生成參數說明
了解溫度、Top P、輸出長度、思考模式和預設值。
The Generation settings section is available in every AI writing and planning dialog. It is collapsed until you open it, and each field can inherit the setting from the prompt or provider.
The workspace keeps the prompt, model and generation controls together so you can review them before writing.
Override one run
Open Generation settings, change the fields you need, and generate. The values apply immediately to this request. A blank field means inherit, so you can change one parameter without copying the rest of a provider profile.
What each parameter does
| Parameter | What it changes | Good starting point |
|---|---|---|
| Temperature | How varied the next token choices are. Lower values are steadier and more literal; higher values are more surprising but can drift. | 0.7–0.9 for prose; 0.2–0.5 for structured plans |
| Top P | Limits choices to a cumulative probability range from 0 to 1. Lower values narrow the candidate pool. | Leave inherited unless you have a specific reason; usually tune Top P or Temperature |
| Maximum output tokens | Caps the response length in tokens. It is not a character limit, and a value that is too low can cut off a scene or JSON response. | Use the provider default, then raise it for long scenes |
| Thinking mode | Lets supported reasoning models spend extra tokens planning before they write. It increases latency and usage, and never inserts the hidden reasoning into your manuscript. | On for difficult planning; off for quick rewrites |
These controls cannot override limits imposed by the selected model or your plan. Structured operations may turn Thinking off automatically so their output remains valid.
Save a default for this entry point
After choosing values, click Save as defaults. Defaults are stored separately for each entry point (for example, scene generation and chat) in this browser. Later dialogs start with those values, while a blank field still inherits the provider setting.
Use Reset to inherited when you want to remove the local override. The status message confirms whether the defaults were saved.
A practical tuning recipe
- Start with inherited settings and generate a short sample.
- If the prose feels repetitive, raise Temperature by 0.1; if it loses continuity, lower it by 0.1.
- Keep Top P inherited while tuning Temperature so you can tell which change helped.
- Increase Maximum output tokens only when the response is being cut short.
- Save the combination once it works for this type of task.