Prompt Management

View and maintain product-level AI task prompt assets, with safe-slot editing, advanced template debugging, and version rollback.

Prompt Management is used to view and maintain the prompt assets used by product-level AI tasks. It is suited for developers, debuggers, and advanced users who need to understand task behavior. Ordinary writing users do not need to visit this page frequently; the main creative chain should be advanced through Auto Director, the novel workbench, and the chapter workbench.

Prompt Management only affects prompts themselves, not model vendor configuration, API keys, or model routing. For model-related settings, please go to System Settings.

Safe-Slot Editing

Modify local rules in a visual interface without changing the overall prompt structure.

Advanced Template Debugging

Reference context and slots through Chinese labels, with source-code view switching.

Preview and Test

Generate previews to inspect messages and context, and test output to verify real model effects.

Version Rollback

Save book-level override versions, and restore official templates or roll back historical versions at any time.

When to Open It

Prompt Management is not a daily creative entry point. Only enter in the following scenarios:

• You want to understand how a certain AI task organizes its input.
• A task’s output is persistently unstable and you need to inspect the prompt version.
• You need to confirm the binding relationship between the prompt and model routing.
• A developer is preparing to adjust structured output or task contracts.

Ordinary users should prioritize adjusting novel materials, the knowledge base, the Style Engine, and model routing. Only when a task type has been performing abnormally for a long time and both data and model have been checked should you consider the prompt.

Interface Structure

The Prompt Management page is divided into three main areas:

Area	Function
Left prompt directory	Lists prompt assets by task type; for example, prose generation corresponds to novel.chapter.writer
Middle editing area	View and modify system / human templates, generate previews, and test current draft output
Right context panel	View what materials are injected into this preview, such as book-level contract, chapter brief, character hard facts, timeline, style contract

The bottom action area is used for generating previews, testing output, saving book-level overrides, resetting modifications, or restoring official templates.

Prompt Categories

Prompts are usually organized by task family, with each family corresponding to a clear task:

Task Family	Typical Use
Book starting direction	Auto Director generates new book ideas and directions
World and character preparation	Worldview and character setting generation
Volume planning and chapter scheduling	Volume structure and chapter brief arrangement
Prose generation	Chapter prose writing, corresponding to novel.chapter.writer
Review and repair	Prose review and issue fixing
Book breakdown analysis	Reference work structure breakdown
Knowledge base summary	Material summary and retrieval
Style extraction	Extract style rules from sample text

Each prompt category should correspond to a clear task, rather than piling large amounts of unrelated instructions together.

Editing Prose Generation Prompts

Prose generation prompts are the most frequently debugged prompts. Entry method:

1. Open Prompt Management

   Select novel.chapter.writer in the left directory.

2. Select scope

   Select the book scope, and specify the novel and chapter to preview.

3. Switch editing mode

   In the middle area, switch between Safe Slot (安全槽位) or Advanced Template (高级模板).

4. Generate preview

   Click Generate preview (生成预览) to check the final messages and right-side context injection results.

5. Test output

   When you need to verify effects, click Test output (测试产出), select a test model in the popup, and start testing.

Safe Slot

Safe slots are suitable for modifying local rules, such as tone, rhythm, Anti-AI rules, or supplementary writing requirements. They do not change the overall prompt structure and suit most debugging scenarios.

Advanced Template

Advanced templates are suitable for mature users to fully adjust the system / human templates of novel.chapter.writer. They only affect prose generation for the current book and do not open schema, context policy, post-processing validation, or free replacement of other prompts.

Visual Reference Tags

Advanced templates use visual reference tags by default. In the editor you see labels like Book-level contract (书级合约), Chapter brief (章节任务), Character hard facts (角色硬事实), Timeline (时间线), Chapter title (章节标题), and Tone and rhythm (语气与节奏), rather than requiring users to memorize {{context.book_contract}}, {{input.chapterTitle}}, or {{slot.tone_and_pacing}}.

How to Use

• In the editor, type @ and select context, runtime variables, or slots from the reference menu.
• After selection, a tag is inserted; when saving and previewing, the system still compiles it into underlying template tokens.
• Hover over a tag to view its key, original token, description, and whether it is required context.
• Delete a tag and re-insert with @, suitable for replacing incorrect references.
• If the template contains unrecognized tokens, the editor preserves the original content and shows an abnormal status, making it easy to troubleshoot without silently losing text.

Common Tag Meanings

Type	Tag Example	Purpose
Context	Book-level contract, Chapter brief, Timeline, Character hard facts	Inject current book and chapter materials into prose generation
Runtime variables	Chapter title, Target word count, Chapter sequence number	Use parameters passed in during this runtime
Slots	Tone and rhythm, Anti-AI rules, Custom supplementary rules	Reuse editable rules from safe slots

Source Code View

Advanced templates retain a source code view, suitable for debugging underlying templates:

• The visual view is for daily editing and prioritizes displaying Chinese labels.
• The source code view displays real template text, such as {{context.chapter_mission}}.
• After switching from source code view back to visual view, registered references are re-displayed as tags.

If you are only adjusting writing effects, prioritize the visual view and the @ reference menu. Only switch to source code view when troubleshooting tokens, copying templates, or confirming compilation results.

Context Injection Panel

The right-side context panel is used to confirm what materials the model will actually see:

Display Item	Description
Context block name and grouping	Clarifies material source
Whether it is required context	Locked by system governance rules, cannot be disabled in the interface
Token estimate	Helps judge context length
Current status	Selected, trimmed, summarized, or missing
Preview content	Checks whether materials come from the current novel and current chapter

Common required blocks for prose generation include book-level contract, chapter brief, character hard facts, obligation contract, style contract, and timeline. This prevents users from accidentally deleting key materials, causing chapters to deviate from the main plot or violate prior state.

Preview, Test, and Save

We recommend following this sequence after each modification:

1. Generate preview

   Check whether system / human messages and context injection are complete.

2. Confirm context

   View the right-side context and confirm materials come from the target novel and target chapter.

3. Test output

   In the popup, select a test model and view the return under the current draft on the real model.

4. Save override

   If it is only a local adjustment, prioritize saving as a book-level override.

5. Version management

   When an advanced template is broken, you can roll back to a historical version or restore the official template.

Difference Between Generate Preview and Test Output

Action	Purpose	Calls Model?
Generate preview	View final messages, context injection, missing materials, and diagnostic info	No
Test output	Use the current draft and selected model to generate one result, judging the effect of changes	Yes

Test output does not require saving the draft first. Structured prompts display parsed JSON and repair counts; text prompts display model-returned prose. If the test is unsatisfactory, you can continue adjusting before saving.

Saving an advanced template generates a book-level version history. Restoring the official template does not delete historical versions; you can still view and roll back later.

Prose Effect Lab

The Prose Effect Lab is used to compare “what the writing will look like after template modification.” The entry is in the chapter editor and the Style Engine. It automatically locates the prose generation prompt and brings in the current novel and chapter.

Recommended workflow:

1. Open the lab

   Open the Prose Effect Lab (正文效果实验室) from the chapter editor or the Style Engine.

2. Confirm scope

   Select the target novel and chapter, and confirm the right-side context comes from the correct project.

3. Adjust draft

   Adjust the current draft in safe slots or advanced templates.

4. Start trial writing

   Use the real model to generate comparison prose, and continue adjusting based on the streaming display results.

5. Save version

   Save the book template version only when satisfied.

Trial writing uses the current unsaved template, does not overwrite chapter prose, and does not modify chapter status. It calls the real model and consumes quota; if you only want to check final messages and context, use Generate preview.

Action	Suitable For	Overwrites Prose?
Generate preview	Check whether template, references, and context are complete	No
Start trial writing	Compare reading feel after template modification	No
Save as new version	Make subsequent book prose generation use this template	Only affects subsequent generation

Usage Boundaries

Prompt Management does not bypass system governance rules:

• Cannot disable required context.
• Cannot modify contextPolicy in the page.
• Cannot modify structured output schema.
• Cannot disable validation, repair, or state synchronization after prose generation.
• Advanced templates first-priority override prose generation novel.chapter.writer; other prompts still focus on safe slots and previews.
• Test output only executes registered prompts and the current draft, and does not provide an arbitrary system prompt bypass entry.

If the issue stems from missing novel materials, incomplete chapter briefs, insufficient model capability, or unstable structured output, simply changing the prompt may not solve it. You should first return to the corresponding module to complete materials or inspect task failure causes.

Confirm Before Editing Prompts

Before modifying prompts, we recommend confirming the following:

• Whether the current issue truly comes from the prompt.
• Whether model routing is appropriate.
• Whether input data is complete.
• Whether output schema is being followed.
• Whether failure can be resolved through retry or switching models.

The prompt is not the first fix for all problems. Structured output, model capability, context materials, and task status can all affect results.

Versions and Rollback

Prompt adjustments affect subsequent AI tasks. We recommend maintaining version awareness:

• Clearly record the purpose of adjustments.
• Only modify prompts related to the target task.
• Validate results with a test novel.
• Keep revertible versions.
• Avoid changing model, data, and prompt simultaneously, which makes problem sources hard to identify.

Public product-level prompts should go through unified registration and schema management; we do not recommend temporary inline embedding in business code.

Binding to Tasks

The value of Prompt Management lies in understanding “which task uses which prompt.” When troubleshooting, you can follow this path:

1. Confirm task type

   Confirm the failed task type in the Task Center.

2. Check model routing

   Confirm whether the model used by the current task is appropriate.

3. View prompt

   In Prompt Management, confirm the task prompt and output requirements.

4. Test and verify

   Return to the test novel to verify modification effects.

If the task requires structured output, the prompt and output schema must be considered together.

System Settings

Configure model vendors, model routing, and Auto Director global preferences.

Model Configuration

First-time quick setup, third-party vendor access, and model routing details.

Style Engine

Extract style rules from sample text, prioritizing writing effect adjustments.