Rule 1. A clear task and result
An AI does not know what is in your head. Name the action and the result: what to do, for whom, and how you will know the answer is ready. The more precise the request, the fewer redos and the faster you get what you need.
A request to an AI is called a prompt. The Anthropic guide has a simple test for a prompt: show it to a colleague who does not know the task. If they get lost, the model will miss too.
Bad: “Write about vacation”.
Good: “Write an email to my manager asking for vacation from 3 to 14 August. Polite tone, 5–6 sentences, and at the end suggest who will cover for me during that time”.
The rules on this page come from the guides of OpenAI, Anthropic and Google and work with any AI. ChatGPT, Claude and Gemini themselves are officially unavailable in Russia: for which services work there, see the page on AI services in Russia.
Rule 2. Context: who you are, who it is for, limits
Context is what you know and the model does not: who you are, who will read the answer, what already exists and what must not be done. Google’s guide advises putting this information in the prompt instead of hoping the model will guess.
Anthropic adds: explain why a limit matters. When the model understands the reason, it hits the target more precisely and applies it in similar places by itself.
Bad: “Explain what percentages are”.
Good: “I am a math teacher. Explain to 5th-grade students what percentages are, using a store discount as an example. No formulas: we will cover them later. 4–5 sentences and one example with numbers”.
Rule 3. An example or a sample
A sample shows tone, length and structure more precisely than any description. Google advises always adding examples to a prompt, and Anthropic calls them one of the most reliable ways to steer an answer.
This technique is called few-shot prompting, that is, a prompt with several examples. Anthropic recommends 3–5 varied examples similar to your task: with only one example, the model may copy it too literally. A sample can be your earlier text, an email you liked or a table of the right shape.
Bad: “Write captions for product photos”.
Good: “Write captions for 5 photos of mugs in this style. Sample: ‘A blue mug for your morning coffee. Dishwasher safe.’ Each caption up to 12 words, no exclamation marks”.
Rule 4. Answer format
If you do not set the form of the answer, the model chooses it, and often it is a long text with headings. The Google and Anthropic guides advise naming the form directly: a list, a table, a paragraph, a length. If the answer will go into an email, a presentation or a spreadsheet, say so: the model will adjust the layout.
Anthropic advises saying what to do, not what to avoid. Instead of “do not write too much”, write “answer in 3 points, one sentence each”.
Bad: “Compare three mobile plans”.
Good: “Compare the three plans from the text below in a table: price, gigabytes, minutes, what is not included. Under the table, one line: which plan is cheaper if I need 10 gigabytes a month”.
Rule 5. Split complex work into steps
A model handles a big task in one prompt worse: it loses something or mixes something up. Google advises giving one instruction per prompt instead of many in one, and OpenAI advises splitting a task into parts and making a plan first.
A convenient order is: a plan, then each part in turn, then a check of the whole. This sequence is called prompt chaining. This way you see a mistake at the step where it appeared.
Bad: “Write a business plan for a coffee shop”.
Good: “Step 1: list the sections of a business plan for a small coffee shop near a subway station. Next I will ask you to expand each section separately, and at the end to check whether costs and income add up”.
Rule 6. Refine instead of starting over
The first answer is rarely exact. Google writes directly that you sometimes need to rework a prompt several times before the answers become reliably useful. If the task is complex, ask the model to ask you clarifying questions first and only then answer.
In the same chat, the model sees what you have already discussed. Say what you liked and what to change: this is faster than explaining everything from scratch in a new chat. Anthropic also advises giving the model its own answer and asking it to check or extend it.
You need a new chat when the topic has changed or the conversation has drifted. A very long chat may not fit into the context window, that is, the amount of text the model takes into account at once.
Bad: “Wrong. Write it again”, five times in a row.
Good: “The start is good. Cut the second paragraph in half, remove the word ‘unique’ and add a question to the reader at the end”.
Rule 7. Check facts, numbers and links
A model can confidently write something false: invent a number, a date, an article of law or a link. This is called a hallucination, and Anthropic writes that it happens even with strong models.
The techniques in the guide reduce such errors but do not remove them. Allow the model to say “I don’t know”, ask for a source for each claim, and check what matters yourself against the original source.
One more technique: ask the same question twice, for example in a new chat, and compare the answers. If they differ, that is a signal to check the fact especially carefully.
Without search, a model does not know about recent events. Each model has a knowledge cutoff date: Anthropic defines it as the date up to which the model’s knowledge is complete and reliable.
Bad: “What is the fine for paying tax late?”, and then copying the number straight into a report.
Good: “What is the fine for paying tax late? Give the article of the law and a link to the official source. If you are not sure, say so”. Then open that article on the official website.
Rule 8. Do not send personal data or work secrets
Everything you write in a chat goes to the server of the company that made the service. In its Gemini privacy help, Google writes that human reviewers read some chats and asks you not to enter data you would not want them to see.
Do not write passport details, social security numbers, card numbers, passwords, diagnoses, data about clients and colleagues, or texts under a non-disclosure agreement. Before sending, replace names and numbers with placeholders: “Client A”, “contract X”.
What counts as personal data and what a company risks is covered on the page on personal data and AI and in the free course on AI safety. At work, first find out which service your company allows and which documents you may upload there.
Bad: “Check the contract with Acme Ltd, here are the company details and the director’s passport: ...”.
Good: “Check the terms of this supply contract: deadlines, penalties, how to terminate it. I replaced the names of the parties and the amounts with placeholders”.
What does not work: 5 common mistakes
1. One word instead of a task. Prompts such as “advert”, “essay” or “table” give a generic answer because the model does not know the goal. Write the action, the result and who it is for.
2. Trust without checking. Smooth text is not the same as correct text. Check numbers, dates, laws and links against the original source, especially in studies, reports and documents.
3. Personal data in a prompt. You cannot take back a passport, a password or client data once it is in a chat. Anonymize the text before sending.
4. “Make it nice” without criteria. The model does not know what nice means to you. Name the features: shorter, simpler, more formal, no bureaucratic words, for teenagers.
5. Expecting fresh news from a model without search. A model’s knowledge ends at its training date. For news, exchange rates and prices, turn on search or go to the official website.
A prompt builder helps you write a prompt without these mistakes: you fill in the fields, and it puts them together into a ready prompt.
Test yourself: a short quiz
Five questions on the rules above. Answers are checked right in your browser and sent nowhere, and after each answer you will see a short explanation. If you get one wrong, go back to the rule from the explanation and try again.
Answers are checked in your browser and are not sent anywhere.