humanizer-paper

# Humanizer (Academic): register-aware AI-tell removal Polish an academic draft so it reads like a careful human author **and** conforms to its register's norms. Two modes: English journal articles (`en-journal`) and Chinese doctoral dissertations (`zh-dissertation`). The analysis kernel is the "Signs of AI writing" taxonomy from [Wikipedia](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) (WikiProject AI Cleanup), **re-gated** for academic register: several general-prose defaults are backwards in a journal or dissertation, so each tell is kept, calibrated, or disabled rather than applied blindly. The heavy content lives in `references/`; this file is the router and the behavioral rules. > In the `python` command below, `<skill-dir>` is this skill's base directory, > announced when the skill loads. Substitute that literal path; it is not an > environment variable. The bundled script self-locates, so only the path needs > to resolve. ## 诚信边界 (behavioral hard rule) This skill polishes **the author's own draft** for clarity and for compliance with academic norms. It is **not** a detector-evasion tool. - The legitimate, transferable techniques here — varying cadence, calibrating hedging, tightening argumentation, enforcing terminology consistency, protecting citations — apply only to text the author actually wrote or substantively authored. - If the request is framed as "rewrite this generated text so it passes Turnitin / 知网 AIGC / an AI detector", **refuse that framing**. Redirect to the legitimate goal: improving real originality (add real data and citations, deepen analysis, fix norms). High AIGC rate is treated by many institutions as academic misconduct; do not position this skill as a way around that. ## When to use - Polish an English journal/conference manuscript or section to remove AI tells while keeping formal register, hedging, and section-appropriate passive. - Polish a Chinese 学位论文 (摘要/引言/方法/结果/讨论/结论/致谢) for AI tells **and** GB punctuation, terminology unification, and AIGC quantitative self-check. - Norm-check a draft (`--check-only`) and report mechanical tells without rewriting. ## When NOT to use - **Generic / non-academic prose** (blog posts, essays, marketing). This skill's defaults are tuned for academic register; for general "make it sound human" use an upstream generic humanizer, not this one. - **Paper intake, reading, synthesis, or literature review** (skim, deep-read, card, compare, gap map, review outline, normalize a DOI/arXiv source). That is `paper-workbench`, not this skill. This skill edits _language_; `paper-workbench` analyzes _content_. - **Implementing a paper's method into code.** Out of scope. ## Routing ### 1. Target (which norm pack) - If `--target` is given, use it. - Otherwise infer from the text's CJK ratio: a high count of CJK characters (roughly > 2% of characters, like `infer_language` in `paper-workbench/scripts/normalize_paper.py`) means `zh-dissertation`; otherwise `en-journal`. - If the language is genuinely ambiguous (mixed-language, very short), use `AskUserQuestion` to pick `en-journal` vs `zh-dissertation`. Do not guess silently. ### 2. Section (which gates) - Infer the section from headings or structure: abstract, introduction, methods, results, discussion, conclusion. `--section` overrides. - Section drives the calibrated tells (most importantly passive voice: keep in Methods, may activate in Discussion). If undeterminable, treat as "general academic body". ## Core loop 1. **Classify.** Read the input. Resolve target and section (Routing above). Identify each AI tell using `references/ai-tells-academic.md`, applying its keep / calibrate / disable gate — do **not** apply general-prose defaults blindly. 2. **Load the norm pack.** Read `references/en-journal.md` or `references/zh-dissertation.md` for the active target. It decides per-section behavior and adds register-specific norms. 3. **Register-aware draft.** Rewrite to remove the tells while obeying the norm pack. Cover everything the original covers (same number of paragraphs/claims). Calibrate hedging (do not delete it), keep section-appropriate passive, unify terminology, fix ghost citations by supplying real `(author, year)` / `[n]` or lowering claim strength, replace hollow generalities with specific data/method/citation, and vary sentence cadence. 4. **Mechanical lint.** Run the script (below) for the quantifiable tells the model tends to miss (dash/quote characters, sentence-length and burstiness stats, "首先/综上" comma-clauses, optional terminology variants). The script is a copilot: it gives coordinates; it does **not** rewrite. 5. **"Still-AI" audit.** Ask: _what still reads as AI-generated here?_ List the remaining tells briefly (uniform cadence, residual ghost citations, templated paragraphs, over-stacked hedges). 6. **Final.** Revise to address the audit and the lint hits. In `zh-dissertation`, conform punctuation to GB and keep full-text terminology consistent; in `en-journal`, keep tense and citation style consistent per the norm pack. ## Mechanical check Run the linter for quantifiable tells (planned entry; if the script is not yet present, do the mechanical scan by hand from the same checklist): ```bash python "<skill-dir>/scripts/polish_lint.py" \ --target "<en-journal|zh-dissertation>" \ --file "<path-or-omit-to-read-stdin>" \ --json ``` The linter reports surface tells (em/en dash, curly quotes, AI high-frequency words, Chinese "几字+逗号" short clauses), cadence stats (sentence count, mean length, burstiness, over-long ratio with the zh > 28 字 threshold), and optional terminology variants when a `--glossary` is supplied. It is a reporter (exit code always 0); the rewrite stays with the model, guided by `references/`. ## Output contract Deliver, in order: 1. **Draft rewrite** — register-correct, all covered content preserved. 2. **"Still-AI" audit** — brief bullets of remaining tells. 3. **Final rewrite** — addresses the audit and the lint hits. 4. **Change summary** — what was changed and why (and, optionally, the lint report). For `--check-only`, skip the rewrite: return the lint report plus a short list of the tells found, by gate, with no edited text. ## References - `references/ai-tells-academic.md` — the re-gated 33-pattern kernel (keep / calibrate / disable) plus 5 new academic-specific tells, with academic before/after. - `references/en-journal.md` — English journal norm pack (register, hedging, section-gated passive, IMRaD/CARS, tense, citation styles, dash/comma, spelling). - `references/zh-dissertation.md` — Chinese dissertation norm pack (语体, 术语统一, GB/T 15834 标点, GB/T 15835 数字, 法定计量单位, 摘要/结论/标题/致谢, AIGC 量化特征). ## Attribution The pattern taxonomy is from [Wikipedia: Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing), maintained by WikiProject AI Cleanup. This skill (MIT) re-gates that taxonomy for academic register and adds two norm packs and a mechanical linter.