---
title: "Prompt"
metaTitle: "Prompt"
description: "A prompt is a text instruction for an AI model. The five parts of a working prompt and how to check the answer."
date: 2026-11-17
tags: [ai-training, prompting]
cover: cover.webp
coverAlt: "A young woman with pink hair at a service hatch holds her task card with clipped pages while a finished document comes out of the hatch"
draft: false
sources:
  - "https://ru.hexlet.io/qna/vaybkoding/questions/chto-takoe-prompt"
  - "https://mediacritica.md/ru/разъясняем-что-такое-промпт/"
---

A prompt is the text instruction a person sends to an AI model: a question, request, or directive together with the relevant input. The model works only from what is in its context - your prompt, any attached files, and whatever the service itself adds, such as conversation history, system instructions, or retrieved material. It does not know your client, your rules, or what a good answer looks like until you write those things down. That is why a working prompt describes five things: the task and who will read the result, the source material, the answer format, what is off-limits, and how to verify the output.

## What a prompt is and related terms

The word "prompt" comes from everyday English, where it means a cue or signal that triggers a response. In work with AI models it refers to the text instruction you give the model - a question, request, or directive that tells it what to do. You may also see people search for this concept as "AI prompt."

Two related terms appear nearby. Prompting - also called prompt engineering - is the skill of writing effective instructions for AI models. A system prompt is a standing instruction that a service or developer attaches to every request you send, for example "reply in English, keep it brief, do not invent facts."

Majento teaches prompting to employees in [corporate AI training](/ai-training/). The rest of this article covers the foundation: the five parts of a working prompt and a step-by-step example.

## Why a short prompt produces unpredictable answers

A language model continues text in the most probable way given what it received. If the prompt says "write a letter to the client," the model has no information about the client, the occasion, the tone, or what promises are acceptable. It fills the gaps with generic language and sometimes invented details - deadlines, discounts, contract terms. The answer reads smoothly, but there is nothing to check it against.

A second issue is that outputs vary between runs. The same prompt today and tomorrow can return different wording. A useful prompt therefore includes criteria that let a person immediately judge whether the answer is acceptable. Without those criteria, the employee evaluates the result by feel, and errors pass through unnoticed.

## What a working prompt contains

The five parts below cover most professional tasks. Order does not matter; completeness does.

| Part | What to write | Example phrase |
| --- | --- | --- |
| Task and reader | What the output should be and who will read it | "Draft a reply to a client asking about delivery timing" |
| Source material | The text, table, or document the answer should draw from | "Below is the client's message and a row from the shipping schedule" |
| Format | Length, structure, language, tone | "3-4 sentences, formal register, no marketing language" |
| Restrictions | What the model must not do | "Do not promise discounts or dates that are not in the schedule" |
| Verification | How the answer should show its basis and flag gaps | "If the delivery date is not in the data, say so. State which row of the schedule the date comes from" |

A role instruction such as "you are an experienced sales manager" can help with tone. On its own it does not replace source data or restrictions. A model given an expert role but no facts invents details with greater confidence.

## Teaching example: one prompt in three versions

This example is constructed for analysis; the company, client, and data are fictional. A manager has received the following message.

**Client message**

> Good afternoon. When will our order 1042 arrive? We need to plan the warehouse receiving. Also - is it possible to pay for the next order in installments?
>
> Dilnoza, Purchasing, Example LLC

The manager's shipping schedule contains one relevant row: "Order 1042, dispatch 20 Nov, delivery 2-3 business days." The manager's materials contain nothing about payment terms.

**Version 1.** "Write to the client about the delivery."

The model is missing everything: which order, what date, what language, what tone. The likely result is a polite template with an invented date or a phrase like "shortly" that the client has already heard.

**Version 2.** "Write a polite reply to the client at Example LLC about the delivery date for order 1042. We are dispatching on 20 November, delivery takes 2-3 days."

Facts are now present, so the answer will be specific. Gaps remain, however. The model may add an apology for a delay that did not happen, or promise a courier call. It is also unclear what to do with the client's second question about installment payments.

**Prompt (version 3)**

```
Draft a reply to the client's message below. The reader is the client's purchasing contact, who needs a date to plan warehouse receiving.

Source data: the client's message and this row from the shipping schedule: "Order 1042, dispatch 20 Nov, delivery 2-3 business days."

Format: 3-4 sentences, formal register, no apology unless the data shows a delay.

Restrictions: do not promise discounts, phone calls, or dates that are not in the schedule. Do not answer questions for which no data is provided.

Verification: if the client asks about something not covered by the source data, list those questions separately after the letter under the heading "needs manager input."

Client message: "Good afternoon. When will our order 1042 arrive? We need to plan the warehouse receiving. Also - is it possible to pay for the next order in installments? Dilnoza, Purchasing, Example LLC."
```

The following reply shows what version 3 can produce. It was written by the article's author for analysis and is not a recorded model output.

**Reply (written by the author for analysis)**

> Hello Dilnoza. Order 1042 is scheduled for dispatch on 20 November. Delivery takes 2-3 business days from dispatch, so please plan receiving for that window.
>
> Needs manager input:
> - Whether the next order can be paid in installments - payment terms are not in the source data.

The reply now has a clear basis and a clear boundary. The manager checks three things.

1. The date matches the schedule: the order number, 20 November, and 2-3 business days all come from the schedule row. An exact arrival date is not in the reply because it is not in the data.
2. No promises go beyond the data: no apology, no discount, no courier call.
3. The installment question is not buried in the letter - it is pulled into a separate list. The manager answers it by confirming the terms with whoever owns that decision, then adds a paragraph before sending.

Because the criteria were written in advance, the manager checks only the date, the promises, and the unanswered questions. To repeat this exercise, paste version 3 into an approved service and check your output against the same three points. Your wording will likely differ. If the answer fails any point, that is the signal to extend the prompt.

![A young woman with pink hair clips the third of three notes with the same request to a line, from a torn scrap to a filled-in card with a receipt and a check mark below it](/blog/promt/fig-three-requests.webp)

## How to check an AI model's answer

Every claim in the answer that looks like a fact must appear in the source data. In practice, checking involves four steps.

1. Match every number, date, and name against the source. If the source is a table, recalculate the total for at least one row manually.
2. Find statements that are not in the source material. Remove them or send them to someone who can confirm the facts.
3. Check for omissions. Models often silently skip an awkward item - for example, the client's second question.
4. Record what went wrong, add a rule to the prompt, and run the same example again. The error can recur, so keep the test case even after editing the prompt.

![In a warehouse room during the day, a hand replaces an order tag on a rail above a box with a matching colour label; a young woman with pink hair holds a clipboard beside it](/blog/promt/fig-replace-order.webp)

If the answer fails verification twice, the problem is usually in the prompt itself - missing source data, a missing restriction, or a missing criterion. Rewriting the answer by hand at that point takes longer than fixing the prompt.

## What data must not go into a prompt

Everything you paste into a prompt is sent to the service that processes the request. Before working with real business documents, a team needs to decide which service is approved, which data may be shared, and who is responsible for that decision. Client personal data, commercial terms, and internal reports must not go into a public chatbot without such a decision in place. Training exercises work well on anonymized examples. How to prepare that kind of material for a team is covered in the article on [data safety when working with AI](/blog/safe-materials-team-ai-training/).

## Eight ready-made prompts for common tasks

Below are eight templates for typical office tasks: client reply, meeting summary, document comparison, review analysis, table check, product description, policy-based reply, and meeting agenda. The first is version 3 from the teaching example turned into a reusable template.

Text in square brackets marks the fields to replace with your own data: [client message], [table], [policy]. Replace the entire bracketed phrase and delete the brackets. Paste source material after the prompt or attach it as a file if the service supports that. Under each template is a short checklist of what to verify before passing the answer on.

### 1. Client reply based on known facts

```
Draft a reply to the client's message below. The reader is [who the client is and what they need].
Base the reply only on this data: [facts: dates, prices, order status].
Format: [3-5 sentences], formal register, no promotional language.
Do not promise anything not in the data: no discounts, no dates, no phone calls.
If the client asks about something not covered by the data, list those questions after the letter under the heading "needs manager input."

Client message: [text]
```

**What to check.** Every date and number in the reply exists in your data. Questions the client asked but the data does not cover are listed separately, not silently dropped.

### 2. Meeting summary with decisions

```
Below are the transcript or notes from a meeting. Write a summary for [who will read it - for example, a manager who was not present].
Structure: decisions made; action items with owner and deadline; open questions.
Write only what is explicitly stated in the text. If an action item has no owner or deadline, write "not assigned."
Do not add conclusions or judgments of your own.

Text: [notes or transcript]
```

**What to check.** Every decision in the summary was actually made at the meeting. Action items carry the names and deadlines that were stated. Models often turn "let's think about it" into a decision.

### 3. Comparison of two documents

```
Compare the two documents below: [Document A] and [Document B].
Find differences in [quantities, dates, prices, payment terms, responsibilities of each party].
For each difference, quote the exact wording from both documents and cite the section or clause number.
If a clause appears in only one document, flag it separately.
Do not interpret the terms or recommend which version is correct.
```

**What to check.** Search the source files to confirm each quoted passage. Go through the clauses you already know are disputed - the model may have missed them. Legal interpretation of the differences is for a lawyer, not the model.

### 4. Customer review analysis by theme

```
Below are [number] customer reviews. Group them by complaint and praise themes.
For each theme, state the number of reviews and include two verbatim quotes.
A theme counts as recurring if it contains at least [3] reviews. Put the rest in a section called "one-off."
Do not invent causes for the problems. If a review is unclear, put it in a section called "unclear."

Reviews: [list, one per line]
```

**What to check.** Count the reviews in at least one large theme manually. Confirm that the quotes are verbatim. If the total number of reviews exceeds what fits in a single request, split them into batches and consolidate the results separately.

### 5. Table check before a report

```
Below is a table [what it contains - for example, monthly sales by branch].
Find rows with missing values, duplicates, negative numbers, and figures that differ from adjacent rows by more than [10] times.
For each finding, state the row number and column name.
Do not correct any data or calculate totals.

Table: [paste or file]
```

**What to check.** The model flags suspicious entries but does not calculate totals - sums and percentages are more reliably checked with a formula in the table itself. Open a couple of flagged rows in the source file and confirm that the row numbers match. A detailed walkthrough of this kind of check is in the article "[Analyzing Excel tables with AI](/blog/ai-excel-dashboard-audit/)."

### 6. Product description from a spec sheet

```
Write a product description for [website, marketplace] based on the spec sheet below.
The reader is [who buys it and what they use it for].
Length: [up to 600 characters]. Tone: [calm, no superlatives].
Use only the attributes in the spec sheet. Do not add properties, certifications, or comparisons with competitors.
At the end, list any attributes that are missing but would be needed for a complete description.

Spec sheet: [name, materials, dimensions, intended use]
```

**What to check.** The text contains no properties that are not in the spec sheet. Names, dimensions, and materials are spelled exactly as in the spec sheet. Remove any claims like "suitable for everyone."

### 7. Policy-based reply to an employee question

```
Answer the employee's question using only the policy text below.
Question: [question text].
Quote the relevant clause number and the exact wording from the policy that supports the answer.
If the policy does not cover the question, write "not addressed in the policy" and name who to contact: [department or role].

Policy: [text]
```

**What to check.** The clause number exists and the quote matches. If the answer looks reasonable but there is no quote to support it, treat the policy as not covering the question.

### 8. Meeting agenda from a list of topics

```
Create an agenda for a [duration] meeting with [participants and their roles].
Topics for discussion: [list].
For each topic, state the goal (decide, approve, or inform), the time allocated, and who prepares the material.
If there are more topics than the time allows, suggest which to defer and explain why.
```

**What to check.** The time allocations add up to no more than the meeting length. Topics that require a decision have a named decision-maker.

## When a single prompt is not enough

A prompt solves a one-off task for one person. Problems arise when the same task recurs across ten employees. Each person writes their own version, checks the output differently, and the result depends on who is working that day.

At that point, a successful prompt becomes a team template - like the eight examples above. A team template needs four things.

1. An owner. This is the person who knows what a correct result looks like and updates the template when rules, prices, or policies change.
2. A sample input and a sample verified output. These show a new employee what the result should look like.
3. A verification checklist, like the ones under each prompt above.
4. A date of last edit and a brief note of what changed and why.

Without an owner, a template library goes stale quickly. Employees keep using a template that references last year's price list, and the error multiplies. How to structure such a template with ten fields and a change log is shown in the article "[Prompt Engineering](/blog/teach-employees-ai-task-briefs/)." When the task is handled not by a chat interface but by an AI agent with access to business systems, the prompt also needs defined permissions and approval checkpoints - covered in "[How to brief an AI agent](/blog/ai-agent-task-brief/)."

Majento runs [corporate AI training](/ai-training/) in Tashkent and online. Teams build a template library around their own work: employees bring recurring tasks, write prompts on anonymized examples, and check outputs against agreed criteria. At the end of the program, the department has tested templates with named owners and a list of tasks worth handing to a system. Those can be developed further during [AI implementation](/ai-implementation/) or [business process automation](/business-automation/).

## Frequently asked questions

### What is a prompt in simple terms

A prompt is a plain-text instruction for an AI model. The person writing it explains what needs to be done, what material to use, what form the answer should take, and what is not allowed. The less the model has to guess, the more accurate and verifiable the output.

### What is the difference between a prompt and a system prompt

A prompt is the instruction you write for a specific task. A system prompt is a standing instruction that a service or developer attaches to every request automatically - for example, "always reply in English, keep answers brief, do not invent facts." You usually do not see the system prompt; it runs in the background.

### Where can I find ready-made prompts

Open collections of prompts are useful as examples of structure. For a real work task you need a template with fill-in fields, explicit restrictions, and a verification checklist. Eight such templates are in the section "Eight ready-made prompts for common tasks" above. Before using any of them, test it on your own approved data. The most reliable option is a library of prompts that have already been tested on your team's actual tasks and each has a named owner.

### Do these prompts work in any AI model

The templates are written in plain language and are not tied to a specific service. Different models follow formatting instructions and restrictions to different degrees, so when you switch services, run the template on two or three known examples to confirm it still behaves as expected.

### What is a prompt engineer

A prompt engineer is someone skilled at writing effective instructions for AI models. In a working team it is more practical for everyone who assigns tasks to a model to have a basic level of that skill. Reusable templates should be owned by the process owner - the person who knows what a correct result looks like.

### Does the language of a prompt matter

The language of a prompt matters less than people expect, because current models handle many languages well. Write in the language your source material is in and the language you need the answer in. Mixing languages within a single prompt without a reason can cause the model to reply in the wrong one.

*We can train your team to write prompts for their own tasks and verify the outputs. Reach us on Telegram at [t.me/shimaoz](https://t.me/shimaoz) or by email at [hello@majento.ai](mailto:hello@majento.ai).*
