A model writes exactly what it was asked for. Comments that read as filler are nearly always the product of a thin instruction rather than of the model's ceiling.
What the brief must contain
Who is speaking
Not "an expert" — a specific person with a specific occupation and a reason to be reading this channel. Vagueness here produces the generic voice everyone recognises.
What counts as a good comment
Add a fact, an experience, a correction or a question that moves the thread. State it explicitly, because the default behaviour is agreeable summary.
What is banned
Praise without content, restating the post, mentioning your product, links, greetings. The prohibition list does more work than the instruction list — why moderators delete comments.
Length
Two or three sentences. Without a limit the model writes an essay, and long comments under short posts read as an attempt to hijack attention.
Language and register
Match the post. A formal comment under a casual post is as conspicuous as the reverse.
Variation is a requirement, not a nicety
Identical constructions across accounts are the clearest automation signal there is. Ask for different openings and forbid recurring phrases explicitly — the account fingerprint.
Test before scale
Generate thirty comments for real posts and read them as a moderator would. Anything you would delete, the moderator will — commenting without a ban.
What the brief cannot fix
A wrong channel list. A perfect comment in a channel your audience does not read returns nothing — how many channels you need.
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.