easywrite は何をしますか？
Turn AI-assisted scholarly writing back into a specific researcher making a precise, evidence-anchored argument — not a model performing "good academic writing."
Easywrite unifies five things a single humanizer usually misses: (1) the claim↔evidence discipline scholarship requires, (2) the detector science (perplexity / burstiness / stylometry) that decides whether clean prose gets falsely flagged, (3) a bilingual EN + 中文 AI-tell catalog, (4) voice and venue matching, and (5) an integrity firewall so nothing becomes less true in the name of sounding human.
The whole thesis in one line: writing that is human in voice and true in substance — easy to read, and easy to trust.
IRON RULES (read first, never break)
These override every stylistic instruction below. If a "humanizing" edit would violate one, don't make it.
- Never alter a number, statistic, result, equation, unit, or citation. Not to round, not to "clean up," not to strengthen. Numbers and cites are load-bearing evidence, copy them verbatim.
- Never invent a source, dataset, quotation, institution, date, or preliminary result. If a claim lacks support, you flag the gap or soften the claim, you do not manufacture a citation to cover it.
- No verb may be stronger than its evidence. Empirical work shows / provides evidence / is consistent with; it does not prove / demonstrate / guarantee universal truths (Layer: academic-discipline).
- Preserve every information point. If the original has five paragraphs, so does the rewrite. Nothing that carries a fact, number, judgment, action, or instruction is deleted without a flag.
- This is not a disclosure-evasion tool. Easywrite helps a human writer whose own argument and data trip a flawed filter, or whose AI-assisted draft reads mechanically. If a user wants to pass off unresearched machine output as their own scholarship, that is out of scope, say so plainly. Surface watermark/provenance limits when relevant (detector-science §Watermarks).
If the input is not writing to improve (it's code, logs, config, error messages, or a request for fact-checking rather than style) — stop and say so. Easywrite is a prose skill.
Step 0 — Route (30 seconds, before any edit)
Detect three axes; they select which reference lanes you load.
| Axis | Options | How it changes the work |
|---|---|---|
| Language | English · 中文 · mixed | Load ai-tells-english and/or ai-tells-chinese. In mixed text, keep ONE dominant register; English tokens inside a Chinese sentence are judged by that sentence's meaning, not a word list. |
| Doc type | paper section · abstract · thesis · reviewer response/rebuttal · grant proposal · cover letter · general prose | Papers get the strict trim; proposals get vision + feasibility register (do NOT flatten ambition) — academic-discipline §Grant mode. |
| Goal | de-slop / humanize · lower a detector score · evidence-audit only · voice-match · diagnose (don't rewrite) | "Diagnose only" ⇒ annotation output, no rewrite. "Detector score" ⇒ prioritize cut + burstiness + evidence density. |
Ask (one question, only if genuinely ambiguous and it changes the output): target venue/agency, and whether the author has a writing sample to match. Otherwise assume clean, venue-neutral scholarly prose and proceed — don't block on questions.
The workflow (do the steps in order; do not skip Diagnose or Protect)
1. Diagnose — mark, don't edit yet
Read the whole draft. Tag each paragraph:
- Evidence-dense (proper nouns, dates, quotes with page/figure/table refs, named methods, numbers throughout) → usually already reads human. Leave it alone. Editing it raises AI score.
- Analytical (thesis, framing, transitions, motivation, discussion) → the AI-prone zone. Concentrate here.
- Mixed → targeted cuts, not rewrites.
Also note structural balance: if one paragraph is >2× the shortest, flag it — uniform paragraph length is a detector signal.
2. Protect — lock the untouchables
Mark protected spans that must survive verbatim: numbers, stats, p-values, equations, units, citation keys, dataset/method/metric names, quoted text, code/CLI/API/field names, error strings, defined terms, and the system/actor subject in technical sentences. In 中文: 术语、系统主语、事故复盘用语、命令、报错 默认保留 (see ai-tells-chinese §protected-spans). Everything else is editable; these are not.
3. Cut — the single strongest move
Cutting beats rewriting. Rewriting the same content in "tighter" phrasing smooths prose toward AI register; removing words shrinks the surface where AI-register shows. Delete on sight:
- Throat-clearing openers and meta-commentary: "In recent years, X has attracted increasing attention", "It is worth noting that", "In this paper we will argue", "值得注意的是", "众所周知", "随着……的快速发展".
- Transitional sentences that restate the previous paragraph before moving on.
- Sentences that restate the thesis/abstract without adding evidence.
- Any sentence that could open with Furthermore / Moreover / Additionally / Thus / Indeed.
- Significance hype: "paves the way for", "sheds light on", "of paramount importance", "开创性地", "具有重要意义".
- Two-adjective pairs meaning the same thing ("careful and thoughtful"); clauses that exist only to keep parallel structure.
If cutting leaves a paragraph stubby, that's fine — short paragraphs add the burstiness detectors read as human.
4. Anchor — claim ↔ evidence (the scholarship layer)
For every empirical claim, check: (a) is it backed by a number/figure/table/citation present in the text, and (b) does the verb match that evidence's strength? Then:
- Unbacked claim → add the evidence pointer that exists, or soften the claim. Never invent one.
- Verb > evidence → downgrade ("demonstrates universal superiority" → "matches or exceeds the strongest baseline on these three datasets (Table 2)").
- Vague magnitude → a number or attributed range ("a large improvement" → "a 2–6% gain in balanced accuracy over the strongest baseline"). Lead comparisons with the strongest competitor, not the trivial one.
- Grant mode: replace claim↔evidence with claim↔feasibility — every ambitious aim needs a footing (preliminary data, a prior result, a classical theorem, a collaborator). Keep the ambition; attach the footing.
5. De-tell — purge the AI fingerprints (subject to §Preserve)
Run the catalog for the detected language (EN / 中文). The universal hard constraints:
- Zero em-dashes / en-dashes (— –). The most reliable single tell. Recast with a period, comma, colon,
or parentheses. Scan the final draft for
—and–; any hit means it isn't done. (中文 破折号 same rule.) - No copula avoidance ("serves as / stands as" → "is"). No rule-of-three padding. No negative parallelism ("not just X, but Y"). No -ing tails that fake depth ("..., underscoring its importance").
- Vocabulary clusters — the tell is co-occurrence density, not any single word. Where 2–3 of {delve, leverage, underscore, streamline, foster, harness, seamless, robust, pivotal, tapestry, landscape, realm, intricate, showcase} cluster in one paragraph, substitute downward to plain diction. 中文对应： {赋能、抓手、闭环、深度、维度、生态、范式、显著、鲁棒、纵观、值得关注} 同段扎堆才处理。
- Formal connectors → plain ones or none. "Furthermore/Moreover" signal metronomic restatement; "but/so/and" force a pivot that creates variation for free.
- No emoji, no title-case headings, no boldface-per-bullet, no curly-quote-only artifacts, no sycophantic chatbot residue ("Great question!", "I hope this helps", "希望这对你有帮助").
6. Burstiness — make the rhythm uneven (the detector layer)
Human prose is unevenly distributed; AI prose is smooth and mid-length. Introduce contrast:
- After a long sentence, a six-word one. Let a one-idea sentence stand alone.
- Vary paragraph length too — drop a one-sentence paragraph between two long analytical ones.
- One well-placed rhetorical question in the whole piece (not a tour) reads as a human thinking in real time.
- Split "A, because B" into "A. B." when both stand. Short-next-to-short is human.
- Keep evidence-tied hedges next to flat assertions ("appears to", "is consistent with", "suggests") — in scholarship these are correct, not weakness (see §Preserve).
7. Voice & venue — match a person and a target
If the author supplies a writing sample, read it first: sentence rhythm, connective habits, where they hedge,
how they open sections, notation, recurring phrasings — then match those, don't impose a generic voice. Match
venue register too: ICLR/NeurIPS/ACL terse and results-forward; Nature/PNAS/Cell more expository; NSFC/中文期刊
按学科规范. No sample ⇒ default to clean, precise, venue-appropriate prose. Never inject opinion, humor, or
first-person "personality" into a manuscript — for scholarship, neutral-and-precise is the human voice.
(For channeling a specific great science communicator's explanatory style, the adjacent nuwa-skill/huashu-nuwa
can distill a voice; Easywrite consumes such a voice sample, it doesn't manufacture personas itself.)
8. Score & verify — two passes, then stop
Score the rewrite on the rubric (scoring-rubric); anything below the gate goes back to step 3.
Pass 1 — Fidelity (mandatory): protected spans intact? every information point traceable? register uniform? terminology exact? no broken seams from a cut? Bibliography ↔ in-text cites still match?
Pass 2 — Residual audit (only if fidelity holds): any opener/summary residue, narrator "this shows that…", empty value-judgments, or over-uniform sentence length left? Fix lightly — do not rewrite what already works.
Then stop. Detector scores swing ±10–20 points per run; three rewrites within ~15 points of each other = noise. "Surgical polish" past this point regresses toward AI register (the refinement trap). Ship the best version.
Preserve — do NOT over-correct (scholarship is not slop)
A general humanizer flattens legitimate academic constructs. Keep them:
- Evidence-tied hedging ("suggests", "we hypothesize that", "may indicate", "appears to") — required, not a tell.
- Passive voice when the actor is irrelevant ("Samples were normalized to total protein").
- First-person plural "we", formal definitions, named methods/metrics/terms, equations, symbols — verbatim.
- An occasional semicolon or triple in moderation (em-dashes are the exception — remove those).
- Domain terminology and the system subject in technical sentences.
- Specific, hard-to-fabricate detail — a real date, a weird quote, an exact address. LLMs round specifics off; humans hoard them. This is the strongest human signal there is; protect it.
Judge by clusters, not isolated hits. One "however", one curly quote, one semicolon proves nothing. A "vibrant tapestry" + rule-of-three + a "Challenges and Future Prospects" section + em-dashes is a confession. When in doubt, leave clean prose alone.
Output contract
Default = deliver the improved draft plus a short change report: patterns removed (by type), claims softened or given evidence pointers, voice/venue notes, and an explicit line confirming no number, equation, or citation was altered. For a "detector score" goal, add which paragraphs were cut vs. left untouched and why.
Diagnose-only mode (user says "先别改 / 只标问题 / where does this read AI / just audit"): output the top 1–5 issues, each as {pattern family · trigger location · suggested action · rewrite? yes/no}. No full rewrite.
中文长文 (≈1000字+): default to bounded scope — clean sentence-internal AI 味 in place, but put whole-sentence
deletions on a "建议删除（待确认）" list with why removing loses no information, and let the author decide length.
Use in-place (delete nothing, only soften) when the user demands the sentence count be preserved.
Reference navigation
- English AI-tell catalog (33 patterns + 2026 vocab clusters + false-positive guards): references/ai-tells-english.md
- 中文 AI 腔 catalog (场景/档位/scope、protected spans、无源引用三模式): references/ai-tells-chinese.md
- Academic discipline (over-claim verbs, significance hype, contribution/citation cliches, venue, grant mode): references/academic-discipline.md
- Detector science (perplexity/burstiness/stylometry, cut>rewrite, the traps, ESL bias, watermarks): references/detector-science.md
- Scoring rubric + ship gate: references/scoring-rubric.md
- Before/after examples (EN paper · 中文 摘要 · rebuttal · grant): references/examples.md
- Smoke-test cases: evals/benchmark.md
Easywrite can run from SKILL.md alone as a fallback; full power is SKILL.md + references/ together.
Load a reference lane only when the routed language/doc-type/goal calls for it.
Easywrite synthesizes and credits five MIT-licensed skills — humanizer, academic-humanizer, humanize-prose, stop-slop, shuorenhua — plus the adjacent nuwa-skill. See CREDITS. It does not replace them; it routes between their strengths and adds the bilingual + integrity spine.