A language model knows how to write and knows nothing about what you sell. Everything it says to a customer comes from your instructions — and from invention wherever those instructions stop. The second is where the problems are.
Start from real conversations
Not from a product description. Take twenty actual enquiries and see what people asked, in their words. That list is the specification; the marketing copy is not.
Four things the brief must contain
What you sell, in the customer's language
Not internal product names. What the person gets and what problem it solves.
Prices, or how they are formed
Fixed price, or the parameters it depends on. Without this the model either avoids the subject or invents a number.
Boundaries
What you do not do. This matters more than it sounds: with no explicit list the model agrees to everything, and you inherit the promise.
What to do when it does not know
An explicit instruction: do not guess, say you will check, call a person — when to hand over.
Tone
State it, or you get neutral corporate English, which reads as foreign in a Telegram conversation. Two or three examples of how you would answer are worth more than adjectives.
Testing before launch
Ask your own configuration twenty questions including the uncomfortable ones: a discount, a comparison with a competitor, something you do not offer. The gaps show up immediately.
Then read the first few dozen real conversations end to end. That is the only way to catch phrasing that quietly damages deals — what not to trust it with and the knowledge base.
How Neurogram does this
Everything above is manual work that runs into the number of hours in a day. Neurogram does it on your own accounts and without you: it collects channels and audiences, writes comments and replies with a language model, and keeps limits and quiet hours so the accounts survive. See pricing or look at your case.