How to write AI prompts that actually work

We asked an AI to write a product description for a candle. It invented a candle. A name (“Serenity Bloom”), a scent (lavender and ylang-ylang), a burn time (60 hours), even the packaging, none of it real, all of it wrapped in “Indulge your senses.” Thirty seconds later we sent the same model a five-sentence brief instead: the actual product, the actual buyer, a sample of the shop’s voice, the banned cliches, the exact format. It returned copy the shop could paste in today. Same model, same price, same half-minute.

The difference was not a magic word. If you type four words and get mush, the fix is not hiding in a $9 PDF of “500 best prompts,” and it is not “act as a world-class copywriter.” A prompt is a brief, and everyone reading this already knows how to brief a competent human. This guide is that one skill, split into five moves, each shown with real before-and-after outputs we generated for this post. The moves work in ChatGPT, Claude, Gemini, and every image and video model you will meet, because they are not tricks about software. They are the mechanics of asking well.

The magic-prompt industry thrives because the field looks occult from outside. A sweeping academic survey known as the Prompt Report catalogs 58 named prompting techniques for text models and 40 more for other modalities. You need about five of them. The vendors agree: OpenAI’s, Anthropic’s, and Google’s own prompting guides, all linked below, converge on the same short list this post teaches.

Move 1: Make the decisions before the model does

A vague prompt is not a smaller version of a good prompt. It is a delegation order. “Write a product description for a candle” contains no candle, no buyer, and no voice, so the model supplies all three from the statistical average of every candle description it has ever seen. That average is “Serenity Bloom.” The model did not fail; it answered the question we actually asked, which was “what does a generic candle description look like?”

Read the two outputs side by side and notice what the brief bought. Every fact in the strong version is true because we supplied it. The voice is the shop’s because we showed it. The cliches are absent because we banned them by name. None of that required technical skill. It required deciding what we wanted before asking, which is the part the magic-prompt PDFs let you skip, and the part that does the work.

The practical test: could a stranger produce what you want from your prompt alone? If a freelancer would need to email you three questions back, the model needed them too. It just doesn’t get to ask.

Move 2: Brief it like a capable new colleague

Anthropic’s official guidance uses a frame worth stealing: treat the model as “a brilliant but new employee who lacks context on your norms and workflows.” Their golden rule follows from it: show your prompt to a colleague with minimal context and ask them to follow it. If they’d be confused, the model will be too.

Context is the answer to three questions you’d volunteer to any human helper without being asked. Who is this for? A resignation letter for your manager of eight years is a different document from one for a boss you met in March. What is it for? A summary that feeds a slide deck wants different compression than one that feeds a lawsuit. What have you already tried? “I wrote a draft and it sounds stiff” plus the draft beats starting cold, every time, because now the model is editing toward a target instead of guessing at one.

This is also where the fear of long prompts dies. Beginners keep prompts short out of politeness, as if the model were busy. It is not busy. Two paragraphs of honest context cost you forty seconds of typing and routinely save five rounds of “no, not like that.”

Move 3: One example beats ten adjectives

Adjectives describe style badly. “Warm, plainspoken, not salesy” narrows the space a little; one pasted sentence of the style you want nails it. Our candle brief included a single line from the shop’s existing listings (“Our porch in October smells like this…”), and that line did more steering than every instruction around it. The model matched its temperature, its sentence length, its refusal to sell hard.

This move, called few-shot prompting in the literature, is the closest thing prompting has to a cheat code, and it is the one beginners use least. Google’s prompting guide is blunt about it: “We recommend to always include few-shot examples in your prompts. Prompts without few-shot examples are likely to be less effective.” OpenAI’s guide recommends two or three varied examples, which the model implicitly picks up the pattern from. The examples do not need to be masterpieces. They need to be yours: the email that sounded like you, the caption that worked, the summary your boss praised.

One caution, learned the hard way: examples steer hard, so a bad one steers hard in the wrong direction. If you paste a sample whose facts differ from your task, say explicitly what to take from it (“match the voice, not the details”), or the model may helpfully import the details too. And when the output clings too tightly to your example, reusing its phrases verbatim, add a second, different example. Two samples define a range; one sample defines a template.

Move 4: Say the format out loud

Models are strangely sensitive to form. A research team led by Melanie Sclar found that trivial formatting changes in a prompt, things as small as separators and casing, swung task accuracy by up to 76 percentage points on an older open model. You cannot control that sensitivity entirely, but you can stop leaving format to chance: state it.

In text, format means length, structure, and register: “one paragraph, max 60 words, then 3 bullet points” is a spec the model can hit, and in our candle test it hit every element exactly. It also means naming what to avoid; our ban list (“no luxury cliches like ‘indulge’ or ‘elevate’”) deleted the worst of the weak output in one clause. In images, format is aspect ratio and medium. In video, it is duration and shot count. In spreadsheets, it is “return a table with these three columns.” The model can only match the shape in your head if the shape leaves your head.

Move 5: Treat the first answer as a draft

The single most expensive habit in prompting is accepting or rejecting the first output. People type once, skim, sigh, and either settle or close the tab. Professionals treat the first result as a draft and the second prompt as the actual work: “Keep the structure and the first bullet. The middle paragraph is too formal, match the sample I gave you. Cut the last line entirely.” You are not starting over; you are directing.

Two habits make iteration cheap. First, change one thing at a time, the same advice Google gives for image generation: establish the core idea, then refine toward your vision rather than rerolling from scratch. Second, say what to keep, not just what to change. An instruction like “same scene, warmer light” preserves the 90% that worked. A fresh prompt gambles it.

There is a limit, and it is worth knowing where it sits. Iteration polishes a draft that is roughly right; it cannot steer one that is fundamentally wrong. If the third revision is still fighting you on the same axis, the problem is upstream: a decision you never made (back to move 1) or context you never gave (move 2). Go fix the original brief and resend it. Restarting with a better prompt feels like defeat and is usually the fastest path to done.

Images and video: same briefing, new vocabulary

Everything above transfers to pictures; only the vocabulary changes. Where a text brief specifies audience and voice, an image brief specifies subject, context, and style, the exact three-part structure Google’s Imagen prompt guide teaches: what the thing is, where it is, and how it’s rendered (photo, watercolor, isometric 3D), plus lighting and lens language if you want photographic control. We ran the experiment to show the spread:

“a coffee shop”

subject + context + style + lighting + lens

Neither image is “wrong.” The left one is even pleasant. But the left one is the model’s decision and the right one is yours, and if you were making a poster, a menu, or a product shot, only one of them is usable on purpose. Piling on adjectives (“beautiful, stunning, 8k”) does far less than adding decisions: whose hands, which light, what lens.

If you don’t have the photography words yet, describe it the way you’d describe a photo to a friend on the phone: what’s in it, what it’s made of, where the light comes from, how close you’re standing, what mood it leaves. That plain-English version outperforms a pile of borrowed jargon, because the model needs your decisions, not your vocabulary. The technical terms are compression, not admission requirements; you pick them up naturally after a dozen generations.

Video adds one more layer: motion. Google’s prompting guide for its Veo model recommends a five-part brief (cinematography, subject, action, context, style) and, crucially, real film language: dolly in, tracking shot, slow pan, close-up, low angle. These terms work because the models trained on footage described by people who used them precisely. Prompt like you’re directing a shot, not describing a vibe, and say what moves: the subject, the camera, or both. We put current video models through this kind of prompting, clips included, in Sora vs Veo vs Kling.

What a great prompt can’t fix

Honesty box. A perfect brief cannot rescue the wrong tool: no phrasing lets a chat model read the sales spreadsheet you never attached, and no phrasing makes a model without web access know yesterday’s news. It cannot make a model certain: a confident answer about a niche legal question deserves verification no matter how well you asked. And it cannot push past a model’s actual ceiling: if four careful iterations haven’t gotten close, the fifth rewrite is usually the wrong move. Switch models, switch modalities, or accept that this one is a human job. Knowing which tool fits which task is its own skill, one we walk through in the plain-English tour of what AI can actually do.

Everything else is learnable in an afternoon, and it compounds, because the five moves are one habit wearing five hats: say what you want, say who it’s for, show what good looks like, name the shape, then direct the revision. Here is the whole guide in a form you can keep:

The next time an output disappoints you, resist the two reflexes: the model is broken, or you lack the secret words. Reread your prompt as if it landed on a stranger’s desk. Nine times out of ten you’ll find the missing decision in ten seconds, add a sentence, and watch the answer snap into focus. That’s the entire discipline. Not magic words. A better brief.