AI chatbot prompt: how to write instructions and add a knowledge base

Without an AI chatbot prompt, the AI replies with generic phrases and makes up prices: it simply has nothing to rely on. The client leaves for a competitor, and the manager wastes time rewriting the bot's replies. If the price list is written right into the text, it goes out of date — and the arguments begin: “but the bot quoted a different price”.

A prompt, or instruction, is plain-language text the bot uses to decide what to reply and how. It turns the bot from an auto-responder into a sales consultant. There is no need to draw button flows: in plain words, you define who the bot introduces itself as, what it must not do, how it persuades and how it leads the client to a booking or payment. The bot takes prices and schedules from a Google Sheet with every reply: change a price, and the bot already quotes the new one. And the same instruction can be moved to a voice bot for phone calls.

Watch the two-minute video, and below you will find step-by-step setup.

Why a chatbot prompt matters for your business

• Online school: the facts hold graduates' results and reviews, the plan covers the goal, level, course selection and a trial lesson. The bot handles “too expensive” and “I'll think about it” on its own and gets the client to sign up.
• Clinic or salon: the rule “do not make diagnoses”, the price list and specialists' schedule from a sheet, the branch's time zone. The bot doesn't promise too much and doesn't book anyone “for yesterday”.
• Real estate: a plan of questions about budget, area, timing and mortgage, plus a property catalog from a sheet. The agent gets a client who has already been asked the key questions.
• Online store: stock and prices from a warehouse export via a link. The bot quotes today's stock, not yesterday's.
• Support: FAQ in a Google Doc and the rule “if the answer is not in the knowledge base, don't make it up; offer a manager”. The bot doesn't invent answers and honestly hands the question over to a person.

How a chatbot prompt works

With every reply, the bot sends the AI the whole instruction, the knowledge base data, the current date and time and the latest messages of the chat. Whatever is missing from the instruction, the bot fills in on its own. That is why prices, timelines and terms must be in the instruction or in the knowledge base.

If the bot was created by the assistant, all the blocks are already filled in — how that works is covered in the article AI chatbot for website. All that's left is to check them and adjust them to your business.

How to write an AI chatbot prompt: step by step

1. 1
   Editor

   Open “AI chat bot” → the bot's gear icon → “Edit” and go to the “Instruction” tab.
2. 2
   Role and style

   In “Sales role”, describe who the bot is, where it works and what its goal is. In “Speaking style”, set the tone, a formal or casual way of addressing the client and the reply length.
3. 3
   Rules

   Click “Add rule” for each ban or mandatory action: what must not be promised, when to call a manager, what data to collect.
4. 4
   Facts

   Click “Add fact”: prices, timelines, guarantees, cases, addresses.
5. 5
   Plan

   On the “Conversation plan” tab, fill in “Starting phrases” and “Conversation plan”. The last step of the plan is a call to action.
6. 6
   Save

   Click “Save changes” and test the bot in “Training AI” (more in the article how to train an AI chatbot).

Each item holds up to 2,000 characters, and the number of items is unlimited. You change the order by dragging the icon on the left.

If the bot has to qualify clients, build the plan as a short survey. In “Starting phrases”, say why you ask and how many questions there are: “To give you an idea of the price and timing, I’ll ask 3 questions — it takes a minute, okay?” Then in “Conversation plan”, ask about the task, budget and timing, with 3–4 answer options for each question, and make the phone number the last step. Clients answer more willingly when they can see the end of the survey. In Telegram, without this step the manager won’t get a number: the bot can’t see it.

Prompt templates, client data and triggers

The “…” button next to a field opens a window with the “Templates” and “Lead parameters” sections. Templates hold ready-made rules and objection-handling replies; save your own in “My templates”. In “Lead parameters”, choose a field, for example “Name”, and click “Insert”. The bot fills in the value from the client's card and addresses the client by name.

To make the bot send a price list or a video by itself, type “@” in the rule text and choose a saved trigger (how to create triggers is covered in the article Chatbot CRM integration). When the AI adds it to a reply, the trigger fires, and the client doesn't see the service code.

How to connect a knowledge base to a chatbot: the “Webhook” tab

The bot's knowledge base is your price list, schedule, stock and FAQ in a Google Sheet or Doc. A webhook here is a link the bot uses to read them with every reply.

1. In Google, set access to “Anyone with the link” and copy the link to the sheet you need.
2. On the “Webhook” tab, click “Add webhook”, give it a name, for example “Price list”, and paste the link into “URL”.
3. If you like, turn on “Caching” — replies will be faster. The “Reset-URL” link clears the cache, and your accounting system can call it too.
4. Add a rule: “Take prices only from the ‘Price list’ block; if an item is missing, say you will check with a manager.”

One webhook is one sheet of a spreadsheet, up to 100 KB of text.

Useful chatbot settings

On the “Settings” tab, below the channels:

• “Time zone” — the bot correctly understands “tomorrow” and “on Saturday”. The default is Moscow.
• “How many last messages to take from the bot” — the dialog memory, 100 by default. More means the bot remembers longer but replies slower and costs more.
• “Quick replies after the bot's answer” — hint buttons under the reply: the client answers with one tap, and triggers fire on that too. Add a permanent “Book now” button in “Custom buttons added to every reply”.
• “Add a random delay to the reply” — the reply arrives in one to three minutes, as if from a real manager.
• “Login blacklist” — spammers and competitors don't trigger the AI and don't create leads.

Read also: How to train an AI chatbot without prompts when it answers wrong, Chatbot CRM integration: creates leads, sends files and follows up, AI chatbot for website: how to create a bot that answers 24/7