What Is a Prompt and How to Write One: Structure, Examples, Ready-Made Templates

A prompt is the text you give an AI model: a question, a task, an instruction. Almost everything depends on how it is written. The same model gives you filler for a poor prompt and a result you can use right away for a good one.

Below: what a prompt is in plain words, what it is made of, how to write one step by step, which techniques actually work, and ready-made prompts for common tasks.

A prompt in plain words

Imagine briefing a new hire. They are smart, fast and read at lightning speed, but know nothing about your company, your customers or what exactly you want. A prompt is that brief. The more precise it is, the fewer rewrites you need.

The word means a cue or a nudge. In AI work it covers any text sent to a model: a short question in a chat, a long task with examples, or a standing instruction for an AI agent. Whether you call it a prompt, an AI prompt or a request to a model, it is the same thing; only the place differs: a chat like ChatGPT, DeepSeek or Claude, an image generator, or an automation.

What a good prompt is made of

01Rolewho the model should be
02Taskwhat to do
03Contextfor whom and why
04Datawhat to work with
05Formathow to shape the answer
06Limitswhat is off limits

Part	Why it matters	Example
Role	sets the angle and level	You are the editor of a city news channel
Task	says what to do	Retell the news item for Telegram
Context	explains for whom and why	Readers are city residents reading on a phone
Data	provides the material	The news text below
Format	defines the shape of the answer	2-4 short sentences, the first is the main fact
Limits	protects against mistakes	Only facts from the source, no conclusions, no emoji

You do not need all six parts every time. A short question needs only a task and context. But when you dislike the answer, one part is almost always missing, most often context or format.

How to write a prompt: step by step

Define the result. Not "help with a post" but "a 4-sentence Telegram post that brings people to the promotion".
Describe the audience. Readers, platform, tone. A post for lawyers and one for teenagers are written differently.
Provide the material. Facts, figures, source text, links. The model does not know your data, and without it the model will invent it.
Set the format. Length, structure, list or paragraph, table, heading. For automation, a strict format such as JSON helps.
Set limits. What is not allowed: inventing facts, promising deadlines, clichés, more than N words.
Show an example. One or two samples of the result beat any description of style.
Check and refine. Look at the answer, find what is off and add a rule for that exact mistake.

Example: a bad prompt and a good one

Bad: "Write a post about our promotion."

The model knows neither the promotion, nor the audience, nor the platform. It will invent everything, and you will rewrite it.

Good: "You write for a coffee shop's Telegram channel. Readers are locals aged 20-35. Write a post about the promotion: from 1 to 7 November, second coffee free from 8 to 11 am. 3-4 sentences, friendly, no hype, no emoji. End with the address: 1 Main Street."

Another example, for customer work:

Bad: "Reply to the customer."

Good: "You are a manager at an online store. Reply politely and briefly, up to 3 sentences. Use only the information in the base below. If the answer is not there, say you will check and get back within an hour. Do not promise discounts or delivery dates that are not in the base. Question: ... Base: ..."

Prompts for different tasks

Task	What to specify	Typical mistake
Posts and copy	platform, audience, tone, length, style examples	no examples, so the text is generic
Customer replies	knowledge base, "unsure, hand over to a person" rule	the model invents prices and terms
Document review	what to extract, table format, "not found"	the model fills in missing clauses
Code and scripts	language, input data, sample output, limits	too big a task in one go
Images	subject, style, light, angle, frame format	vague words like "beautiful"
Analysis	goal, data, criteria, output format	"analyse this" with no criteria

Prompts for image generation

Image models understand a description of a scene. A good image prompt answers: what is in the frame, where, in what style, what light, what angle, what format.

"A cappuccino on a wooden table by a window, soft morning light, side view, blurred background, photorealistic, horizontal 16:9 frame."

The more specific, the less randomness. If unwanted elements appear, many generators let you list separately what should not be in the frame.

Ready-made prompts for work

Copy them and replace the square brackets with your data.

Retelling news for a channel: "You are the editor of a [topic] channel. Retell the text below for [platform]. Reader: [who]. [N] short sentences, the first is the main fact. Only facts from the text. If something is missing, do not add it. No emoji, no conclusions."

Customer reply from a knowledge base: "You are a manager at [company]. Answer the customer's question using only the information below. Up to 3 sentences, polite. If the answer is not there, say you will check and pass it to a specialist. Information: [base]. Question: [question]."

Contract review: "Read the contract below and put into a table: parties, subject, amount, deadlines, liability, non-standard clauses. If an item is not found, write 'not stated'. Do not assume anything."

Content plan: "You are a marketer in [niche]. Suggest 10 post topics for a month on [platform]. Audience: [who]. For each: headline, one line on what it is about, format (post, carousel, video)."

Meeting summary: "Below is a meeting transcript. List: decisions made, tasks with owners and deadlines, open questions. Format: three lists. Do not add anything not in the text."

Improving your own text: "Edit the text below: remove filler and repetition, shorten sentences, keep the meaning and facts. Do not add new claims. Show the result and a list of what you changed."

Prompt engineering techniques that actually work

Examples instead of descriptions. Two or three samples set the style more precisely than a paragraph of adjectives. This is called few-shot prompting.
A rule for missing data. "If the text has no date, do not invent one." The main protection against made-up facts.
A plan before the answer. For complex tasks, ask for the steps first, then the execution.
Strict format. A list, a table, JSON. Then the answer slots straight into your work or automation.
Self-check. Ask the model to compare its answer with the source data at the end and fix discrepancies.
Splitting the task. Not "build a marketing strategy" but step by step: audience, topics, first post.
Iteration. A good prompt rarely works first time. Turn each mistake into one new rule.

What prompt engineering is and who needs it

Prompt engineering is the skill of writing and testing prompts so the model reliably gives the result you need. For a one-off chat question, common sense is enough. But when the model runs inside an automation and handles hundreds of requests a day, the prompt becomes a working tool: it is tested on dozens of real examples, the share of good answers is counted, and it is fixed mistake by mistake.

The system prompt of an AI agent

When a model works inside an automation or an AI agent rather than a chat, the prompt becomes a standing instruction: role, rules, priorities, prohibitions, output format. It is written once and refined on real mistakes. In my AI agent newsroom every role has its own instruction, and that is where most of the system's "intelligence" lives.

Using AI to write prompts

You can ask the model to write the prompt itself. What works best: describe the task in your own words and ask the model to ask you 5-7 clarifying questions. Answer them and let it assemble a prompt from your answers. Then test it on a few real examples.

Common mistakes

Too short. "Make it nice" means something different to the model.
Contradictions. "Detailed but short": the model picks one, not necessarily the right one.
No style examples. Without them the text comes out generic and AI-sounding.
Everything in one message. Ten tasks in one wall of text get an average result on all of them.
No data. The model does not know your prices and terms and fills in plausible inventions.
Not checking facts. Even with a good prompt, check numbers, dates and names.

How to tell a prompt is good

Collect 20-30 real task examples and run them through the prompt. Count how many answers are usable without edits. Note typical mistakes and add one rule for each. If the share of good answers grows after the edits, you are on the right track. That is how I test agent instructions before launch.

FAQ

What is a prompt in simple terms?

It is the task text for an AI model: what to do, for whom, with what data and in what form to return the result.

How do I write a good prompt?

State the role, task, context, data, output format and limits. Add two or three examples of the result and a rule for what to do when data is missing.

What is prompt engineering?

The skill of writing and testing prompts so the model reliably gives the result you need. In production systems it also means testing prompts on dozens of real examples.

How long should a prompt be?

Long enough to include everything needed and nothing extra. A simple question needs a couple of sentences; an agent's working instruction can run to a page.

Where can I get ready-made prompts?

Start with the templates in this article and add your data. Prompts from the internet should be adapted to your task and audience, or the result will be generic.

Should prompts be written in English?

Modern models understand many languages well. Write in the language you want the result in; it makes the style easier to control.

Why does the AI ignore my prompt?

Usually the prompt has contradictions, lacks data or the task is too big. Split it into steps, remove contradictions and add an example of the result.