审写 · ReviewWrite

把不该交付的 AI 过程残留，变成可核验的专业正文。

不拿幻觉换“人味”：不编造经历、数据、来源、案例或细节。

Review before rewrite. Keep the facts. Lose the slop.

不是 AI 检测规避器；是专业写作的交付质量控制。

可复现证据：4 份合成案例的严格预检从 14 fail / 23 warn 降至 0 / 0；38 项受保护字面量通过逐字回归检查。另有一项 324,015 字符、55 个分块的长文回归，能将问题定位回全文第 9,001 行。查看完整报告

中文 · English

审写 · ReviewWrite is an evidence-preserving review-and-rewrite Skill for professional writing. It provides compatibility and installation guidance for Codex, Claude Code, Hermes Agent, Gemini CLI, GitHub Copilot, OpenClaw, WorkBuddy, and other Agent Skills-compatible systems.

中文

ReviewWrite 0.9.0 增加来源与交付完整性只读审计，并保留标题句号检查和确定性更新策略：新会话首次调用只读本地版本与缓存，缓存过期后在当前任务完成时联网刷新，不随机检查、不在每次调用时联网。结构化响应仍必须经过正文抽取和严格终检，通过后才输出不带内部标签的正文；它不会为了“具有人味”编造经历、数据、来源或细节，也不会移除或伪装 AI 来源标记。

先审后写，保护事实，交付正文

审写 · ReviewWrite 面向中英文专业写作：先判断文本和目标，再评审、改写、复核。它保护事实、引用、限定条件和作者声音，不把“像人”理解为制造错误或规避检测。

运行环境要求 Python 3.10 或更高版本；正文审写默认只读取本地输入，不联网。更新检查是可选能力，只有用户或平台明确启用时才访问 Release。

超长文档也能审，但不夸大完成度

面对十万字、几十万字的报告、论文、政策材料或长篇研究文本，ReviewWrite 不把全文粗暴塞进一次上下文，也不把几段抽样结果包装成“全文已审核”。它按章节和段落分块，保留全文行号与块号，建立数字、标题和术语一致性索引，并明确报告尚未完成的跨章主张—证据复核。

python3 scripts/long_document_review.py path/to/long-draft.md --format json

分块预检解决上下文长度、定位和一致性线索问题；事实、引用、法律责任和跨章论证仍需全局模型或人工复核。

核心边界：不拿幻觉换人味

ReviewWrite 不会为了让文字更像人，而主动编造经历、数据、来源、案例、情绪或细节。无法确认的内容，会保留限定条件，指出需要核实的部分，或降低主张强度。文字可以更自然，但事实不能靠虚构来补齐。

它背后的方法论来自教授团队与专业研发团队长期从事高质量写作、审稿和交付的经验积累，不是把几个“去 AI 痕迹”词表拼在一起。ReviewWrite 关注的是主张是否有证据、限定条件是否完整、体裁是否合适，以及修改后能否真正交付。

来源透明度：只审计，不清除

ReviewWrite 可以检查文件中可观察的 Content Credentials、C2PA/JUMBF 或 SynthID 相关线索，并定位可能影响交付完整性的 Unicode 格式控制字符。审计结果严格区分“已发现”“未发现”和“无法判断”，不把来源线索缺失解释成来源不存在；输入文件保持不变，不提供水印移除、来源伪装或检测器规避功能。

名称与调用：短，但不混乱

• 技术 Skill ID：reviewwrite。支持 slash 调用的平台使用 /reviewwrite。
• 产品名：审写 · ReviewWrite。
• 自然语言简称：审写、ReviewWrite。它们只用于支持自然语言触发的宿主，不是第二份可安装 Skill，也不保证成为 slash alias。
• 不要并行安装旧 review-write 和新 reviewwrite；安装器发现旧目录时会拒绝并行安装，先报告路径，再由用户决定迁移或删除。

想了解不同模型常见 AI 味、普通 humanizer 的风险，以及如何把“去 AI 痕迹”前移到写作契约，见 AI 写作痕迹与 Humanizer 策略。

一句话交给智能体安装

复制下面代码块中的内容：

请从官方仓库 https://github.com/songhai-dg/review-write 安装 ReviewWrite Skill；如当前网络无法访问 GitHub，使用同步镜像 https://gitee.com/cufe01/songhai-dg。如果当前智能体已经安装，不要重复安装，也不要搜索或替换成名称相近的其他技能。请使用所在平台自己的 Skill 安装或导入机制；如果仓库已在当前目录，直接以当前目录为安装来源。平台不支持、没有权限或无法访问来源时，说明限制并停止。

国内 SkillHub 入口：审写 · ReviewWrite。

正式版本会自动生成一份分发包，统一提供 GitHub、Gitee、SkillHub 和其他 Agent Skills 目录所需的安装指令、版本信息与宣传文案，避免各平台出现不同版本。分发包说明见 distribution/README.md。

它主要解决：

• 提示词、推理过程、工具调用和编辑指令混入正式正文；
• 表面“去 AI 腔”导致事实、数字、引用、限定条件和作者声音漂移；
• 用一套通用规则误改论文、基金、公众号、政策、公文等不同体裁；
• 把中文规则直接翻译到英文，或忽略本语种、地区和机构的表达规范。

核心闭环

Understand → Review → Plan → Rewrite → Verify → Evolve
理解文本      独立评审    约束修改     生成正文      双重复核      保留经验

ReviewWrite 严格分离四个写作表面：评审报告、修改计划、正式正文和复核报告。只有正式正文可以成为对外交付文本。

核心能力

• 适配论文、基金、政策、报告、公文、备忘录、营销和双语写作；
• 0.6.0 增加跨语言专业审写：中英互译时保护数字、引用、术语、主张、限定条件和责任边界，并按目标地区、体裁和话语共同体独立审校；
• 支持超长文档审写：对十万字及更长文本进行章节优先分块，使用一次性全文行索引定位发现项、中文数字和标题，并限制结果体积，避免重复扫描和把片段检查误报成全文完成；
• 根据语言、读者和真实体裁选择表达，不强套模板；
• 可作为宿主智能体的写作质量控制层：风险路由负责决定是否接入，可执行门禁负责抽取、严格终检和“失败时不输出正文”，问题最多回改两轮后再交付；
• 增加技术解读/产业评论路由：检查模型、推理、设备、内存和性能主张中的叠加式模板信号与技术口径；
• 默认保护数字、引用、专名、义务和限定条件；
• 将评审、修改稿和复核分开，并对正式正文做泄漏预检；
• 泄漏预检支持体裁与语境感知：AI 安全论文讨论 system prompt、公文使用结构化条目等被授权情形，用 --context/--genre 声明后降级或放宽，避免正当写作被误判；
• 识别空泛强化和公式化转折，但必须按体裁、上下文与证据判断，不使用生硬禁词表；
• 识别技术评论中的叠加式模板信号，以及公众号文章中反复使用“第一、第二、第三”的机械枚举；
• 识别“值得注意的是”“总的来说”、无来源“研究表明”、连续“我们可以看到……这说明……”、翻译腔和对称句式堆叠；
• 检查 Markdown 标题末尾误带句号的情况；正文中的正常句号不触发此项，问号、感叹号和版本号标题保留；
• 对 DOCX/PPTX 提供可选的只读字体与渲染质检：检查中英文字体声明、主题/继承不确定项、profile 不匹配和目标字体库存；不静默改文件；
• 对超长 UTF-8 文本提供分块预检和全局一致性索引；明确报告覆盖范围，不把部分审核声称为全文完成。

# 结构化响应通过后，只输出不带标签的正文；失败时不输出正文
python3 scripts/reviewwrite_gate.py path/to/response.txt

AI 风格与过程残留审查清单

下表列的是可观察的文本信号和交付风险，不是“AI 生成概率”，也不据此判断作者身份。单个词语通常不会触发处理；技能结合体裁、重复频率、证据和段落功能判断。

覆盖范围：中文 zh-CN 与通用专业英语；论文、基金、公众号、技术解读、政策、公文、报告、备忘录、营销和双语文本。规则输出评审信号，不生成“人味分数”，也不承诺规避任何检测器。

四个内容案例与一项长文回归

四份合成案例直接展示问题稿、修改稿和验证结果：严格预检由 14 个硬失败 / 23 个警告 降至 0 / 0；38 项受保护字面量（数字、日期、角色与限定词）通过逐字回归检查。查看完整复现报告。另有长文回归测试：对 324,015 个字符切分为 55 个块，故意放入一处模型自述，预检将其定位到全文第 9,001 行；同时验证数字索引和重复段落索引。查看测试代码。

安全与来源

发布包包含 Skill、参考资料、一个配置结构样例、本地文本/Office QA 脚本和可选更新器。默认审写路径不联网、不读取凭证、不修改或上传输入正文与 Office 原件；只有明确启用更新检查/下载时，更新器才访问 Release、读取可选令牌或调用 gh。Office QA 只有在显式要求渲染时才向预览目录写入 PDF/PNG。安装前请核对官方 Release、SHA-256 与 artifact attestation，并在已有安装时拒绝覆盖未知目录。完整边界见 SECURITY.md。

Office 交付前质检：先审计，再决定是否修复

当用户交付 Word 或 PPT 时，ReviewWrite 可以在正文审写之后增加一个可选的 Office QA 环节。它不是“自动统一字体”：单位模板优先，默认只读审计，原件不会被覆盖。

# 默认只读结构审计；不生成渲染预览
python3 scripts/office_qa.py path/to/submission.docx --format json

# 模板已确认时再对照字体 profile，并要求完成渲染
python3 scripts/office_qa.py path/to/deck.pptx \
  --font-profile examples/office-qa/font-profile.example.json \
  --render required --output-dir /tmp/reviewwrite-preview --format json

报告会区分“字体缺失/回退风险”和“字符本身损坏”，不会把后者伪装成字体问题。结构审计或生成 PDF/PNG 预览都不等于视觉通过：必须逐页查看缺字、字体回退、截断、溢出和异常换行。没有确认模板、目标字体库存或视觉检查时，Skill 不会声称文件在所有设备上均可正常显示。详见 Office QA 说明 与 字体 profile 规范。

使用示例

使用 ReviewWrite 评审并修改这份政策简报。先给问题清单，再给修改稿，不能改变数字和政策口径。

使用 ReviewWrite 修改这篇英文论文摘要。保留术语、样本量、统计结果和引用，不要把编辑过程写进摘要。

多语言，不只是翻译

当前核心支持 zh-CN 和通用专业英语。语言能力使用三个坐标：语言、地区规范和话语共同体。英文不能默认等于美式商业写作，中文也不能默认增加政策口号、谦辞或成语。

0.6.0 开始支持跨语言专业审写：翻译前确认源语言、目标语言、地区规范、体裁、读者和术语；翻译后独立检查数字、引用、主张强度、限定条件、情态和责任边界。重点是把专业判断完整带入目标语言，不能用逐句替换、回译或流畅度代替准确性证明。详见跨语言专业审写。

新增语言必须经过独立语言包、locale-aware few-shot、事实保持回归和本语种专业 review，才能从 planned 进入 experimental，再进入 core。详见语言包规范。

few-shot 如何设计

few-shot 不是越大越好的模板库。每次最多选择三个：一个修复最高风险问题，一个校准体裁和 locale，必要时一个保持作者声音。授权、语言、体裁或文化语境匹配不足时宁可不用。示例只传递修改原则，不迁移事实、观点、专名或标志性句式。详见 few-shot policy。

检查与测试

python3 scripts/reviewwrite_lint.py path/to/draft.md
python3 scripts/reviewwrite_lint.py path/to/draft.md --strict
python3 scripts/reviewwrite_gate.py path/to/structured-response.txt
python3 scripts/validate_skill.py
python3 -m unittest discover -s tests -v
python3 scripts/package_skill.py

更新与发布

ReviewWrite 不随机检查，也不在每次调用时联网。新会话第一次调用只读取本地版本和缓存；缓存超过 24 小时时，当前写作完成后再联网刷新。默认只提示新版；可信环境可以自动下载 patch，但安装和会话刷新仍独立执行。minor/major 需要明确升级，宿主平台已有原生版本管理时优先使用平台渠道。

python3 scripts/reviewwrite_update.py status --format json  # 始终只读本地
python3 scripts/reviewwrite_update.py check --format json   # 仅在任务结束且缓存过期时

版本采用双轨制：有证据的小改进进入 edge patch；通过完整回归和人工 review 后提升为 stable minor；没有有效变化就不发布。详见运营策略和更新策略。

站在已有项目之上

ReviewWrite 吸收并重新组织 Stop Slop、Humanizer-zh 和 skill-deslop 的可复用能力，同时增加文档本体、体裁与语言路由、泄漏硬门、few-shot 选择和改后验证。精确来源、版本和许可见 NOTICE.md 与 sources.lock.json。

贡献与许可

欢迎提交体裁包、语言包、授权 few-shot、误报样例和事实保持回归。请先阅读 CONTRIBUTING.md。项目采用 MIT License。

English

Review first, then rewrite

审写 · ReviewWrite is built for professional Chinese and English writing: it identifies the goal and context, then reviews, rewrites, and verifies the text while preserving facts, citations, qualifications, and authorial voice. It does not promise detector evasion or add mistakes to imitate a human author.

Its central boundary is explicit: ReviewWrite does not invent experiences, data, sources, examples, emotions, or details to make text appear more human. When something cannot be confirmed, it preserves the qualification, marks it for review, or narrows the claim.

Its method is informed by a professor-led and professional product team with long-term experience in high-quality writing, review, and delivery. ReviewWrite is not a bundle of “AI trace removal” word lists; it evaluates claims, evidence, qualifications, genre fit, and whether the revised text is ready for real use.

Naming and invocation: short without ambiguity

• Technical Skill ID: reviewwrite. Use /reviewwrite on hosts that support slash invocation.
• Product name: 审写 · ReviewWrite.
• Natural-language names: 审写 and ReviewWrite. They are only natural-language triggers on hosts that support them, not separately installable Skills or guaranteed slash aliases.
• Do not install both legacy review-write and reviewwrite. The installer refuses a parallel install when it detects the old directory, then reports its path for a user-directed migration or removal.

For AI writing signals, common humanizer risks, and ReviewWrite's writing-before-rewriting strategy, see AI writing signals and humanizer strategy.

One-sentence agent installation

Copy the text in the code block:

Install ReviewWrite from the canonical repository https://github.com/songhai-dg/review-write. If GitHub is unavailable on the current network, use the synchronized Gitee mirror https://gitee.com/cufe01/songhai-dg. If the current agent already has it, do not install it again or search for a similarly named replacement. Use the current platform's native Skill installation or import flow; if the repository is already in the current directory, use that directory as the installation source. If the platform is unsupported, permissions are unavailable, or the source cannot be reached, state the limitation and stop.

It is not an AI-detector bypass tool; it is delivery-quality control for professional writing.

It never trades hallucinations for a more “human” tone: no invented experience, data, sources, cases, or details.

It addresses four recurring failures:

• prompts, reasoning, tool calls, or editorial instructions leaking into the deliverable;
• surface-level “humanization” changing facts, numbers, citations, qualifications, or authorial voice;
• one generic style rule being applied across papers, grants, public articles, policies, and official documents;
• translating rules between languages without respecting locale, discourse community, or institutional convention.

Core loop

Understand → Review → Plan → Rewrite → Verify → Evolve

ReviewWrite separates four writing surfaces: the review report, revision plan, deliverable body, and verification report. Only the deliverable body belongs in the publication-ready text.

Core capabilities

• genre-aware Chinese and English writing support;
• cross-language professional review between Chinese and English, preserving terminology, claims, qualifications, and responsibility boundaries across translation;
• long-document review for 100k+ characters: heading-aware chunking, one reusable line-offset index, bounded output, global locations, Chinese-context number/heading indexes, term variants, and duplicate-paragraph signals;
• technical commentary review for model, inference, device, memory, and performance claims, including composite template signals and scope checks;
• context-sensitive routing by language, audience, locale, and discourse community;
• an executable delivery gate that extracts the unique body, runs strict checks, emits body-only output on success, and emits no body on failure;
• protection for facts, numbers, citations, names, obligations, and qualifications;
• separated review, revision, deliverable, and verification surfaces with leakage checks;
• genre- and context-aware preflight: authorized cases such as an AI-safety paper discussing system prompt, or an official document using structured items, are declared with --context/--genre and downgraded or relaxed instead of hard-failing legitimate writing;
• contextual checks for empty intensifiers and formulaic pivots, rather than a rigid blacklist of “AI words”;
• composite checks for stacked technical-commentary signals, vague authority, inference narration, translationese, symmetric structures, and repetitive sentence-initial enumeration, with genre-aware exemptions;
• optional audit-only DOCX/PPTX delivery QA for Chinese/Latin font declarations, theme/inheritance uncertainty, confirmed-profile mismatches, target font inventories, and a render gate; it never silently rewrites a source Office file.

python3 scripts/reviewwrite_gate.py path/to/response.txt

AI-style and process-residue coverage

These are observable writing and delivery signals, not an authorship classifier. A single word is rarely enough; ReviewWrite combines genre, recurrence, evidence, and paragraph function.

Coverage includes zh-CN and general professional English across papers, grants, public articles, technical commentary, policy, official documents, reports, memos, marketing, and bilingual writing. Findings are review signals, not a “human score” or detector-evasion promise.

Four reproducible cases

Four synthetic cases show the draft, revision, and verification result. Strict preflight moved from 14 hard failures / 23 warnings to 0 / 0; a literal regression check retained 38 protected items (numbers, dates, roles, and qualifiers). A long-document regression also processes 324,015 characters in 55 chunks, maps an intentional model self-description to line 9,001, and checks number and duplicate-paragraph indexes. Read the test code.

Security and provenance

The release bundle contains the Skill, references, one configuration-structure example, local text/Office QA/provenance-audit scripts, and an optional updater. The default writing path stays local: it does not read credentials, modify or upload input prose or source Office files. The provenance audit is read-only: it reports observable Content Credentials/C2PA/JUMBF or SynthID references and Unicode format controls, distinguishes detected, referenced, and indeterminate states, and never removes or disguises source signals. Only an explicitly enabled update check contacts a Release source; only an explicitly requested Office render writes PDF/PNG previews to an output directory. Verify the official Release, SHA-256, and artifact attestation before installation, and never overwrite an unknown existing directory. See Security and provenance-integrity for the full boundary.

Office delivery QA: audit before any repair

When a Word document or PowerPoint deck is part of the deliverable, ReviewWrite can add an optional Office QA pass after text review. It is not an “apply one font everywhere” button: the confirmed institutional template wins, the default is read-only, and the source file is never overwritten.

# Default read-only structural audit; it does not generate a render preview.
python3 scripts/office_qa.py path/to/submission.docx --format json

# Use a confirmed profile and require that a render can be produced.
python3 scripts/office_qa.py path/to/deck.pptx \
  --font-profile examples/office-qa/font-profile.example.json \
  --render required --output-dir /tmp/reviewwrite-preview --format json

The report separates font-availability/fallback risk from damaged characters. A structural pass or generated PDF/PNG is not visual approval: inspect every page or slide for missing glyphs, fallback, clipping, overflow, and unexpected wrapping. Without a confirmed template, target font inventory, and visual inspection, ReviewWrite does not claim that a file will display correctly on every recipient device. See Office QA and the font-profile specification.

Usage examples

Use ReviewWrite to review and revise this policy brief. List the issues first, then provide the revision. Preserve every number and policy qualification.

Use ReviewWrite to revise this academic abstract. Preserve terminology, sample size, statistical results, citations, and the author's English variant. Do not leak editorial process into the abstract.

Multilingual by design

The current core supports zh-CN and general professional English. Language behavior is modeled through three coordinates: language, locale, and discourse community. English is not assumed to mean US business prose, and Chinese is not assumed to require slogans, honorific padding, or idioms.

Version 0.6.0 adds cross-language professional review between Chinese and English: confirm source and target locale, genre, audience, and terminology before translation, then independently check numbers, citations, claim strength, qualifications, modality, and responsibility boundaries. It is not word substitution, and fluency or back-translation is not proof of professional accuracy. See cross-language professional review.

A new language must pass a dedicated language pack, locale-aware few-shots, fact-preservation regressions, and professional native-language review before moving from planned to experimental and then core. See the language-pack specification.

Few-shot design

Few-shots are not a growing template dump. A task may select up to three examples: one for its highest-risk failure, one for genre and locale, and, when authorized, one for author voice. If authorization, language, genre, or cultural context does not match, ReviewWrite uses no example. Examples teach transformations; they do not transfer facts, opinions, names, or signature phrasing. See the few-shot policy.

Validation and tests

python3 scripts/reviewwrite_lint.py path/to/draft.md
python3 scripts/reviewwrite_lint.py path/to/draft.md --strict
python3 scripts/reviewwrite_gate.py path/to/structured-response.txt
python3 scripts/validate_skill.py
python3 -m unittest discover -s tests -v
python3 scripts/package_skill.py

Updates and release cadence

ReviewWrite uses a deterministic update policy: the first invocation in a new session reads only local version/cache state; a due network refresh runs after the active writing task and at most once per 24 hours. It does not use random checks or replace a Skill during an active task. The default is notify; trusted environments may auto-download patch releases, while installation and session refresh remain explicit. A host platform's native update channel takes precedence when available.

Qualified small improvements enter the edge channel as patches. A stable minor is promoted only after full regressions and human review. No meaningful improvement means no release. See operations and the update policy.

Built on existing work

ReviewWrite learns from and reorganizes reusable ideas from Stop Slop, Humanizer-zh, and skill-deslop. It adds a writing ontology, genre and language routing, a delivery-leakage gate, few-shot selection, and post-rewrite verification. Exact sources, revisions, and licenses are recorded in NOTICE.md and sources.lock.json.

Contributing and license

Contributions of genre packs, language packs, authorized few-shots, false-positive examples, and fact-preservation regressions are welcome. Read CONTRIBUTING.md first. ReviewWrite is released under the MIT License.