IMO this is because responding within a json response with all those other fields shifts the distribution to text that's staler.

If you told it to roleplay a pixel knight fighting to the death and provided it responses like "Authur (renamed claude sonnet 5) moves west towards you menacingly", the whole thing would come alive.

I've been working on an AI roleplaying game for a while now and it had felt stale until I separated the story output from the plumbing output (importantly, I do story first and then base the plumbing on that). I think the two just have incompatible distributions.

I've been working on AI roleplaying too on and off, but after...

"Authur moves west towards you menacingly"

...RP is imemediately broken in the reasoning chain, and it's in that mode the prose is generated, e.g.

"Alright. Arthur moved west towards me menacingly. As Merlin, I should do X. No, I should do Y. Hmm. Maybe I should do Z. Let me reconsider. OK, I'll do X. How would Merlin respond? Maybe 'I'm a cool wizard'. No. 'I'm Merlin the wizard!' Yes, that's the load-bearing literary masteripiece. Let me respond with that!"

"I'm Merlin the wizard!"