What does this skill do, and when should you use it?
Shuorenhua (说人话) is a rewrite skill for Chinese text that removes templated AI phrasing, marketing packaging, engineer-posturing, translation-ese, and unsupported claims. Its core rule is to lock down facts before editing: metrics, versions, commands, paths, causal conditions, and attribution are always preserved; only excess flourish is removed. It works with Claude Code, Codex, Cursor, and ChatGPT, and ships scene packs (README, Release Notes, Issue replies) plus long-form scope controls. The repo includes a 120-case benchmark plus 20 scenario samples, with the latest release at v2.4.0.
After reading the input text, it applies minimal edits per SKILL.md rules: deletes pure greetings, empty teasers, and exact duplicates; simplifies nominalizations, bloated function words, and stiff syntax; removes unsupported credential claims. Facts, terminology, attribution, numbers, commands, and quotations stay intact throughout. It supports structural / bounded / in-place scopes, minimal / standard / aggressive intensities, and an annotation mode that only flags problems without rewriting. Pure prompt-based skill — no scripts or network needed.
- A developer publishing release notes who wants PR-style declarations removed while keeping concrete p95 latency figures
- A maintainer replying to an issue who wants empathy-padding and credential claims gone, leaving reproduction status and next steps
- A technical writer drafting a README who wants the first screen to explain the project without empty slogans
- An engineer de-translation-ese-ing a long AI-drafted document without altering conditions, negations, or intensity
- A user who wants problems flagged but no edits, using annotation mode
How do you install this skill?
- All eval results are self-reported; no independently executable verification is included in this review — spot-check before adopting.
- CHANGELOG uses future 2026 dates; treat timeline with care.
- Grok covered only 11/58 normal-text false-edit checks; full false-edit rate unknown.
- Release gating partly depends on maintainer's unilateral decisions; second-model coverage is incomplete.
Claude Code: run /plugin marketplace add MrGeDiao/shuorenhua, then /plugin install shuorenhua@shuorenhua. Codex: clone the repo and run codex exec -C . "Read ./SKILL.md and rewrite the following text per its rules: …". Agents supporting the skills protocol: npx skills add MrGeDiao/shuorenhua. A web version, the 说人话 GPT, is also available (requires ChatGPT Plus / Pro).
git clone --depth 1 https://github.com/MrGeDiao/shuorenhua.git ~/.claude/skills/shuorenhuaGenerated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- remove the AI tone from this passage
After install, type something like "remove the AI tone from this passage" and paste your text. Optional parameters: annotation mode (flag only), scope (structural / bounded / in-place), intensity (minimal / standard / aggressive), and rewrite-safe (delete unsourced claims). Default output is one publish-ready rewritten draft, with no scores attached.
What are this skill's strengths and limitations?
- Strict fidelity design: numbers, commands, attribution, conditions, and quotations are preserved by default, avoiding over-generalizations like "clearly optimized"
- Backed by a 120-case benchmark with an L1 hard-constraint gate (fabricated facts and protected-span drift must be zero)
- Multi-platform install paths and tiered scope control; pure prompt with zero dependencies
- Explicitly distinguishes quoting a claim from endorsing it, so discussed wording is not deleted as boilerplate
- Chinese-first; English rules cover only common stock phrases and mixed-language redundancy
- Effectiveness depends on how well each model follows the rules; cleanup rates vary by model (tracked separately in the repo)
- Positions itself as rule-driven redundancy cleanup and does not mimic a specific author's personal voice
- v2.4.0 evals cover two models; false-positive rates for other models lack public data
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Shuorenhua: Chinese AI-Tone Cleanup Skill this page | 65 · Recommended | ★ 2.5k | 1d ago | MIT |
| HumanAI — Text Humanization Engine | 52 · Use with care | ★ 39 | 3mo ago | MIT |
| Humanizer — AI Writing Humanizer & Detector Skill | 67 · Recommended | ★ 302 | 2mo ago | MIT |
| e2e Docs Authoring Skill | 54 · Use with care | ★ 12k | 6d ago | Apache-2.0 |
| Humanizer Writing Skill | 53 · Use with care | ★ 65k | 9d ago | MIT |
The README lists related projects: stop-slop (English AI-redundancy rules and evaluation), humanizer (English AI-pattern taxonomy), and speak-human-tw (Traditional Chinese). For mainly English text those are better fits; for Simplified Chinese this skill is more complete.
How did FollowSkills review this skill?
SKILL.md is pure text-editing rules: no external calls, no dependencies, no sensitive-data handling; minimal edits by default, code blocks preserved verbatim, transparent data flow. Deducted for: no rollback mechanism stated, unverified publisher identity.
High internal consistency: explicit edit-action list, scope parameter table, pre-delivery verification, defined failure behavior (revert uncertain edits, state no remaining content). Deducted for: static review cannot execute; abnormal-input behavior rests only on rule text, capped at 10.
Clear scenarios (READMEs, release notes, work syncs, personal writing), declared non-fit boundaries (not an AI detector, no fact-checking), precise semantic triggers, Chinese-first and fully local with no overseas service dependency. Deducted for: boundaries conveyed via long prose rather than structured trigger conditions; English is secondary.
MIT license explicit, thorough CHANGELOG honestly recording failures and unreleased candidates, README install instructions, runtime-files. defining runtime boundary, maintenance path described. Deducted for: future-dated 2026 changelog entries are questionable; release gates partly rest on maintainer's unilateral judgment.
Clear value claim with inline example showing directly usable output. Deducted for: static review cannot verify actual output quality; improvement evidence (5/7) is self-reported, capped at 7.
Repository provides detailed eval records, failure archives, criteria files and a CI workflow; negative results honestly recorded. Deducted for: all evidence is author-reported; key result files not present in the reviewed sources; cannot independently reproduce, capped at 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →