That's hard for humans too. Perhaps if one tries to interspers "internal monologue" with the temperature cranked up, with the main output, and prompting the LLM

That's hard for humans too. Perhaps if one tries to interspers "internal monologue" with the temperature cranked up, with the main output, and prompting the LLM to consider and weave in aspects from the inner monologue when continuing while taking care to make it coherent... (I'm sure it'd be harder than that - but basically inject random internal "thoughts" without letting them control the output)