shu​ren​hua – Chinese‑first “Say‑it‑in‑human‑words” editing skill

What it is

shu​ren​hua is a text‑editing skill (a prompt / rule‑set) that you load into a generative‑AI chat assistant (Claude, Grok, ChatGPT, Cursor, etc.). Its purpose is to clean up Chinese prose – removing boiler‑plate, repetitive set‑ups and stiff phrasing while preserving the original facts, conditions, promises and the author’s voice. It can also work on English text and on “review‑only” requests that only flag problems without rewriting.

Who might use it

• Developers writing README files, release notes, issue comments, or internal sync messages.
• Technical writers and product managers who want a concise, jargon‑free Chinese version of a document.
• Anyone who drafts Chinese articles and wants an automated “human‑sounding” polish without losing meaning.

How it works (from the repo)

1. Skill definition – The core editing logic lives in three markdown files:
   • SKILL.md – the entry point that the AI loads.
   • references/editing‑guide.md – detailed boundaries (what may be changed, what must stay).
   • references/examples.md – side‑by‑side before/after examples.
2. Installation – Copy the three files listed in runtime-files.json into the directory that your AI‑assistant scans for skills. Three distribution flavours are provided:
   • mini – a single self‑contained markdown (dist/shuorenhua‑mini.md) for one‑off prompts.
   • lite – only SKILL.md (fallback to the reference files when needed).
   • full – all three files for the most complete behaviour. Installation instructions are supplied for several popular tools (Codex, Claude Code, Cursor/Windsurf, OpenClaw, ChatGPT).
3. Running – You feed the original text to the AI together with a short instruction, e.g.

   把这段去掉套话，保留信息和我的口气：……
   

   The model then rewrites the passage according to the rules, returning the cleaned version (or just a list of issues if you asked for “review‑only”).
4. Control granularity – The skill recognises three scopes:
   • structural – free editing, sentences may be merged or reordered.
   • bounded – only in‑sentence clean‑up; whole‑sentence deletion is allowed but no merging.
   • in‑place – never delete or merge sentences; only word‑level tweaks. Users pick the scope with a keyword in their request.
5. Safety checks – The rules explicitly keep:
   • Numbers, conditions, commitments, and attribution.
   • Code blocks (including comments) unchanged unless the user explicitly asks to edit them.
   • Quotations that are citations, UI strings, or titles.
   • Statements that lack a source (“研究表明”) are kept unless the user requests removal.

Evaluation & quality

• Benchmark: 123 test cases (65 “to‑improve” texts, 58 normal texts) plus 20 scenario samples and 10 long‑form human articles.
• Model coverage: Claude Opus 5 passed 136 targeted cases; Grok 4.6 passed 56.
• Metrics: Zero hard failures (information loss or missed task), normal‑text mis‑edit rate < 10 %, and at least 5 of 7 positive‑sample edits show a clear benefit.
• Detailed results are in evals/results‑v2.5.0.md and the criteria file evals/v2.5/criteria.md.

When it’s not a fit

• If you need a full‑blown AI writing detector or a tool that guarantees “AI‑free” text, this skill does not provide that.
• It is primarily for Chinese; English is supported but the rule set is tuned for Chinese phrasing.
• It does not rewrite code logic – only documentation‑style text inside code fences.

License & community

• Open‑source under the MIT license.
• Contributions are welcomed via the Bad‑Case template; issues should include the original text, the request, and the model’s output.
• The repo tracks star growth and provides a changelog; the latest stable version is v2.5.1 (maintenance release).

---

All information above is taken directly from the repository’s README; no external assumptions have been added.