How to Write AI Prompts: the Five-Block Formula

A working prompt has five blocks: role, task, context, format, constraints. When all five are there, the model stops guessing on your behalf and gives you something usable on the first or second try. Drop two of them and the model fills the gaps with an average answer: generic phrasing, random length, and the eternal "intro - body - conclusion" structure.

The practical minimum looks like this: "You are [who]. Write [what] for [whom and why]. Here are the inputs: [facts]. Format: [length, structure]. Do not use [what gets in the way]." Below we break down each block, show how a bad request turns into a good one, and give templates for text, images, video and sound. All the models named in this article run in one interface in Moleculs, so you can test the same prompt on several of them.

What a prompt is and why the result depends on it

A prompt is a text request to a model. The model does not know who you are, what you need the text for, or what a good result looks like for you. It takes your words and completes the most probable answer. The less you say, the wider the field of options and the more average the output.

Hence the main principle: a model answers what you wrote, not what you meant. "Tell me about marketing" fits a schoolkid and a product director equally well, so the model picks something in between - exactly the kind of text people later call filler. The moment you add who the reader is and what they should do after reading, the answer narrows down to something useful.

Second point: the model has no access to your internal data. It does not know your company, your prices, your competitors or your brand tone unless you hand them over. Invented details in the answer almost always mean the model filled an empty slot in your prompt.

The prompt formula: role, task, context, format, constraints

Role

Who the model should be in this answer. "You are a business editor," "you are a physics teacher for 15-year-olds," "you are an HR specialist in tech." The role sets the vocabulary, the depth of explanation and the level of assumptions. One line changes the answer more than you would expect: an editor cuts, a teacher unpacks step by step.

Task

What exactly to do, in one verb. Not "write about our service," but "write three headline options for the home page." It helps to add the goal: why this text exists, what the reader should understand or do.

Context

The inputs the model cannot have: the product, the audience, facts, numbers, examples, business limits. Good and bad samples belong here too. One solid sample in the prompt works better than three paragraphs of "write in a lively tone."

Format

Length in sentences or units, structure (lists, table, headings), where you want paragraphs and where you want bullets. Models are bad at counting characters, so structure is the reliable lever: "five points, two sentences each" is followed more accurately than "about 1,500 characters."

Constraints

What is off limits: bureaucratic phrasing, clichés, invented numbers, exclamation marks. One nuance: constraints work better in positive form. "Don't write filler" is weak; "write only facts, numbers and concrete actions" is strong.

Before and after: turning a vague request into a working one

Before: "Write a post about our CRM."

What you get: a generic text with abstract benefits and a closing "choose the best solution." Technically on topic, unusable in practice.

After:

You are a copywriter writing for owners of small service businesses (salons, auto repair shops, studios). Write a post for a Telegram channel about our CRM. The point of the post: show that leads stop getting lost. Inputs: the CRM collects requests from messengers, social media and the website into one inbox, sends a reminder a day before the appointment, and keeps a client history. Audience: people who currently track bookings in a paper notebook or in phone notes. Format: 900-1200 characters, open with a concrete situation from a salon, then three paragraphs with one idea each, end with a single question to the reader. Simple words, informal address, concrete details only, no promises about percentage growth in sales.

The difference is not the length. It is that the model has nothing left to invent: it knows who reads this, what matters, how long it should be and what the tone is.

If you don't feel like assembling all of that by hand, ask a model to do it: open a chat and write "You are a prompt editor. Rewrite my request: add role, goal, audience, format and constraints, and ask me three clarifying questions if the inputs are thin. My request: ..." You get a structured prompt plus the questions you forgot to answer.

Prompts for text: controlling length, tone and structure

With text, three things usually break: length, tone and structure. Each has its own fix.

Length. Set it in units, not characters: "six paragraphs," "ten points," "three sentences per point." If you truly need a character limit, add "no longer than N characters, shorter is better" - at least the model will not overshoot by much.

Tone. Three levers work: name the author's role, give a sample, and ban specific constructions. For example: "no warm-up sentences, no words like unique or innovative, no rhetorical questions at the start of paragraphs."

Structure. Describe it as a list, or better, as a skeleton: "Headline. Two-sentence lede. Three subheads, each with one paragraph and a three-item list. Closing paragraph with a single takeaway." The model reproduces a skeleton far more accurately than a verbal description.

Formats with a fixed shape are the easiest case, because you can describe the shape once and reuse it. A conference abstract needs the topic, the gap it addresses, the method and the finding; a talk script needs the speaking time in minutes rather than a character count, so the model can pace it for speech; a summary needs a sentence limit and an instruction not to restate every conclusion. If any of those slots is empty, you get the generic version.

For long pieces, split the work into steps: ask for an outline first, fix the outline, then ask for section-by-section writing. That way you correct the structure before the model writes three thousand words in the wrong direction. Presentations work the same way: first the list of slides with one idea per slide, then the content.

Prompts for images: subject, scene, composition, light

An image prompt is built differently. There is no dialogue and no follow-up questions, just one description read as a whole. Word order matters: whatever comes first carries more weight.

A working sequence: subject → scene → composition → light → style → technical details.

• Subject. Who or what is in the frame, in what state and action. "An elderly craftsman at a workbench repairing a watch," not "a craftsman."
• Scene. Where it happens, what surrounds it, time of day, weather.
• Composition. Angle and shot size: close-up, top-down, waist-up portrait, wide shot. Aspect ratio goes here too: vertical, square.
• Light. Soft daylight from a window on the left, backlight, neon, golden hour. Light changes the picture more than half the other words combined.
• Style. Photography, watercolour, 3D render, editorial shot, flat illustration. Do not mix three styles in one prompt; you get mush.
• Technical details. Shallow depth of field, muted palette, film grain.

A note on bans. In image models they are unreliable: "no text" sometimes produces letters in the background anyway. Describing what should be there beats listing what should not.

If you are editing an existing image rather than generating one, invert the logic: describe only the change and leave the rest alone. "Replace the background with a plain grey studio wall, keep the pose and lighting" works; a full re-description of the subject usually shifts things you wanted to keep. The difference between models shows up most on complex scenes with several objects, so the same prompt is worth running through Nano Banana, Midjourney, Flux 1.1 PRO, Ideogram v3 and Seedream v4.5 before you settle on one.

Prompts for video and audio: what to add to the frame description

Video is an image plus time. Three things get added to the frame description that static images do not have.

Motion in the frame. What the subject does during those seconds: walks, turns, pours coffee. One action per clip, not three.

Camera motion. Push in, pull out, pan left, orbit, locked-off tripod. If you don't specify it, the model picks for you, and it is usually a slow drift.

How the scene develops. What changes from start to finish: the light goes out, the door opens, the crowd disperses. A short clip holds one change.

For models like Veo 3.1, Kling 3.0 or Seedance 2.0 it helps to write the prompt as a description of a single continuous take - no cuts, no location changes. If you need a second shot, generate a second clip.

Image-to-video is a separate case, available in Kling 3.0 (image-to-video), Hailuo 02 Standard (image-to-video), Wan 2.6 (image-to-video) and others. The picture is already fixed, so the prompt should describe motion only: what comes alive and how the camera behaves. Re-describing appearance and surroundings just pulls the result away from your reference.

Music works on another structure. In Suno V5 you set genre, mood, instruments, tempo, vocals and track structure (verse - chorus - bridge). For sound effects in ElevenLabs Sound Effects, a short precise description of the event and the environment is enough: "creaking wooden door in an empty hallway, echo." For text-to-speech, the text itself is effectively the prompt, and the delivery is set by voice settings and pause markup rather than by wording.

How prompts differ across text, images, video and audio

The key difference: with a text model you hold a conversation and fix things as you go, while an image, video or music generator reads the prompt once and in full. So with text you can start short and refine, and everywhere else it pays to write the complete description up front.

Eight mistakes that make the model answer the wrong thing

1. No audience. The model doesn't know who the text is for, so it writes for everyone. Add one line about the reader and their level.
2. Only negations. "Don't be long, don't use clichés, don't pad" - models hold instructions better than prohibitions. Rephrase as what to do.
3. Contradictory requirements. "Keep it short but cover every aspect in detail" - the model picks one, usually the wrong one. Scan your prompt for mutually exclusive conditions.
4. Everything in one paragraph. Ten requirements buried in flowing text get lost. Move them into a numbered list.
5. No facts. Without numbers, names and specifics the model either writes in generalities or invents. Supply the inputs yourself.
6. Format implied but not stated. You expect a table and get paragraphs. Say it outright: "answer as a table with columns X, Y, Z."
7. Several tasks in one request. "Analyse this, write the text and suggest headlines" - quality drops on all three. Split it into steps.
8. Abstract judgements instead of samples. "Write it in an interesting way" means nothing. Attach a text you like and ask the model to match its manner.

Iterating: how to fix a prompt when the answer misses

The first answer is rarely perfect, and that's fine. The mistake is rewriting the whole prompt from scratch. Change one parameter at a time and watch what moves.

Step 1. Name the problem in words. Not "I don't like it," but "too generic," "wrong tone," "no structure," "invented facts." Naming the problem almost always points at the empty block in your prompt.

Step 2. Patch, don't rebuild. "Too generic" is cured by context and facts. "Wrong tone" by role and a sample. "No structure" by format. "Invented facts" by one line: "use only the data in my message; if something is missing, tell me what."

Step 3. Give feedback on a specific chunk. With text models this works well: "The second paragraph is too generic, rewrite it with a concrete example. Leave the rest as is." More precise than "rewrite it better."

Step 4. Let the model diagnose the prompt. "What in my request prevents a precise answer? Ask me three clarifying questions." It usually turns out you skipped something that was obvious only to you.

Step 5. Save the prompt that worked. As soon as you land a good version, keep it in your notes as a template with square brackets for substitution. That saves more time than any prompt collection online.

For images the loop is different: change one block of the description at a time - light first, then angle, then style. Change everything at once and you won't know what improved the shot.

Ready-made prompt templates

Learning something new:

You are a teacher of [subject] for first-year students. Explain the topic "[topic]" so that someone with no background gets the point. Format: a two-sentence definition, three key principles with real-life examples, one common misconception and why it's wrong, five self-check questions. Skip formulas where they can be avoided.

Work, business email:

You are a department head. Write an email to [recipient] about [situation]. Goal: [what the recipient should do]. Inputs: [facts, dates, amounts]. Format: subject line, three paragraphs, one concrete question or request at the end. Neutral business tone, no boilerplate, no apologies for bothering them.

Work, reading data:

You are an analyst. Here is the data: [paste]. Find the three main patterns, explain each in two sentences and state the action it implies. Format: a table with columns "observation," "what it means," "what to do." Do not invent numbers that are not in the data.

Creative, children's story:

You are a children's author. Write a story for a child aged [age] about [characters]. Plot: [setup]. Format: 2500-3000 characters, short sentences, dialogue, a gentle ending without an explicit moral.

With stories for kids the decisive inputs are not the plot but the listener's age and the length: they drive vocabulary and pacing. Poetry adds a technical requirement - state the meter, the rhyme scheme and the number of stanzas, or the model drifts into free verse and loses the rhythm by the third stanza.

Image, universal skeleton:

[Subject and action], [place and surroundings], [angle and shot size], [light], [style], [palette], [technical details].

Video, universal skeleton:

[Frame as in a still], [what the subject does], [camera motion], [what changes by the end of the clip], [shooting style].

One practical habit: run the same prompt through two or three models and compare who understood the task. The gap between Claude Opus 5, Gemini 3.1 Pro and ChatGPT-5.6 Sol on an identical request is often wider than the gap between two versions of your prompt.

FAQ

How long should a prompt be?

Three to seven sentences is the sweet spot: role, task, context, format, constraints. A single line usually yields a generic answer, while a full page means part of your requirements gets dropped. If you have many conditions, break them into a numbered list.

Why does the model ignore some of my instructions?

Usually one of three reasons: the requirements contradict each other, there are too many of them crammed into one paragraph, or they are phrased as prohibitions. Remove the contradictions, move the conditions into a numbered list, and replace "don't write filler" with "write only facts and numbers."

Does the same prompt work in every model?

The core transfers, the details don't. A text prompt for Gemini 3.1 Pro will work in DeepSeek V4 Pro with almost no edits, but a frame description for Veo 3.1 is useless in Suno V5, which needs genre, mood and track structure.

Can I ask a model to write the prompt for me?

Yes, and it's a solid habit: describe the task in your own words and ask the model to assemble a structured prompt from it and ask clarifying questions. What comes back needs a light edit, and then it becomes your template.

How do I know what a generation will cost?

Inside a subscription you spend credits, and each model has its own cost per run, shown next to the model in the catalogue. Video and some image models cost more than text ones, so the number is worth checking before a long batch. Quotas per plan are listed on the pricing page.