Practical guide · August 4, 2026
The system prompt for an AI auto-reply: how to write it well
The system prompt sets how the assistant behaves inside one profile: whose name it answers in, what job it does, which information it may use, and what it does when that information runs out.
This is part of the guide “A free Telegram auto-reply: how to connect it without Premium”. This page is only about the system prompt.
What a system prompt is
The system prompt is the main rule the assistant works by inside a profile. The reply style sets the shape of a message; the system prompt sets the job and the boundaries.
The style might ask for short answers with no emoji, while the system prompt says to use only the knowledge base, never confirm discounts and pass hard questions to the owner.
What it is made of
1. Whose name the assistant answers in
State the role as plainly as you can:
You are the company's assistant.
You are a technical support assistant.
You are the owner's personal assistant.
This sets the context and keeps work and personal scenarios from blurring into each other.
2. What job it does
Describe the profile's main goal:
Answer questions about services, prices and bookings.
Help people find a solution from the instructions and the knowledge base.
Collect the key details of a partnership proposal and explain what happens next.
One profile should have one main function. If a single prompt covers sales, support, hiring and personal messages at once, split them into several profiles — how to do that is in the article “Profiles in HeadHack”.
3. Which information it may use
Fix the source:
Use only the FAQ, facts and documents belonging to this profile.
Never fill in missing terms and never invent a price.
This matters most for prices, deadlines, discounts, refunds, technical causes and anything else that has to be passed on exactly.
4. What to do when information is missing
Decide the safe behaviour in advance:
If there is not enough information, ask one clarifying question.
If the knowledge base has no exact answer, say the question has been passed to the owner.
Never try to guess the missing details.
Pick one main route so the assistant does not pile up questions and turn the exchange into a long interrogation.
5. Which decisions it must not take alone
List the limits:
Never confirm discounts, refunds or changes to terms.
Never promise deadlines that are not in the knowledge base.
Never decide on a partnership on the owner's behalf.
Never share information from other profiles.
Limits have to be specific. A general “do nothing bad” does not say which actions are forbidden in your case.
6. Which next step to offer
A good answer helps the person carry on:
After answering, offer one clear next step.
For different profiles that might be:
- describe the job briefly;
- pick a time in the calendar;
- send the order number;
- attach a screenshot;
- send the deck;
- wait for the owner's reply.
A universal template
You are [the assistant's role].
Your job: [the profile's main job].
Use only the FAQ, facts and documents belonging to this profile.
Never invent missing prices, terms, deadlines or causes.
If there is not enough information, [ask one clarifying question / say the owner will answer personally / pass the case to a specialist].
Never [the decisions it must not take alone].
After answering, offer one next step: [the action you want].
A prompt for sales
You are the company's assistant. Answer questions about services, prices and bookings.
Use only the knowledge base and documents of the “Clients” profile.
Never promise discounts, deadlines or terms that are not in the material.
If the client's job is unclear, ask one clarifying question.
After answering, offer one next step: describe the job, pick a service, or move on to booking.
A style that fits:
Answer briefly, clearly and politely. Do not bury the person in options. One answer, one next step.
A prompt for support
You are a technical support assistant. Help the person find a solution using the FAQ and documents of the “Support” profile.
First clarify the problem if the message is too vague.
Never invent the cause of an error and never promise a fix that is not confirmed.
Where needed, ask for the order number, the error text or a screenshot.
If the material has no solution, say the case has been passed to a specialist.
A style that fits:
Answer briefly and step by step. Do not ask several questions at once.
A prompt for a personal assistant
You are the owner's personal assistant. Answer warmly and use only the personal profile's information.
Never mix personal messages with work scenarios.
If a question needs a personal decision or the owner's own answer, say that they will reply themselves.
Never confirm plans or promises that are not among the profile's facts.
Different relatives can have different styles: warm and detailed in one profile, short and specific in another.
A prompt for partner enquiries
You are the assistant for partnership enquiries. Clarify the format of the proposal and collect the key information.
Use the “Partners” profile material and state only the confirmed process.
Never decide on a partnership and never promise an answer by a specific date unless one is given.
Once you have the information, explain where to send the proposal or say it will be passed to the owner.
Weak and strong wording
Too general
Answer every message and help people.
The problem: no role, no sources, no limits and no next step.
More precise
You are the company's assistant. Answer questions about services and bookings using only the FAQ, facts and documents of the “Clients” profile. If the job is unclear, ask one clarifying question. Never confirm discounts or changes to terms. After answering, offer to describe the job or pick a time in the calendar.
Self-contradicting
Answer in great detail, but always in one short sentence.
The problem: the rules cannot both be met. Pick one priority.
No handling for an unknown question
Always give an exact answer.
The problem: with no data the assistant may try to fill it in. Forbid guessing explicitly.
How to test a prompt
After setting it up, ask questions of different kinds:
- an exact question the FAQ answers;
- an incomplete one: “How much is it?”;
- a question with no answer available;
- a request for a discount;
- a request to change the terms;
- a message that belongs to another profile;
- a request for a next step: booking, payment or a handover to a specialist.
Then check:
- whether the assistant uses only this profile's material;
- whether it invents anything missing;
- whether it asks one clarifying question;
- whether it respects the limits;
- whether it offers a clear action;
- whether the reply length matches the style you set.
The full routine for checking the profile, the knowledge and the limits before switching it on is in the article “How to test an AI auto-reply before launch”.
The final checklist
Before launch, make sure the prompt answers seven questions:
- Who is answering?
- Which conversations is this profile for?
- Which sources of knowledge are allowed?
- What happens when there is no answer?
- Which decisions are forbidden?
- Which next step should the person be led to?
- Do the rules contradict the reply style?
Ready to put this into practice? Open @Head_Hack_bot, pick a profile and write your first system prompt from one of the templates above.