Introduction

DeepSeek Harness (DSH) is a platform that supports plugin-based extensions. dsh-novel-craft is an open-source plugin that runs on DSH and focuses on solving the problem of style transfer in AI novel writing. It does not rely on prompt engineering or model fine-tuning; instead, it records users’ reading preferences and enables the model to distill and reuse writing patterns.

Installation and Activation

The plugin depends on the DeepSeek Harness environment. The installation command is as follows:

dsh plugin --profile web add dsh-novel-craft

After installation, restart the dsh service for the plugin to take effect.

Core Features

The plugin provides the following main capabilities:

  • Preference Calibration: Use upvotes and downvotes to help the model learn the author’s style preferences.
  • Candidate Draft Generation: Generate multiple versions of candidate chapters using the preference profile.
  • Annotation Fine-Tuning: Support adding annotations to specific paragraphs and fine-tuning them.
  • Plot Checkup: Generate views for plot tension and foreshadowing management through local statistics.
  • Local Pattern Storage: The preference profile is stored locally only and does not include source text.

Usage Example

In the DSH interface, users can write using the following operations:

  1. Reading and Marking: In candidate drafts, use j/k to move between paragraphs, G to mark good paragraphs, and B to mark bad paragraphs.
  2. Distilling Patterns: Click “⚗️ Distill Patterns” to summarize the marked content into a preference profile.
  3. Generating a New Chapter: Use the “✍️ Start New Chapter” feature together with the preference profile to generate candidate drafts.
  4. Draw Cards and Select Paragraphs: Use the “🎴 Draw Cards and Select Paragraphs” flow to choose paragraphs you are satisfied with.
  5. Annotation Fine-Tuning: Select a paragraph and press A to add an annotation, then use the “Fine-Tune by Annotation” feature to modify it.
  6. Plot Analysis: On the “📈 Plot” page, view the tension curve and global plot management.

Notes

  • The plugin depends on DeepSeek Harness; ensure the environment is correctly configured.
  • During writing, the main text does not enter the model context; only distilled patterns enter, to maintain lightweight operation.
  • Platform approval is not guaranteed; review your published content yourself.
  • The interface is in Chinese.

References

  • Project address: https://github.com/shiyan688/dsh-novel-craft
  • Directory page: https://www.skillhub.cn/plugins/shiyan688/dsh-novel-craft