lynote-ai/dsh-humanizer

A writing skill for DeepSeek Harness (dsh). The core idea is not "write like a human", but "write like me": strip generic AI-sounding patterns on one side, learn your own writing fingerprint on the other, and turn any draft into something you wrote.

This is a skill for the agent, not an LLM wrapper: it performs no model calls. It only produces rules, fingerprints, scores, and rewrite briefs — the agent does the actual rewriting itself.

Install

# from GitHub (ships prebuilt lib/)
dsh plugin add github:lynote-ai/dsh-humanizer

# or, once published to npm:
# dsh plugin add dsh-humanizer

Optional config (all fields have defaults):

- insert:
    - id: dsh-humanizer
      name: 'dsh-humanizer'
      config:
        strength: standard                          # light | standard | aggressive
        storagePath: ~/.dsh/voice-profiles.json
        maxExcerpts: 3

Tools (8 — all deterministic, no model calls)

De-AI:

Tool	Purpose
humanize_scan	Detect AI patterns and return an AI-ness score plus the matched rules
humanize_rules	Export the full rule catalogue (transparent, editable)
humanize_rewrite	Return a rewrite brief (rules + issues found in this text); the agent applies it

Personal voice clone:

Tool	Purpose
voice_import	Import samples → extract a style fingerprint → persist a profile
voice_profile	Read one profile, or list all
voice_remove	Delete a profile
voice_score	Similarity between text and a profile (0–100 + per-feature breakdown)
voice_rewrite	Return a rewrite brief (fingerprint + few-shot samples + issues found)

How it works

src/core/ is a pure, dependency-free, unit-tested library:

• rules.ts — AI-writing pattern catalogue (English + Chinese, modeled on stop-slop / Humanizer-zh)
• analyze.ts — de-AI scan: empty openers, clichés, hedging, template transitions, mechanical parallelism, summary endings
• fingerprint.ts — style fingerprint: sentence length / burstiness, punctuation habits, stance (person, adverbs, contractions), preferred vocabulary, lexical richness
• score.ts — similarity scoring (0–100 + per-feature breakdown)
• render.ts — rewrite-brief construction

humanize_rewrite / voice_rewrite never call a model — they return a brief, and the agent rewrites in its own turn. voice_score's per-feature output is the hook for future "learn from feedback" iteration.

Example (agent's point of view)

User: I have a dozen tweets I wrote. Build me a "my voice" profile.

Agent:
  1. voice_import(name="me-x", samples=[...])   → extract & store fingerprint

User: Rewrite this AI-written release post in my voice.

Agent:
  1. voice_score(text=draft, name="me-x")        → 41/100
  2. voice_rewrite(text=draft, name="me-x")      → get brief (fingerprint + samples + issues)
  3. agent rewrites following the brief
  4. voice_score(rewritten, name="me-x")         → 83/100

Development

npm install
npm run check       # typecheck + build
npm test            # unit tests (node --experimental-strip-types)
npm run build       # tsc → lib/

Design decisions

• No LLM calls — rules, fingerprints, and scores are deterministic and reproducible.
• Single self-contained bundle — no unpublished shared dependency, so dsh plugin add just works.
• File-based persistence — ~/.dsh/*.json, no storage backend required.

License

BSD-3-Clause

---

中文说明

面向 DeepSeek Harness（dsh）的写作插件。核心理念是——不是「像人写」，而是「像我写」：一边去掉通用 AI 腔，一边学习你本人的写作指纹，把任意草稿改成「你写的」。

这是一个给 Agent 用的 skill，不调模型：只产出规则、指纹、打分和「改写 brief」，真正的改写由 Agent 自己完成。

• 去 AI 味：humanize_scan（检测）/ humanize_rules（规则库）/ humanize_rewrite（改写 brief）
• 文风克隆：voice_import / voice_profile / voice_remove / voice_score（0–100 相似度）/ voice_rewrite（改写 brief）

规则库中英双语覆盖（参考 stop-slop / Humanizer-zh）。指纹提取与打分均为确定性计算，可复现、可单测；Profile 默认持久化到 ~/.dsh/voice-profiles.json。

安装：dsh plugin add github:lynote-ai/dsh-humanizer