Preface

In the usage scenarios of DSH, prompts are often not one-time text. You might want the model to remember constraints it habitually omits, common expressions, and which phrases are prone to causing misunderstandings. If you previously used -fin to manually trigger learning, Hua1Q1nG/dsh-prompt-self will integrate daily learning into an automatic process: rewriting the prompt according to the persona before sending a new request, continuing to learn the persona after the round ends, and providing visual toggles and access points to view the persona.

What is this

Hua1Q1nG/dsh-prompt-self is a dual-sided client plugin for DeepSeek Harness (DSH). It is maintained by Hua1Q1nG and licensed under MIT.

This plugin does not modify the DSH application code; it is installed via web profile patches, user presets, skill files, and AGENTS.md.

Please note: It has a dual-sided structure; the web profile side and the Agent layer engine need to be installed/enabled separately. Running only the marketplace installation will set up the host layer Web routes, settings page, and conversation input dock; installing user presets is also required to enable message-level rewriting and automatic learning.

Core Features

Below are the capabilities of the plugin.

  • Message-level automatic rewriting: Before sending a new request, the LLM rewrites the prompt based on the personal persona, then injects the optimized version into the model for execution.
  • Automatic persona learning: After a round ends, analyzes the original request and the final answer, distills writing habits, anti-hallucination rules, and optimization experience, and deposits them into the persona file.
  • Retain -fin: Can still serve as a compatible trigger for “immediate learning and confirmation”.
  • Visual control: The “Prompt Persona” section in the settings page provides two toggles, “Auto Rewrite Optimization” and “Auto Learn Persona”, which save instantly and take effect in real-time.
  • Persona 3-column view: Habit list, anti-hallucination rules, learning records (containing count, update time, and collapsible state).
  • Quick access in conversation area: A persistent status capsule above the input box; clicking it pops up a floating panel to view/switch.
  • Zero intrusion: Does not modify DSH application code.

Installation and Enablement

First perform the host-side installation, then enable the Agent layer. The plugin runs with the current DSH process permissions; you should check the source code and the MIT license before installing.

  1. Marketplace Installation
dsh plugin --profile web add github:Hua1Q1nG/dsh-prompt-self

This step installs the host layer Web routes, settings page, and conversation input dock.

  1. Install User Presets

Install the user presets and switch the default preset to code-prompt-self. This step is used to enable message-level rewriting and automatic learning.

  1. Register web profile patch layer plugin row

Register the plugin row in the web profile patch layer to allow the web profile side to load the plugin.

  1. Optional installation of skill and global commands

Optionally install skill and AGENTS.md.

  1. Restart and Verify

After restarting DSH Desktop, check the toggles and persona in Settings -> “Prompt Persona”, and view the status capsule above the input box. After completing the steps above, send a new message and wait for the round to end to view the learning records on the persona page.

The peerDependencies declared in package.json are as follows:

{
  "peerDependencies": {
    "@deepseek-ai/cordis": "^4.0.1",
    "@deepseek-ai/dsh-llm": ">=0.0.1-rc.1 <0.1.0 || >=0.1.0-rc.1 <0.2.0-0",
    "@deepseek-ai/dsh-scope": ">=0.0.1-rc.1 <0.1.0 || >=0.1.0-rc.1 <0.2.0-0",
    "@deepseek-ai/dsh-host-webserver": ">=0.0.1-rc.1 <0.1.0 || >=0.1.0-rc.1 <0.2.0-0"
  }
}

Typical Usage

Below are several daily entry points.

Automatic rewriting occurs after you normally send the prompt: the plugin rewrites it based on the personal persona and injects the optimized version for model execution.

Automatic learning occurs after a round ends: the plugin analyzes the original request and the final answer, and deposits writing habits, anti-hallucination rules, and optimization experience into the persona file.

When viewing the persona, say in the conversation:

Check my persona

The model reads and displays the persona via the skill.

Input:

-fin

Allows immediate learning of the most recent interaction with a brief confirmation.

Applicable Scenarios and Notes

Suitable for developers who need to consolidate personal prompt preferences and reduce repetitive constraint descriptions within DSH Desktop.

Notes when using:

  • When the auxiliary model call fails, the engine will silently pass through the original request without blocking the conversation.
  • Rewriting is skipped when the persona is empty (no habits/rules/records).
  • If you run the pnpm install operation in the plugin management, the dsh-prompt-self-client directory without registered dependencies might be cleaned up, requiring you to copy the plugin package again.
  • Check the source code, dependency declarations, and license before installation.

Conclusion

dsh-prompt-self breaks down personal prompt preferences into message-level rewriting, post-round learning, and visual toggles, suitable for continuously accumulating your own expression rules within DSH.

GitHub Repository: https://github.com/Hua1Q1nG/dsh-prompt-self