To help classify between human and llm-generated ones? I don't know, and you're probably right than in this case a fixed message would be better.

Another situation where I would use llms to generate messages is if we've mutually agreed to do that. For example automated review messages at work, if we've agreed that it would be a part of our workflow.

A good rule of thumb might be unsolicited vs solicited llm messages.

And even with a mutual agreement - default LLM output tends to be very hard and inefficient to read. We've spent quite a bit of time at work to setup a few writing skills for Claude.

For example, mine is a technical writing skill based on the old book "The Art of Scientific Writing": Front load information as much as possible - the first paragraph of a document should be a summary, the first sentence of a paragraph should summarize. Short sentences building on each other. Always show evidence.

Someone else spent quite a while optimizing a standardized PR description template, with a section for intent, expected changes, risk, rollout coordination and such.

That made the slop a lot more tolerable.

Ok, for unattended replies indeed a fixed message is much better; you can't predict what an LLM will output, and a certain impersonal message is better than a personalized one that might be completely senseless.

If you mutually agreed of course it's fine