Introduction

DeepSeek Harness (DSH) turns extension capabilities into plugins, and the community directory is a standalone site, not an official app store. Daily interaction with agents involves inputs beyond just the keyboard: long prompts, common templates, local documents, and spoken content all require smoother input paths.

dsh-better-input is an input enhancement plugin for DSH Web UI, maintained by DIAG5 and licensed under the MIT License. Below is an introduction to what it does, how to install it, and several typical use cases.

What is this

dsh-better-input provides capabilities such as voice input, AI polishing, prompt optimization, prompt templates, local file input, and file-to-Markdown conversion.

It can reuse DSH’s configured models, and some functions do not require an extra API Key; prompt templates are saved locally and do not pass through the server. The repository address is:

https://github.com/DIAG5/dsh-better-input

Core Features

Voice Input

After clicking the microphone button, you can transcribe while speaking, and the recognized text will stream into the input box in real-time.

Voice recognition is performed locally in the browser using the Web Speech API; no API Key or server round-trip is required. If Firefox/Safari is not supported, the microphone button will be automatically disabled. The recording duration limit can be customized from 1 to 600 seconds.

AI Polishing

AI polishing is used for automatic cleanup after speech recognition, including removing filler words, fixing homophone typos, adding punctuation, and converting spoken lists into numbered lists.

This feature reuses the DSH configured model and requires no extra Key. If you manually modify the draft during polishing, the result will not overwrite the edit; if it fails, the original text will be retained.

Prompt Optimization

Prompt optimization is used to organize already written prompts to make them more precise.

After writing a prompt in the input box, click the optimization icon in the top-right corner of the input box to pop up a comparison page between the original text and the optimized result. After confirming you are satisfied, you can adopt it, or cancel and keep the original text. This feature reuses the DSH configured model and requires no extra Key.

Prompt Templates

Common prompts can be saved as templates; later, type / in the input box to search and insert the body of the template with one click.

Templates are saved locally on the host:

~/.dsh/better-input/templates.json

Template data is not processed by the server and is not uploaded. There is a limit of 200 templates and 8,000 characters for the body.

File Input and File to Markdown

dsh-better-input supports bringing local files into the input box.

Plain text files can be sent directly. Document files can be converted first and then inserted as a reference using the format @filename. Supported document formats include PDF, DOCX, XLSX, PPTX, HTML, EPUB, CSV, JSON, XML, etc. When sending, the file reference will expand into the converted Markdown body.

OCR Visual Recognition

The usage instructions in the README mention OCR visual recognition: for scanned PDFs, image-based PDFs, or image-based PPTs, you can hand them over to a visual model page by page or by extracting images to convert to Markdown.

It should be noted that this item is not listed separately in the implemented capability list, so it is recommended to refer to the current documentation in the repository before enabling it in practice.

Update Check

The “About and Updates” entry at the bottom of the settings page provides a one-click check for the latest npm version. If a new version is found, the update command will be provided.

DSH does not automatically update third-party plugins when entering, so you need to manually execute the update command.

Visual Settings

The settings page allows you to configure recognition language, recording duration, voice polishing toggle, model, custom prompts, and other content.

If you wish to use a slider-style reasoning effort adjustment, the README mentions that you can install:

@HanaAyane/dsh-reasoning-effort

Installation and Activation

Prerequisites

Confirm that the environment meets the requirements before installation:

  • DeepSeek Harness >= 0.1.0-rc.8
  • Node.js ^22.19.0 || >=24.0.0
  • Chrome/Edge browsers

Global dsh CLI Available

If you have already installed the global dsh CLI, you can use:

dsh plugin --profile web add dsh-better-input

No Global dsh CLI

If you do not have the global CLI or do not want to write to the global environment, you can use npx to temporarily pull it:

npx -y @deepseek-ai/dsh plugin --profile web add dsh-better-input

Alternative Installation Methods

According to the repository information, alternative installation methods also include:

  • Installing from source code
  • Adding insert configuration in the agent preset’s cordis.yml

After installation, refresh the Web UI, and a microphone icon will appear on the right side of the input box.

Typical Usage

1. Voice Input

  1. Open any conversation.
  2. Click the microphone button on the right side of the input box.
  3. Start speaking; recognized text will flow into the input box in real-time.
  4. Click the button again or click Stop to end the recording.
  5. Check, modify, and send.

2. AI Polishing

  1. Go to Settings → BetterInput.
  2. Enable AI Polishing.
  3. Select a model configured in DSH.
  4. After speech recognition, the system cleans up according to the configured capabilities.

3. Prompt Optimization

  1. Write a prompt in the input box.
  2. Click the optimization icon in the top-right corner of the input box.
  3. View the comparison between the original text and the optimized result.
  4. Choose to adopt the optimized result, or cancel and keep the original text.

4. Prompt Templates

  1. Go to Settings → BetterInput → Prompt Templates.
  2. Click New Template.
  3. After saving the template, go back to the input box.
  4. Type / to search for templates.
  5. Select the candidate; the template body is inserted into the input box.

5. Adding Files and Converting to Markdown

  1. Click the add file button in the top-right corner of the input box.
  2. Select a local file.
  3. For plain text files, you can directly type @ to select and send.
  4. For document files, click Convert first.
  5. Type @ in the input box and select the file.
  6. When sending, the file reference will automatically expand into Markdown body.

6. OCR Visual Recognition

  1. Go to Settings → BetterInput.
  2. Select an OCR vision model that supports image input.
  3. Add a .pdf or .pptx file.
  4. Click Convert.
  5. Follow the prompts to choose whether to use OCR.

If an OCR model is not configured, using OCR will only prompt you to go to the settings page, without showing a red error. If the selected model explicitly declares it does not support image input, it will prompt you to change it in advance to avoid invalid results.

7. Check for Updates

  1. Go to Settings → BetterInput → About and Updates.
  2. Click Check for Updates.
  3. If a new version is found, view the update command provided.
  4. Execute the update command using the global dsh CLI or npx method, depending on your current installation method.

Applicable Scenarios and Notes

It is suitable for people who frequently input prompts in DSH Web UI, input spoken content via voice, or need to feed local documents to agents.

Please note:

  • Voice input relies on browser capabilities; if Firefox/Safari is not supported, the button will be automatically disabled.
  • Prompt templates are saved locally and do not pass through the server.
  • DSH does not automatically update third-party plugins; you need to manually execute the update command.
  • DSH has native support for image input since rc.8; the README also states that image-related plugin functions are no longer provided.
  • OCR visual recognition appears in the README usage instructions, but is not listed separately in the implemented capability list. Please verify the current repository documentation before use.
  • The plugin runs alongside the DSH client; actual permissions depend on the environment where dsh is launched. You should check the source code and license before installation.
  • License is MIT.

Links

GitHub:

https://github.com/DIAG5/dsh-better-input

Community Directory Page: The directory page URL has not been confirmed in this information, so this article does not provide that link; if needed, please refer to the actual page of the DSH Community Directory.