dsh-voice-prompt-compressor

Compress verbose voice-dictation text into token-efficient prompts — fully local, zero LLM tokens.

A DeepSeek Harness (DSH) plugin that removes filler words, repetitions, and politeness padding from speech-to-text / dictation output, then helps organize the result into a structured prompt (Context / Goal / Constraints / Deliverables).

Features

• Deterministic local compression — normalize → strip fillers → strip politeness → dedupe. No network, no LLM call, no tokens spent.
• Bilingual wordlists — Chinese and English fillers, hedges, and politeness phrases; auto language detection by CJK ratio.
• compress_voice_text tool — callable by the agent; returns compressed text plus savings stats (estimatedTokensSaved, ratio, removed counts per category).
• Bundled skill voice-prompt-compressor — auto-triggers when the user pastes rambling dictation, and organizes the compressed text into a four-section prompt.
• Configurable — mode (light / balanced / aggressive) and keepPoliteness overridable in your profile patch layer.

Install

# local development install (file: reference)
dsh plugin --profile web add /path/to/dsh-voice-prompt-compressor
# after publishing to npm
dsh plugin --profile web add dsh-voice-prompt-compressor

Note: dist/ is not committed — after cloning, run npm install && npm run build before installing.

Refresh the web page after installing. The plugin registers a tool and a skill; both become available in new sessions.

Usage

Skill (recommended): paste rambling voice dictation into the chat. The agent loads the voice-prompt-compressor skill, calls compress_voice_text, and presents a compressed four-section prompt (Context / Goal / Constraints / Deliverables).

Tool: the agent can call compress_voice_text directly with these parameters:

Parameter	Type	Default	Description
text	string (required)	—	The dictation / transcript text to compress
language	auto | zh | en	auto	Wordlist language; auto detects by CJK ratio
mode	light | balanced | aggressive	config	Compression strength
keepPoliteness	boolean	config	Keep politeness phrases when true

The tool returns:

{
  "compressed": "…",
  "originalLength": 512,
  "compressedLength": 210,
  "estimatedTokensSaved": 76,
  "ratio": 0.59,
  "removedCategories": { "fillers": 18, "repeats": 3, "politeness": 2 }
}

Config

Override in your profile patch layer (same id):

- insert:
    - id: voice-prompt-compressor
      config:
        mode: balanced      # light | balanced | aggressive
        keepPoliteness: false

How it works

The pipeline is purely mechanical and deterministic:

1. normalize — full-width alphanumerics to half-width, unify whitespace.
2. strip-fillers — remove filler words and discourse markers (嗯, 那个, 就是说, 然后, um, like, you know, basically, …). Ambiguous demonstratives (那个 / 这个 / 就是) are only removed adjacent to punctuation or whitespace in balanced mode; aggressive removes them anywhere plus hedges (说实话, frankly, …).
3. strip-politeness — remove politeness padding (麻烦你, please, could you, thanks, …) unless keepPoliteness: true.
4. dedupe — collapse adjacent repeats (不对不对 → 不对, very very → very).

Compression is mechanical only — it never rewrites meaning. Technical requirements, constraints, edge cases, and business rules are preserved verbatim.

Development

npm install
npm test        # builds then runs node --test
npm run typecheck

Publishing

1. Push the repo to GitHub.
2. (Optional) npm publish.
3. Submit to the DSH plugin market: open an issue/PR at dsh-market/dsh-market or register at https://awesome-dsh-plugin.com.

License

MIT