Maximum length of each generated chunk. Higher = longer passages but costs more and risks hitting the model's context limit.
256–4096 for most models. 2048 is a good default.

Controls how creative vs. predictable the output is. Lower = more focused and literal. Higher = more varied and imaginative.
0.6–0.8 for structured prose. 0.9–1.05 for creative flair.
Above ~1.1 the model can pick genuinely junk tokens. Because it then writes based on that token, one slip degrades everything after it — which is why a long chunk can start well and collapse into gibberish. Raise Top-P instead.

Restricts sampling to the most likely tokens making up this share of the probability mass, discarding the junk tail.
0.95 is a good default. This is what makes a higher Temperature safe. At 1 nothing is filtered.

How much of the story history to keep in memory when generating. Older segments beyond this limit get summarised automatically.
Match your model's context limit. 8000 suits most; use 16000+ for larger models.