The academic-writing skill that makes AI-assisted papers read human while staying evidence-true — in English and 中文.
让 AI 辅助的学术写作既有人味、又守得住证据：去 AI 味 + 守证据 + 降 AIGC 检出率，中英双语。
A general "humanizer" strips AI tells but has no idea that your p-value is load-bearing. An academic polisher fixes structure but doesn't know why clean prose still trips GPTZero. And neither one handles 中文 AI 腔. Research writing needs all three at once, plus a hard rule that nothing gets less true in the name of sounding human.
Easywrite is that skill. It routes between five strengths and adds a bilingual + integrity spine:
- 🧬 Claim ↔ evidence discipline — no verb stronger than its data; every claim earns its number, figure, or citation. Papers, rebuttals, and grants each get the right register (grants keep their vision; papers get the strict trim).
- 🔬 Detector science, done right — perplexity / burstiness / stylometry. It cuts instead of "polishing" (polishing raises the AI score), and it tells ESL and technical writers the truth: a high score is often the detector's bias, not their writing.
- 🈶 Bilingual — a full English AI-tell catalog and a 中文 去 AI 味 engine (场景 / 档位 / scope、protected spans、无源引用三模式).
- 🎯 Voice & venue matching — matches your own writing sample and the target venue (NeurIPS terse vs. Nature expository vs. NSFC 规范).
- 🔒 Integrity firewall — never alters a number, result, or citation; never fabricates a source; not a tool for evading AI-use disclosure.
The whole thesis in one line: human in voice, true in substance — writing that is easy to read and easy to trust.
Route (language · doc-type · goal)
└─ Diagnose → tag paragraphs (evidence-dense = leave alone; analytical = fix here)
└─ Protect → lock numbers, results, cite keys, terms, system subjects
└─ Cut → the strongest move; remove filler, restatement, hype
└─ Anchor → claim ↔ evidence (papers) / claim ↔ feasibility (grants)
└─ De-tell → em-dashes, vocab clusters, rule-of-three, 综述腔/翻译腔
└─ Burstiness→ uneven sentence & paragraph rhythm
└─ Voice → match author sample + venue register
└─ Score → 2 hard gates (Fidelity, Evidence) + 5-dim rubric, then STOP
Once installed, just ask in natural language — the skill auto-triggers:
- "Humanize the introduction of my paper, keep every number."
- "降一下这段摘要的 AIGC 检出率，别动数据和引用。"
- "My rebuttal reads too AI and too grovelly — tighten it and tie each reply to a change."
- "Polish Aim 1 of my NSF proposal but keep the vision."
- "这段哪里像 AI？先别改，先标问题。" (diagnose-only mode)
Option A — Skills CLI (any agent):
npx skills add Florent2025/easywriteOption B — Claude Code plugin marketplace:
/plugin marketplace add Florent2025/easywrite
/plugin install easywriteOption C — Manual (Claude Code / Codex / Cursor / OpenCode):
# Claude Code (user-global):
git clone https://github.com/Florent2025/easywrite.git ~/.claude/skills/easywrite
# Codex: ~/.codex/skills/easywrite Cursor: .cursor/skills/easywrite OpenCode: ~/.config/opencode/skills/easywriteOption D — No automation: paste SKILL.md into the conversation. It runs standalone; full power is SKILL.md + references/.
See install/ for per-tool notes.
easywrite/
├── SKILL.md # core router + 8-step workflow + IRON RULES
├── references/
│ ├── ai-tells-english.md # 33 patterns + 2026 vocab clusters + false-positive guards
│ ├── ai-tells-chinese.md # 场景/档位/scope · protected spans · 无源引用三模式
│ ├── academic-discipline.md # over-claim verbs · venue · grant (NSF/NIH/NSFC) mode
│ ├── detector-science.md # perplexity/burstiness/stylometry · the traps · ESL bias · watermarks
│ ├── scoring-rubric.md # 2 hard gates + 5-dim rubric + ship gate
│ └── examples.md # before/after: EN paper · 中文摘要 · rebuttal · grant
├── evals/benchmark.md # should-fix + fidelity-guard smoke tests
├── install/ # per-tool install notes
└── .claude-plugin/ # plugin + marketplace manifests
- Fidelity is a gate, not a dimension. A stylistically perfect draft that altered a number is a failure.
- Cutting beats rewriting. "Polishing" smooths prose toward AI register. Remove, don't re-smooth.
- Preserve what works. Evidence-dense paragraphs and evidence-tied hedges are human — leave them.
- Clusters, not isolated hits. One em-dash proves nothing; four tells together is a confession.
- Honesty about limits. It clears a flawed stylometric filter for a real author; it does not beat SynthID/C2PA watermarks and is not for disclosure evasion.
Easywrite stands on the shoulders of five excellent MIT-licensed skills. It does not replace them — it routes between their strengths and adds the bilingual + integrity spine. Please star the originals:
| Source | Author | What Easywrite took |
|---|---|---|
| humanizer | @blader | The 33-pattern English AI-tell catalog, voice calibration, false-positive guards |
| academic-humanizer | AIScientists-Dev | Claim↔evidence discipline, over-claim verbs, venue & NSF/NIH grant modes |
| humanize-prose | @celestialdust | Detector science (perplexity/burstiness/stylometry), cut>rewrite, the traps, ESL bias, watermark limits |
| stop-slop | Hardik Pandya | Tight core rules + the multi-dimension scoring rubric |
| shuorenhua 说人话 | @MrGeDiao | The entire 中文 去 AI 味 engine: scene/tier/scope, protected spans, 无源引用三模式 |
Adjacent and complementary: nuwa-skill 女娲 (distills a thinker's voice you can then feed to Easywrite's voice-match step) and taste-skill (frontend/UI, not prose).
MIT. If you build on Easywrite, keep the attribution chain intact — credit the five sources above, as this repo does.
Issues and PRs welcome: new AI-tell patterns (with a before/after), venue register notes, additional languages,
and benchmark cases. Every rule should come with an example, and no rule may weaken an IRON RULE in SKILL.md.