MrGeDiao/shuorenhua

shuorenhua (说人话) – Chinese‑AI‑flavor “humanizer”

What it is – A lightweight skill that can be plugged into LLM‑powered agents (Codex, Claude Code, Cursor, ChatGPT, etc.) to rewrite Chinese text so it no longer sounds like a generic AI‑generated response. It removes stereotypical “AI‑tone” patterns (over‑polite openings, engineer‑speak, marketing buzzwords, translationese, unsourced authority claims) while preserving every factual detail – numbers, versions, commands, error messages, attributions, and timestamps stay exactly where they belong.

Why you’d use it

Scenario	What shuorenhua does
Casual chat	Cuts out filler openings and sales‑y sign‑offs, keeping the conversation natural.
Technical status updates	Keeps facts, version numbers, command snippets and error logs untouched, while stripping empty “progress” language.
README / release notes	Shows a clear change list, version, validation steps and limits – no promotional fluff.
Forum posts / issue replies	Reads like a conscientious maintainer, not a canned announcement.
Long Chinese articles	Removes empty sentences, lists them as “suggested deletions” for human approval, and guarantees the original length and key information stay intact.

How it works

1. Scene detection – decides whether the input is chat, status, docs, or public‑writing and loads the appropriate Scene Pack (README, release‑note, forum‑post, etc.).
2. Protected‑span extraction – automatically locks numbers, version strings, file paths, command lines, error traces, names and responsibility statements.
3. Intensity ranking – classifies the text into Tier 1/2/3 and selects a rewrite strength (minimal / standard / aggressive) and a scope (structural / bounded / in‑place).
4. Pattern‑based rewrite – applies a curated dictionary of Chinese AI‑style patterns; it is not a blind find‑and‑replace.
5. Fidelity check – re‑reads the output to ensure all protected spans and factual relations are unchanged.
6. Residual audit – a second, lightweight pass that only fixes remaining “AI‑taste” leftovers.

The rule set lives in references/ (positive‑style guide, protected‑spans list, scene‑pack definitions) and is deliberately strict: numbers must stay attached to their objects, relationships cannot be altered, time spans cannot be shortened, and no new facts may be invented.

Getting started (≈30 seconds)

• ChatGPT (Plus/Pro) – use the hosted “说人话 GPT” link; paste text and ask it to “把这段去 AI 味”.
• Claude Code – run the two plugin commands shown in the README, then issue 把这段去 AI 味.
• Codex – clone the repo and execute a one‑liner like:

  codex exec -C . "读取 ./SKILL.md，按其中规则改写以下文本：…"
  

• Other agents – npx skills add MrGeDiao/shuorenhua (or follow the platform‑specific install/*.md docs).

For long‑form documents you can request annotation mode (只标注不改写) to get a list of suggested deletions without the tool actually removing anything.

Evaluation

• A public benchmark of 103 cases (57 “must‑change”, 46 “must‑stay”) is provided in evals/benchmark.md.
• Results are reported per model (Codex, Claude) with hard‑constraint failures (L1), false‑positive rate (SNF), and style‑target metrics (L2). The latest release (v2.3.0) shows 0 hard‑constraint failures and < 10 % false‑positive rate on the blind test set.
• Automation scripts (automation/eval/hard_metrics.py) compute length‑preservation, dash density, and protected‑span integrity for long‑form runs.

Installation

Platform	Docs
Codex	install/codex.md
Claude Code	install/claude-code.md
Cursor / Windsurf	install/cursor.md
OpenClaw	install/openclaw.md
ChatGPT / custom GPT	install/chatgpt.md

The core of the skill is a single file SKILL.md; for full‑featured use (scene packs, protected‑span lists) include the references/ directory.

License – MIT.

---

TL;DR – shuorenhua is a Chinese‑language “humanizer” skill for LLM‑based agents that strips away AI‑style boilerplate while guaranteeing factual fidelity. It works via rule‑based pattern matching, scene‑aware intensity settings, and a two‑pass verification pipeline, and it ships with a thorough benchmark and easy installation for the major AI‑assistant platforms.