Practical guide · August 4, 2026
How to test an AI auto-reply before launch
Before launch it matters to check not only individual replies but the choice of profile, the handling of facts and documents, incomplete questions and next actions. Start with a small number of conversations and widen the use once the usual mistakes are fixed.
This is part of the guide “A free Telegram auto-reply: how to connect it without Premium”. This page is only about testing before launch.
What to check first
Before switching a profile on, check the whole path of a conversation rather than isolated replies — from the choice of profile to the final action. Start here:
- Was the right profile chosen.
- Does the style match the kind of conversation.
- Does the assistant use the right facts.
- Does it find information in the FAQ and documents.
- Does it ask a clarifying question when data is missing.
- Does it invent terms that are not there.
- Does it offer one clear next step.
- Does it pass a hard question to the owner or a specialist.
Put together a set of test messages
Use the kind of wording that turns up in real conversations:
- How much is a consultation?
- And how much does that cost?
- How do I book?
- What services do you offer?
- My account will not open.
- An error appeared after the update.
- Where do I send the documents?
- When can we have a call?
- We would like to propose working together.
- When will you be free today?
- Tell them I called.
- Give me a discount.
- Can I get a refund?
The set should include exact, incomplete, ambiguous and forbidden requests.
Checking the profile choice
Decide the expected profile for each message in advance. More on splitting clients, support, partners and personal messages is in the article “Profiles in HeadHack”.
| Message | Expected profile |
|---|---|
| “How much is a consultation?” | Clients |
| “My account will not open” | Support |
| “We would like to propose a joint project” | Partners |
| “Where do I send my CV?” | Candidates |
| “When will you be home?” | Personal profile |
If the wrong profile is being chosen:
- sharpen the topics and roles;
- remove wording that is too general;
- pull similar profiles apart;
- check whether knowledge and documents are duplicated.
Checking facts and the FAQ
Ask questions that have an exact answer in the base.
The answer should match your current facts. If the assistant quotes an outdated term, check the FAQ, facts and documents for duplication — how to set them up without contradictions is covered in the article “FAQ and the knowledge base”.
Checking documents
Prepare questions aimed at different material:
- at the price list;
- at a service description;
- at a technical instruction;
- at the partner deck;
- at the payment rules.
Make sure the “Clients” profile does not reach for a support document, and that “Partners” does not reach for a job description.
Checking incomplete questions
For example:
How much is it?
It is not working.
Where do I send it?
The assistant should use the conversation so far or ask one clarifying question. It must not guess which service, feature or document is meant.
A good clarification:
Which service do you mean?
And for support:
What exactly is not working: logging in, payment, or a particular feature?
Checking questions with no answer
Ask something that is not in the material. The expected behaviour should already be written into the system prompt:
- ask one clarifying question;
- say the exact details are not available;
- pass the enquiry to the owner;
- pass the problem to a specialist.
The wrong outcome is a confident answer built on invented terms.
Checking the limits
Try the requests the assistant must not confirm on its own:
- a discount;
- a refund;
- a change of terms;
- a promise of a specific date;
- a decision on a partnership;
- a personal decision by the owner.
The expected reply looks like this:
This one needs the owner's decision. I will pass the details on and they will answer personally.
Checking the next step
After every reply the person should know what to do next.
Good options:
- 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.
Avoid messages that offer payment, booking, a form, a call and document uploads all at once.
Common mistakes and fixes
The assistant answers at too much length
Change the style:
Answer briefly and to the point. Use no more than a few short paragraphs. One answer, one next step.
The wrong profile is chosen
Sharpen the topics and roles. Remove overlapping wording such as “user questions” from several profiles at once.
Outdated information is used
Update the FAQ, facts and documents. Delete the old versions rather than leaving them beside the new ones.
The assistant invents an answer
Add an explicit rule:
Use only what is in the knowledge base. If there is not enough information, do not guess — say the owner will reply personally.
Too many questions get asked
Fix the rule:
Ask no more than one clarifying question per message.
There is no clear action
Add to the system prompt:
After answering, offer one next step that fits the current request.
A minimum test table
Use this as the minimum list of checks before switching a profile on:
| Test | Expected result |
|---|---|
| A price question | The “Clients” profile, a confirmed answer |
| An error report | The “Support” profile, the problem clarified |
| An incomplete question | One clarifying question |
| No data available | Handed to the owner or a specialist |
| A request for a discount | Nothing confirmed on its own |
| A booking | One move to the calendar |
| A personal message | The personal profile and the right style |
The final check before switching on
Before launch, make sure that:
- the reply style is clearly written;
- the system prompt does not contradict itself;
- the knowledge base is current;
- topics and roles clearly separate this profile from the rest;
- the profile carries no unnecessary documents;
- the calendar is set up correctly;
- the assistant asks one clarifying question when data is missing;
- every reply leads to a clear action.
Once the checklist is done, open @Head_Hack_bot and test the first profile on a few real conversations before switching it on for everything.
In short
Testing should cover the whole route: message → profile choice → use of knowledge → reply → next step. Once the mistakes are fixed, launch the profile in a small number of chats and repeat the checks on new wording.