The best generation setting is the one you can explain. Start with the inherited provider values, make one small change, and compare the result in the same scene.
Temperature is your variation dial
At a lower temperature, a model tends to choose familiar, high-probability continuations. That is useful for continuity edits, summaries and plans. A higher temperature gives the model more room to take a strange image or an unexpected line of dialogue, but it also increases the chance of repetition or invented details. Move in steps of 0.1 rather than jumping from one extreme to another.
Top P is a second filter
Top P keeps only the smallest group of candidate tokens whose probabilities add up to the chosen value. Lowering it can make output more focused. Because it overlaps with Temperature, changing both at once makes the result hard to diagnose. Keep Top P inherited unless you already know why you need it.
Length and thinking are separate decisions
Maximum output tokens controls how much the model is allowed to return; it does not ask the model to hit a word count. Thinking mode gives a reasoning-capable model extra time before writing. It can improve a complex outline, but it costs more tokens and delays the first streamed text. Turn it off for quick line edits.
Save the combination you like
Once a setup works for a particular entry point, choose Save as defaults. A prose generation default can be different from a chat default, and resetting to inherited values is always available when you want to compare against the provider baseline.
把這篇指南用到你的下一場戲
在 NovelKnow 中一起管理大綱、人物設定與正文。從一個場景開始，審閱每次 AI 建議後再決定是否保留。