Sonnet 4.6 overactive refusal as conversation goes
Hey everyone,
I wanted to see if this is just a me issue or a common issue. I like creating stories and using AI as a sounding board to develop them.
That naturally leads to longer conversations, more smaller prompts rather than one big prompt, since I start not knowing exactly what I'm even going to build yet.
And often (almost always), by prompt 25, Claude will start saying things like "I've drifted" (he absolutely has not), "I can't continue," or "this is against TOS" (writing a fictional story about a married couple, non-sexual, absolutely isn't).
It doesn't seem to matter what the story, style, or characters are, just after 25 prompts, even when I'm not working on a story, even for something like a website, Claude Sonnet 4.6 is almost guaranteed to hallucinate that it has somehow drifted and that it can't continue. It's been frustrating, though Haiku seems to have no such issue.
It's almost reminiscent of how ChatGPT currently hallucinates guardrails that simply don't exist.
Have you guys experienced the same issue? Or is it just me?