Introduction¶
In the daily use of DeepSeek Harness (hereinafter referred to as DSH), the input box serves both as a conversation instruction and a prompt draft. When writing prompts by hand, common issues are incomplete expression and unclear structure; when keyboard input is restricted or quick note-taking is desired, speech-to-text often lacks real-time feedback.
dsh-smart-input addresses these two problems with a plugin: voice input, and optimization of prompts within the input box.
What is this¶
dsh-smart-input is an intelligent input plugin maintained by v-quest123456, designed for the Web client scenario of DeepSeek Harness. It places speech transcription and prompt optimization near the input box, reducing the back-and-forth switching between “drafting” and “refining expression.”
The project license is MIT, and the repository address is:
https://github.com/v-quest123456/dsh-smart-input
Core Features¶
Voice Input¶
The following introduces the voice input section.
dsh-smart-input supports three recognition engines:
- Browser Web Speech API
- Local Whisper ONNX
- Whisper API (Online)
Verified capabilities include:
- Real-time streaming transcription (typing as you speak)
- Intelligent automatic punctuation
- Support for Chinese, English, and Japanese
- Continuous conversation mode: automatically continues listening after sending
- Shortcuts:
Ctrl/Cmd + Shift + V
Used to start or stop voice input.
Prompt Optimization¶
The prompt optimization section supports one-click optimization of prompts in the input box.
Available strategies include:
- Auto
- Concise
- Detailed
- Technical
- Structured
- Role-play
Verified capabilities include:
- Context-aware: can inject workspace file information
- Optimization before and after comparison preview
- One-click accept or undo
- Shortcuts:
Ctrl/Cmd + Shift + E
Used to open the prompt optimization panel.
Installation and Enablement¶
First, execute the installation command. The verified installation command is:
dsh plugin --profile web add dsh-smart-input
This command will add the dsh-smart-input plugin under the web profile.
After the above steps, you can use smart input related configurations in DSH settings.
Configuration Items¶
After installation, you can configure smart input behavior in settings. Verified configuration items include:
Recognition Mode
Local Model
Model Path
Recognition Language
Smart Punctuation
Auto Send
Continuous Conversation
Optimization Model
Default Strategy
Context-aware
Among them, Recognition Mode is used to select the speech recognition method, Local Model is used to select the local Whisper model, Model Path is used to specify the local model directory, and Recognition Language is used to set the default speech recognition language.
Smart Punctuation, Auto Send, and Continuous Conversation are used to control the post-voice input processing flow.
Optimization Model, Default Strategy, and Context-aware are used to control prompt optimization behavior.
Typical Usage¶
Voice Input¶
First, click the button in the input box toolbar for voice input and start speaking. The text will appear in the input box in real-time.
Click the button again to stop recording, and send the text after confirmation.
You can also use the shortcut:
Ctrl/Cmd + Shift + V
To start or stop voice input.
Prompt Optimization¶
First, type a prompt draft in the input box, then click the button in the input box toolbar for prompt optimization.
Select the optimization strategy and click “Start Optimization”.
After the comparison preview, choose to accept or regenerate.
You can also use the shortcut:
Ctrl/Cmd + Shift + E
To open the prompt optimization panel.
Preparing Local Whisper Models¶
If using the local Whisper ONNX mode, you need to prepare a local model directory.
In settings, set the model path to:
D:\subbatch-local-v2.0.0\models
Then select:
whisper-small
Or:
whisper-large-v3-turbo
If you don’t have a local model yet, you can download:
npx @xenova/convert --model openai/whisper-small --output ./models/whisper-small
Or:
npx @xenova/convert --model openai/whisper-large-v3-turbo --output ./models/whisper-large-v3-turbo
After the download is complete, you need to put the models into the directory corresponding to the Model Path specified in the settings.
Shortcuts¶
The verified default shortcuts are as follows:
Ctrl/Cmd + Shift + V
Used to start or stop voice input.
Ctrl/Cmd + Shift + E
Used to open the prompt optimization panel.
Shortcuts can be customized in settings.
Applicable Scenarios and Notes¶
dsh-smart-input is suitable for the following usage scenarios:
- Want to use voice to quickly enter DSH input box content
- Want to use speech recognition in an offline environment or local environment
- Want to quickly organize prompt drafts in the input box into more explicit expressions
- Want to refer to workspace file information during prompt optimization
Notes before use:
- The plugin runs with the permissions of the current dsh process
- Source code and license should be checked before installation
- Project license is
MIT - Local Whisper mode relies on a local model directory and model files
- The project structure includes:
skills/voice-enhancer/SKILL.md
This file is used for Agent skill instructions.
Conclusion¶
dsh-smart-input concentrates voice input and prompt optimization near the DSH input box, making it suitable as an auxiliary plugin for daily input and prompt organization.
GitHub Repository:
https://github.com/v-quest123456/dsh-smart-input