What you will learn
- How to describe a task so you do not have to redo the answer
- What context to give the model and what you must never pass to it
- How to check the facts, numbers and links in an answer
- How to refine a request step by step instead of starting over
Key terms
- Prompt
- The text of a task for a model: what to do, for whom, in what form and under what rules.
- Context
- The material the model should rely on: source text, facts, examples, constraints.
- Hallucination
- A confident but invented answer: a fact, number or link that does not exist.
- Brief
- A short description of a task: audience, goal, materials, format and the criterion for “done”.
01Start with the result
Before you write a request, answer for yourself: what should come out, for whom, and how will you know the work is done. “Write about our product” is not a task. “A 60-second video script for a creator who is hearing about donations for the first time, in plain language, ending with an invitation to try it” is a task.
- Audience: who will read or watch it
- Goal: what the person should understand or do
- Format: length, structure, tone
- Done: the sign by which you will accept the work
Example. “Prepare a 600–800 character post for a Telegram channel for streamers. The goal is to explain that donations arrive straight in the creator’s wallet. Friendly tone, no jargon, and at the end an invitation to open the donations page.”
02Give the right context
The model does not know your product, audience or constraints until you tell it. Add to the request the source text, verified facts, a good example and the rules: what you must not promise and which words to avoid.
Separate facts from assumptions: mark what is verified and what is your hypothesis. Otherwise the model will present one as the other.
- Do not paste passwords, API keys, recovery phrases or other people’s personal data
- Do not upload confidential documents to public AI services without the owner’s permission
03Get a draft and go through it
The first answer is a draft. Read it through your audience’s eyes: is it clear, is anything unnecessary, is the format respected? Point out specific problems, not a general “I do not like it”.
Tip. Ask the model to suggest two or three versions of the opening or the headline: choosing is easier than explaining what is wrong.
04Check facts, numbers and links
Models sometimes invent things with confidence: studies that do not exist, wrong numbers, broken links. Check every name, number, date and link against your source materials or the original source.
Important. Do not publish an AI answer, or send it to a client, with unchecked facts: you are responsible for the text.
05Refine one problem at a time
Instead of a new request like “make it better”, give a specific edit: “Shorten the introduction to two sentences”, “Explain the term ‘network fee’ in one sentence”, “Remove promises that are not in the facts”. That way the good parts stay and the weak ones get fixed.
Before you use the answer
- The result matches the goal and the audience
- The format and length are respected
- All facts, numbers and links are checked against sources
- The request contained no secrets and no one else’s personal data
- A person has read the whole text
Common mistakes and what to do instead
A short request with no context: “Write a post”
Describe the audience, goal and format, and attach materials.
Copying the answer without checking
Check names, numbers and links against sources.
Rewriting the request from scratch after every answer
Refine one specific problem at a time.
Pasting passwords and keys into the chat
Secrets never go into requests.
Try it yourself
Write a five-line brief for a real task of yours and check whether another person could do the work from it.
- Audience: who the result is for
- Task: what you need to get
- Materials: what to attach to the request
- Format: length, structure, tone
- Done: the sign by which you will accept the work
Result. A brief that both a model and a person will understand in the same way.
Test yourself
Answer in your head first, then reveal the answer.
1Why should you check a model’s answer even when it sounds confident?Answer
Models can invent facts, numbers and links. A confident tone does not guarantee truth.
2What must you never paste into a request?Answer
Passwords, API keys, recovery phrases, and other people’s personal or confidential data.
3The answer is almost good, but the introduction is too long. What should you do?Answer
Ask for a specific edit, such as shortening the introduction to two sentences, instead of rewriting the whole request.
In short
Start with the result, give context without secrets, treat the answer as a draft, check the facts and refine one problem at a time. Then AI saves you time instead of adding mistakes.