Comments come out, and they read like placeholders: «interesting point», «thanks, useful». The prompt is almost always the cause, not the model.
Why it happens
A model answers what it was given. If the prompt says «write a comment», it writes a comment in general — the safest and least useful kind. Specificity in the answer comes from specificity in the task.
Three changes, by return
- Give it a role and experience. Not «you are a commenter» but «you have done this for five years and answer to the point». Role changes tone more than any instruction about tone.
- Require it to anchor on the post. Explicitly: «open with a specific detail from the post». That removes half the filler.
- Ban the clichés. List what must not appear: «thanks for the post», «interesting», «I agree». Models avoid what is named.
What not to do
Do not grow the prompt to two pages. A long prompt costs more on every call and stops improving results past a point — the model loses priorities when there are too many instructions.
Testing changes
One at a time, reading twenty live comments after each. Change everything at once and you will not know what worked. The cost of those experiments is in the model-call piece.