You ask a neural network to help with a complaint to a store — and get a dry template, as if written by a robot for a robot. Or you ask it to analyze a contract, and the response is a retelling of the first two paragraphs. Sound familiar?
The problem is almost never the neural network itself. The problem is how we ask it. A prompt is your request, the instruction you give to the neural network. And if it lacks details, the model fills the gaps with guesses. Sometimes good ones, more often not.
The good news: to get the desired result on the first try, you don't need to be a programmer. There is a simple five-step method: role, task, context, format, check. Let's break it down using everyday examples — a complaint to a store, analyzing a contract, and planning a vacation. And we'll immediately show what the AI twin takes on at each step — an assistant to whom you give access to your email, documents, and calendar, and it does the work for you.
What is a prompt and why the neural network "doesn't understand"
The neural network doesn't read your thoughts. It only sees the text you wrote. If you write "help with a complaint," it doesn't know which store, for what amount, what exactly happened, and what result you need.
Imagine you ask a colleague: "Make a report." He will make one — but most likely not the one you expected. But if you say: "Make a sales report for September, as a table, with the total at the bottom" — you'll get almost what you need. It's exactly the same with a neural network.
Prompt engineering in simple terms is the ability to formulate a request so that it is understood literally and precisely. There's no technique here: only structure and specificity.
Step 1. Role: say who should answer
Start with a phrase like "You are an experienced consumer protection lawyer" or "You are a travel agent with twenty years of experience." This sets the style and depth of the answer. Without a role, the neural network answers "on average" — in general terms.
What the human does: decides whose expertise is needed in a specific situation. A lawyer? An accountant? A consulting physician (to understand terms, but not for treatment)?
What the AI twin does: remembers your typical roles and context. If you've already asked about a product return, it will pull up the history from the knowledge graph and won't ask the same thing again. It can also suggest a role itself — for example, "consumer disputes lawyer" — if it sees you're working on a complaint.
Example prompt:
"You are a consumer protection lawyer. Answer in simple language, without Latin or complex terms."
Step 2. Task: one action — one request
Formulate exactly what needs to be done. Not "deal with the contract," but "find the clauses about penalties and termination periods in the contract and explain them in simple words." The more specific the verb, the more accurate the answer.
Bad: "Help with vacation."
Good: "Create a 10-day vacation plan in Sochi for a family of three with a 7-year-old child, budget up to 150,000 rubles."
What the human does: formulates the task and decides what result they need at the output — text, table, list, letter.
What the AI twin does: at this step it already starts collecting data. For a vacation plan, it looks at your calendar, finds free dates, checks email for bookings and tickets. For a complaint, it searches email for the receipt and correspondence with the store.
Step 3. Context: give facts, not hints
The neural network doesn't know that you bought a kettle for 4,900 rubles on September 12, that it broke after a week, and that you lost the receipt but have a bank statement. All of this needs to be written — or give access to the documents.
What the human does: recalls details: dates, amounts, names, order numbers, what exactly went wrong.
What the AI twin does: finds the documents itself. It reads email and finds the order confirmation, downloads the receipt from the attachment, opens the bank statement and verifies the amount. If something is missing, it says directly: "No receipt, but there is a payment confirmation from September 12, that's enough." It can also search the internet if legal norms or store rules are needed.
Example prompt with context:
"Context: on September 12, 2026, I bought an electric kettle in an online store for 4,900 rubles. After 8 days it stopped turning on. The receipt is lost, but there is a payment confirmation from the bank. The store refuses to accept the return, citing that the item was used."
Step 4. Format: say how the answer should look
The neural network can output the answer as a list, table, letter, step-by-step instruction, or a short paragraph. If you don't specify, it will choose itself, and it's not certain it will suit you.
What the human does: decides in what form it's convenient to receive the result. For example: "Format it as a ready letter that can be printed and taken to the store."
What the AI twin does: assembles the result into the needed file. It can generate a ready PDF with the complaint, save it in your documents, and if needed — send it as an email through your mailbox. For a contract, it will make a table: clause — meaning — what to pay attention to.
Format examples:
- "Format as an official one-page complaint with a header and date."
- "Make a table: contract clause, what it's about, what I risk."
- "Give a list of 5 points, without introductions."
Step 5. Check: don't take it on faith, verify
Even an ideal prompt doesn't guarantee an ideal answer. The neural network may err in the article number of a law, mix up dates, or invent a non-existent norm. Therefore the last step is verification.
What the human does: reads the result and verifies key facts — amounts, dates, references to laws. If something is unclear, asks a clarifying question: "Where did you get this article? Show the source."
What the AI twin does: checks the arithmetic itself — adds up amounts, calculates deadlines. It can find the text of the law on the internet and provide a link to the official source. It also sets a reminder: "If the store doesn't respond within 10 days, remind about the next step." And it monitors the deadline — you don't need to keep it in your head.
Three everyday examples: how it works together
Example 1. Complaint to a store
You bought a kettle, it broke, the store refuses. Your prompt: role of a lawyer, task — draft a complaint, context — dates and amount, format — ready letter, check — verify the article of law.
The AI twin finds the order confirmation and payment in email, verifies the amount, checks the return rules on the store's website, drafts the complaint text in PDF, and sets a reminder for the 10th day. You read it, check the article number, and take it to the store.
Example 2. Analyzing a contract
You received a contract for apartment renovation and don't understand what you're signing. Prompt: role of a lawyer, task — find risks, context — contract file, format — table, check — clarify unclear clauses.
The AI twin reads the document, extracts clauses about penalties, deadlines, and prepayment, lays them out in a table, and marks places where you should ask the contractor questions. It does not certify the contract or sign it — only a human does that.
Example 3. Vacation planning
Prompt: role of a travel agent, task — 10-day plan, context — family composition and budget, format — table by days, check — stay within budget.
The AI twin looks at the calendar and finds free dates, checks email for old bookings, searches for accommodation and ticket options, calculates the total amount and compares it with the budget. If it comes out more expensive, it suggests what to replace.
| Step | What the human does | What the AI twin does |
|---|---|---|
| Role | Chooses expertise | Suggests a role from history |
| Task | Formulates the request | Collects data from email and calendar |
| Context | Recalls details | Finds receipts, contracts, correspondence |
| Format | Chooses the answer type | Assembles PDF, table, letter |
| Check | Verifies facts | Calculates amounts, sets reminders |
Frequently asked questions
Do I need to know English or special words for prompts?
No. Modern neural networks understand Russian well. Structure and specificity are more important than language. Write as you would explain a task to a colleague.
Why does the neural network sometimes invent facts?
It predicts text, not checks a knowledge base. Therefore ask for sources and verify important things — article numbers, amounts, dates. The AI twin helps: it searches for primary sources and checks arithmetic.
What exactly will the AI twin do if I ask it to draft a complaint?
It will find the purchase confirmation and payment in email, check the store's rules, draft the complaint text in PDF, save it in documents, and set a reminder about the response deadline. It will not file an application with government agencies — that requires a configured integration.
Can I give the neural network access to a bank statement?
Yes, if you have connected the service. The AI twin uses access only for a specific task — for example, to verify the purchase amount. It does not make payments or banking operations without a connected service.
How much time does a good prompt save?
Usually one or two clarifications are enough instead of five or six. The more precise the instruction, the fewer edits. In practice, this saves dozens of minutes on each task.
Will the AI twin replace a lawyer or doctor?
No. It prepares drafts, searches for information, and reminds about deadlines, but does not replace a specialized professional. Legally significant documents are signed by a human.
Checklist: what to do right now
- Formulate the task in one sentence — what exactly you need at the output.
- Add a role: "You are an expert in..." (the AI twin will suggest a role from your history).
- Gather context: dates, amounts, details (the AI twin will find receipts and correspondence in email).
- Specify the format: letter, table, list, PDF (the AI twin will assemble the file).
- Ask for sources for important facts and verify them.
- Set a reminder about the deadline (the AI twin will do it itself).
- Save a successful prompt — it will come in handy again.
You can start with one prompt — for example, ask the AI twin to find a receipt in email and prepare a draft complaint. Then it will suggest what's missing. More about how to connect services and delegate tasks — at asibiont.com.
Comments