You ask an AI to help with a letter to the property management company, and you get a dry template from the internet. Or you ask for a vacation plan — and it gives you a list of "visit museums and try the local cuisine." Sound familiar? The problem isn't the AI, it's the request. More precisely, it's that we phrase it the way we talk to a person, while the machine understands it literally.
A prompt is a text instruction for an AI. The more precisely you describe who you are, what you need, and in what form, the fewer iterations you'll need. In my experience, one well-crafted prompt saves 3–5 clarifying requests. And if you bring in an AI twin (for example, ASIBiont), it will turn your everyday task into a structured prompt, ask the missing questions, and take it all the way to a finished document.
This article covers 7 rules that work for any model (ChatGPT, Claude, Gemini, YandexGPT), plus ready-made templates for four real-life tasks. At each step I'll show what you do and what the AI twin takes on at that moment.
Why the AI gives you "the wrong answer"
An AI doesn't read minds. It predicts the next text based on your request and its training data. If the request allows for different interpretations, the model will pick the most probable one — and it's not guaranteed to be yours.
Typical reasons for failure:
- No role or context. "Write a letter" — to whom? why? in what tone? The model will make it up.
- Too general a goal. "Help with vacation" — that could be a plan, a budget, a route, or a packing list.
- No format specified. You expect a table, but you get a wall of text.
- No constraints. The model writes 2,000 words when you need 150.
- No example. Without a sample of "how it should be," the style of the answer is a lottery.
7 prompt rules that work
Rule 1. Set a role
Start with the phrase "You are a …". The role narrows the vocabulary and style. Compare: "Write a letter to the property management company" and "You are a housing and utilities lawyer writing an official letter to the property management company." In the second case, the model uses the right terms ("improper performance," "recalculation," references to regulations).
You: choose who the model should "pretend to be" — a lawyer, a doctor (for explanation, not treatment), an editor, a curriculum designer.
AI twin: selects the role automatically based on the task type. For a letter to the property management company — a legal profile; for a vacation plan — a travel planner. You don't need to remember which role to specify.
Rule 2. Give context
The model doesn't know that your faucet is leaking, your child is in 8th grade, and your vacation is in July. Context is facts: who you are, what you've already tried, what constraints exist.
You: list 3–5 key facts. For example: "I own an apartment on the 5th floor, the leak report was drawn up on October 12, the property management company hasn't responded for 10 days."
AI twin: finds the facts in your connected services. For a letter to the property management company, it searches your email for correspondence with the company, your documents for the report and the management agreement, and your calendar for the dates of your requests. And it asks a clarifying question if something is missing: "Do you have the report? Send a photo or PDF."
Rule 3. State the goal in one sentence
The goal is the result, not the process. Not "tell me about vacation," but "create a 10-day vacation plan in Sochi with a budget of up to 80,000 rubles for two, no excursions, with two rest days." The more specific the result, the less fluff.
You: write what should come out — a letter, a table, a list, a PDF.
AI twin: reformulates your everyday request ("I want a vacation") into a measurable goal and suggests options if there's not enough data. For example, it will ask: "Is budget or comfort more important to you?"
Rule 4. Specify the response format
Models understand formats: "a table with columns," "a numbered list," "a one-page letter," "JSON" (if you're a technical specialist). Without a format, you'll get an essay.
You: specify the desired form: "Table: day, activity, budget."
AI twin: chooses the format for the task and saves the result. Vacation plan — a table in a document; letter — ready text in your email; contract review — a list of risks with explanations. You don't have to copy from the chat into Word afterward.
Rule 5. Show an example (if you have one)
One example of "how it should be" replaces three paragraphs of explanation. If you have a successful letter or plan, paste a fragment and ask it to "do it in the same style."
You: attach a sample — an old document, letter, or table.
AI twin: finds similar documents in your storage (for example, a past letter to the property management company) and uses them as a style template. If there are no samples, it will suggest its own and ask if it works.
Rule 6. Set constraints
Constraints are boundaries: length, tone, prohibitions. "No more than 1,500 characters," "no bureaucratic jargon," "don't suggest paid courses," "only verified sources." They cut out what you definitely don't need.
You: list 2–4 prohibitions or limits.
AI twin: applies standard constraints from your profile (for example, "don't send letters without confirmation") and checks the answer for compliance. If the model violated a constraint, it rewrites it.
Rule 7. Clarify and verify
Even a perfect prompt sometimes fails. A good practice is to ask the model: "Ask me 3 clarifying questions before answering." That's cheaper than redoing a finished text.
You: answer the clarifications and check the facts (amounts, dates, article numbers).
AI twin: asks clarifying questions itself, then cross-checks the result against your data. For example, when reviewing a contract, it verifies that all deadlines and amounts match the source file and flags places where your decision is needed (signature, sending).
Ready-made templates for 4 tasks
Letter to the property management company
Prompt: "You are a housing and utilities lawyer. Draft an official letter to the property management company about a recalculation for a roof leak. Context: report dated 10/12/2026, the company hasn't responded within 10 business days. Goal — get a recalculation and repairs. Format: a one-page letter, header with the addressee, references to the norms of the Housing Code of the Russian Federation. Constraints: no emotions, only facts and demands."
What the AI twin does: finds the correspondence with the company in your email, the report and management agreement in your documents, fills in the dates and amounts, generates the letter, and saves it as a PDF. You confirm the sending yourself — it won't send anything without your command.
Contract review
Prompt: "You are a lawyer. Review an apartment lease agreement. Find 5 risks for the tenant: penalties, termination procedure, deposit. Format: a table — clause, risk, what to do. Constraints: don't give legal advice, only analyze the text."
What the AI twin does: loads the contract from your documents, extracts the text, marks the clauses, and compiles a risk table. Important: it doesn't replace a lawyer and doesn't certify documents — the final decision is yours.
Vacation plan
Prompt: "You are a travel planner. Create a 10-day vacation plan in Sochi for two, budget 80,000 ₽ excluding flights. Format: a table — day, activity, approximate budget. Constraints: no more than 2 excursions, two rest days, no nighttime events."
What the AI twin does: checks the dates in your calendar, searches for current prices for accommodation and tickets via search, compiles the table, and saves the document. If the budget doesn't add up, it will suggest options.
Exam preparation
Prompt: "You are a curriculum designer. Create a 14-day exam preparation plan for history, 1 hour per day. Format: a table — day, topic, what to read, 3 self-check questions. Constraints: only topics from the list (attached), no unnecessary theory."
What the AI twin does: breaks the topics down by day, sets reminders in your calendar, and sends self-check questions. You control the answer checking and discipline yourself.
Frequently asked questions
— How many characters should an ideal prompt be?
There's no strict limit. For everyday tasks, 200–500 characters is enough: role, context, goal, format. If the task is complex (contract review, a monthly plan), the length grows to 1,000–1,500 characters. What matters isn't length but the absence of ambiguity.
— Should I write the prompt in English?
No. Modern models (GPT-4, Claude, Gemini, YandexGPT) understand Russian well. English may give a slightly more precise result on technical topics, but for letters, plans, and everyday tasks, write in your native language.
— What if the AI still gives the wrong answer?
Add a clarification: "You didn't account for X, rewrite it taking Y into account." Or ask the model to ask you 3 questions. If that doesn't help either, split the task in two: first ask for a plan, then for the text based on the plan.
— Can the AI twin compose the prompt for me?
Yes. You state the task in everyday language ("write a letter to the property management company about a leak"), and the AI twin turns it into a structured request, finds the necessary documents and data, asks clarifications, and delivers the finished result — a letter, table, or plan.
— Is prompt engineering dangerous for a beginner?
No. It's just a set of formulation rules. Mistakes don't break anything: at worst you get an inaccurate answer and rephrase. Risk arises only if you trust the model with legal or medical decisions without verification — that requires a specialist.
— Where should I store successful prompts?
In notes or in the AI twin's knowledge base. ASIBiont remembers facts about you and your templates so you don't have to explain everything again next time. Just say "do it like last time."
Checklist: what to do right now
- Choose one everyday task — a letter, a contract review, a vacation plan, or exam preparation. (The AI twin can suggest a task based on your documents and calendar.)
- Write a prompt using the 7 rules: role, context, goal, format, example, constraints, clarification. (The AI twin will draft the prompt from your wording.)
- Add 2–3 real-life facts — dates, amounts, names. (The AI twin will pull them from your email and documents.)
- Ask the model to ask clarifying questions before answering. (The AI twin will do this automatically.)
- Check the result — amounts, deadlines, references to regulations. (The AI twin will cross-check with the source files and flag discrepancies.)
- Save the successful prompt in your notes or the AI twin's memory. (The AI twin will remember the template and reuse it.)
- Connect the necessary services — email, calendar, documents — so the AI twin can work with your data. Connection instructions are available at asibiont.com/courses.
The easiest way to start is with one task: go to asibiont.com, describe it in your own words, and see how the AI twin turns it into a prompt and takes it to a finished document. And if you want to dive deeper into prompts for different fields, check out the article on prompts for infrastructure and the collection of working prompts for ChatGPT.
Comments