.
How long should a prompt be?
As long as it needs to be unambiguous — no longer. A great prompt is often two to four sentences: what you want, the key context, the format. Longer isn't better; every extra paragraph is another chance to bury the actual request. When in doubt, start short and add detail in follow-ups.