Introduction

The plugin architecture of DeepSeek Harness (DSH) allows extension of its Web interface. dsh-prompt-boost is designed to address prompt input quality issues. It adds a ✨ button next to the model selector, rewriting raw drafts into structured instructions.

Core Features

  • ✨ Input optimization: Rewrite the current draft into a clear, structured, and directly executable prompt with one click.
  • ↩ Undo: After optimization, the button changes to a ↩ icon. Clicking it restores the original input and supports iterative optimization.
  • Fidelity first: Preserves concrete details in the input (paths, numbers, names, commands) and does not fabricate information; missing critical information is marked as placeholders.
  • Task awareness: Automatically detects task types (coding, data processing, writing, analysis, etc.) and organizes the content accordingly.
  • Natural output: Uses Trae-style natural instruction text (such as numbered lists); does not use JSON or template headers.
  • Language locked: The output language always follows the input language (Chinese in, Chinese out; English in, English out).
  • Overwrite protection: If the user modifies the draft during optimization, the result will not be written back, and the undo function will not be activated.

Installation

Installation requires manually adding a configuration patch. Insert the following content into the Web user’s configuration file (for example, ~/.dsh/profiles/web/cordis.patch.yml):

- insert:
    - id: prompt-boost
      name: 'dsh-prompt-boost'

Ensure the plugin package is available through node_modules (installed via npm or linked), then reload the DSH host and refresh the Web GUI.

Usage

  1. Enter a rough idea in the conversation editor, for example: “Help me write a script to organize the Downloads folder; it’s too messy.”
  2. Click the ✨ button (a loading animation is shown during optimization).
  3. Review the optimized instruction in the editor and send it after confirming it is correct.
  4. If you are not satisfied with the result, click ↩ to restore the original input and try a different wording.

Technical Details and Limitations

  • Zero build: The plugin is pure JavaScript (ESM host + __ModuleLoader__ Web bundling).
  • Input limits: Maximum supported input is 8,000 characters, with a timeout of 90 seconds.
  • Environment requirements: It must run on DSH with the llm and agentDefaultModel services mounted and a default model configured. Browser Web GUI (conversation editor) support is required.
  • GitHub repository: https://github.com/060625dfy/dsh-prompt-boost
  • Plugin directory: https://www.skillhub.cn/plugins/060625dfy/dsh-prompt-boost