dsh-prompt-enhancer

A DeepSeek Harness (DSH) plugin: prompt optimization + voice input enhancement — rewrites a composer draft into something clearer and more actionable, and adds a hotkey, stop-on-silence, filler removal and correction to voice input.

A draft is often half a requirement, so the model asks again — click ✨ and it is rewritten in place, undoable. With voice, just speak; it stops on a pause and the transcript lands back in the composer.

Install

dsh plugin --profile web add github:Fishsb/dsh-prompt-enhancer#v3.4.2
dsh plugin --profile web update dsh-prompt-enhancer   # update
dsh plugin --profile web remove dsh-prompt-enhancer   # remove

Restart DSH afterwards (also after updating or removing) — the ✨ button appears in the composer toolbar. Requires DeepSeek Harness ≥ 0.1.0-rc.6 installed locally and pnpm in PATH.

What it does

Prompt optimization

• ✨ One-click optimize — rewrite the current draft; keep refining, undo, or cancel while optimizing
• 5 modes — Basic / Lite / Standard / Expert / One-click Publish
• Memory switch — pre-send rounds accumulate into a memory chain, cleared on send (off by default)
• Model chain — try models in order, add/remove/reorder, toggle thinking, test connectivity inline

Voice input enhancement (transcription is provided by DSH's official speechToText; the plugin ships no engine and downloads no models)

• Two entry points — the official 🎤 button, or the plugin hotkey (tap to start / tap to stop; long-press = hold to talk, release to stop)
• Stop on silence — auto-stop with a configurable 1-5 s silence window (default 2 s), can be turned off entirely
• Filler removal — local rule-based cleanup of filler words / repeats / drawn-out sounds; no rewriting, no additions (on by default)
• Transcript correction — word-level fixes after transcription, voice path only (off by default)
• Auto-enhance — trigger enhancement automatically once the transcript is in the draft (off by default)

Buttons and text follow the DSH interface language (中文 / English).

Settings at a glance

Settings → "Prompt Enhancer"; changes save immediately:

Capture, transcription and writing back are all handled by DSH's official voice control — the plugin only adds the value-adds above. Audio is not written to disk.

Releases · CHANGELOG · Compatibility notes · Licensed under Apache-2.0

Privacy: the plugin records and reports nothing; enhanced results come from an external LLM — verify before sending.