Docs · Document Creation, Collaboration, Versioning, and Review

1. Introduction

1.1. Background

Documents are the carrier of organizational knowledge and the medium of organizational collaboration. In knowledge work, document-related activities — drafting, rewriting, reviewing, retrieving, and archiving — consume a large share of working hours, which is also the area where large language models were first validated at scale.

But document scenarios place entirely different demands on AI than Q&A scenarios:

Therefore, AI adoption in the document scenario cannot stop at "plugging in a chat box"; it must be built into a traceable, reviewable, versionable document pipeline.

1.2. Definition

Docs (document creation and collaboration): an activity that, taking documents as its object, engineering-manages the entire process of material collection, structure extraction, drafting, fact-checking with citation tracing, human review, versioning and publishing, and feedback looping.

On the technical foundation side, the official UK government report makes this explicit — the tool used in its trial "uses a method called retrieval augmented generation (RAG) to search and extract relevant information to a user's query". This shows that: the factual reliability of a document assistant comes from RAG, not from model memory.

On the standards side, the document-collaboration field has no AI-specific mandatory standard (none was found in the search). The adjacent norms that can be cited are ISO 30401:2018 Clause 7.5 "Documented information" and Clause 9.1 "Monitoring, measurement, analysis and evaluation" — the former requires documented information to be controlled, the latter requires monitoring and measuring knowledge management performance.

1.3. Positioning in the AI Harness System

图 1-1｜Docs 在 Harness 六层能力模型中的定位

数据来源：基于本文分析绘制的示意图。

Docs is positioned in the Harness six-layer capability model as: primary layer L1 context engineering, secondary layers L3 orchestration & control and L6 governance & security (this mapping is the analysis of this document).

Bottleneck layer: Docs' bottleneck lies at the boundary of L1 and L6.

• On the L1 side: document creation needs long context and multi-source material, constrained by the "Lost in the Middle" positional effect — the quality of material assembly determines the quality of the finished document;
• On the L6 side: the document assistant adopts end-user permissions; if permissions are misconfigured, the AI becomes an "efficient unauthorized-access amplifier".

First-hand evidence of the governance risk: The official UK government report states verbatim — "The tool adopts the permissions of the end user and will only retrieve documents that a user could normally access." It also warns: "Copilot's ability to search and query using a user's access may highlight issues when users have access to files they shouldn't." This is the most typical L6 risk for document-collaboration AI and should be listed separately in every document-type adoption plan.

Value: Docs is the direction with the highest penetration of AI Harness into knowledge workers' daily work, and the touchstone for verifying whether "context engineering + governance" truly takes effect.

2. Glossary

3. Case Studies

3.1. UK Government Cross-government M365 Copilot Trial

3.1.1. Background

The UK government wanted to verify the actual impact of generative AI on the productivity of civil servants in a large-scale, multi-department real work environment, rather than relying on vendor claims or small-scale pilots.

3.1.2. Approach

Per the official GOV.UK publication (first-hand government report):

• Time frame: 2024-09-30 to 2024-12-31.
• Scale: covering 20,000 civil servants across 12 agencies — DWP, HMRC, the Home Office, the Ministry of Justice, DESNZ, DSIT, FCDO, DEFRA, DBT, the Welsh Government, ONS, and Companies House.
• Technical mechanism: the tool uses the RAG method to retrieve and extract relevant information; it adopts end-user permissions and only retrieves documents that the user could normally access.
• Measurement method: self-reported time saved + system activity data + satisfaction and recommendation surveys.

3.1.3. Results

Per the official GOV.UK report (R1):

Limitations also disclosed officially (R1): complex, judgment-intensive, or data-intensive tasks performed poorly; sensitive matters such as policy and personnel require human oversight; Word/Outlook usage declined slightly in the later phase (novelty effect).

3.2. Eaton: Automated Generation of 1,000 SOPs

3.2.1. Background

Eaton is a global power-management company that needs to write a large volume of documentation for standard operating procedures (SOPs). SOPs are highly repetitive in structure, strictly formatted, and depend on existing standards for their content — a typical "high-repetition document".

3.2.2. Approach

Per Microsoft's official customer case-summary page, Eaton used M365 Copilot to generate SOPs automatically, using existing standards and templates as the corpus, with AI producing the draft and humans confirming it.

3.2.3. Results

Per Microsoft's official customer case-summary page (vendor customer-account self-reports): Eaton automatically generated 1,000 SOPs, with per-document creation time dropping from 1 hour to 10 minutes.

Lateral reference from the same batch of cases (all vendor customer-account self-reports):

Note: most figures on Microsoft's official customer case-summary page are not accompanied by methodology notes, so they must be cited as "vendor customer-account self-reports" when referenced.

3.3. Jinjiang Gas Power: AI Knowledge Platform and Document Pipeline Governance

3.3.1. Background

Jinjiang Gas Power has accumulated a large volume of unstructured documents — procedures, drawing notes, and maintenance records — and traditional retrieval methods could not support frontline workers' need for rapid lookup, with no unified document-governance process in place.

3.3.2. Approach

Per the report from the Fujian Electric Power Industry Association Information Network:

• Used RAG technology to parse unstructured documents;
• Repeatedly tuned the generic parsing mode and chunk-token parameters — consistent with the engineering conclusions in the RAG direction: chunking parameters must be determined empirically against your own gold set, never copied from defaults;
• Established a standardized "publish — approve — publish to shelf" process, combined with strict permission management;
• Adopted a visual orchestration engine supporting drag-and-drop composition of nodes such as "generate answer" and "knowledge retrieval";
• Advocated an operating model of "everyone participates, crowd-sourced shared intelligence".

3.3.3. Results

This source did not disclose quantifiable benefit figures (R2). Its value lies mainly in corroborating a standard structure for the document pipeline: parsing and chunking → retrieval and generation → approval and publishing → permission-controlled usage → feedback loop. This structure matches the pipeline definition given in Section 1.2 of this document.

Another Chinese case for reference: Xingye Digital Industry Group launched an intelligent office assistant based on DeepSeek-R1 in February 2025, covering three functions — official-document writing, project-progress analysis, and business-status analysis. According to its public materials, overall efficiency rose by 40%+ and management decision-making efficiency by 50% (R2, republished on a local-government website).

4. Practice Standards

4.1. AGENTS.md Specification

The following is the complete, copyable industry-standard AGENTS.md for the Docs direction, reflecting the dedicated toolchain of document-collaboration systems, versioning and diffing, citation locators, glossary and templates, review trails, and permission inheritance.

# AGENTS.md —— Docs（文档创作、协作、版本与审阅）

## 角色与边界
- 你是文档创作与审阅智能体，负责素材归集、结构抽取、草稿生成、事实核查、引用溯源与版本差异说明。
- 你可以：检索语料与既有文档、生成草稿与改写建议、生成引用表、生成版本差异清单、
  按模板与术语表做格式与用词校验、生成审阅要点清单。
- 你不可以：直接定稿发布、代替审阅人做合规判断、使用你无权访问的文档作为素材、
  删除或覆盖历史版本、把未审阅的内容标记为"已确认"。
- 文档助手采用终端用户权限：只能检索调用者本可访问的文档；不得使用高权限账号绕过。

## 环境假设
- 运行环境提供：文档库（含版本标识）、全文与向量检索、版本快照与差异比对、
  引用定位器、术语表、模板库、审阅与审批流、审计日志。
- 每份文档具备稳定 doc_id 与版本号；无版本标识的文档视为不可引用。
- 检索服务在召回阶段即完成权限过滤（permission-aware retrieval）。

## 上下文加载顺序（Context Budget）
1. 任务契约：文档类型、目标读者、篇幅、模板、验收标准（常驻，不压缩）
2. 模板与术语表（决定结构与用词，必须加载）
3. 素材检索结果（混合召回 + 权限过滤 + 重排，附 doc_id + 版本号 + 片段定位）
4. 既有同类型范文（1~2 篇，按需）
5. 历史版本与审阅意见（改写任务时加载）
6. 参考规范与标准摘录（按需）

## 文档流水线（SOP）
1. 素材归集：确定语料白名单与版本；记录来源清单。
2. 结构抽取：按模板抽取章节骨架；标注每节的素材来源。
3. 起草：逐节生成，每条事实性陈述挂来源标识。
4. 事实核查：数值、日期、编号、条款号、人名与来源逐字符比对。
5. 引用溯源：生成引用表（doc_id / 版本 / 片段定位 / 用于哪一节）。
6. 人工审阅：提交草稿 + 引用表 + 自检表；审阅意见结构化留痕。
7. 版本与发布：写版本号与变更日志；发布须经审批。
8. 反馈回流：人工修改与失败用例进入评测集与术语表更新清单。

## 工具契约
- 检索工具：向量 + BM25 混合召回 + RRF 融合 + 重排；结果已过权限过滤。
- 文档读写：写操作只写草稿；定稿与发布须走审批流。
- 版本工具：每次修改生成快照；差异比对输出结构化清单。
- 引用定位器：校验来源标识是否能在原文定位；定位失败判定引用不合格。
- 术语与模板工具：用词与结构必须过校验；术语冲突时以组织术语表为准。
- 脚本工具：解析、格式转换、批量校验、跑分必须调用 scripts/。

## 验证与证据要求
- 证据包：来源清单、检索记录、引用表、自检表、版本差异清单。
- 引用完整性：事实性陈述 100% 有来源，且来源可定位。
- 数值一致：数值、日期、编号与来源逐字符一致。
- 术语一致：与组织术语表一致；不一致项列出待确认清单。
- 密级一致：产出物密级不低于引用素材的最高密级。

## 失败与升级策略
| 失败 | 处置 |
|---|---|
| 素材零命中 | 输出"未检索到可靠依据"，请人工补充语料；禁止凭记忆写作 |
| 引用无法定位 | 剔除或重检索；不得保留不可定位的引用 |
| 事实一致性不达标 | 回退上一策略版本，重跑评测集 |
| 术语冲突 | 列出冲突项交人工裁决，不自行统一 |
| 疑似越权素材 | 立即停止，记录审计事件，升级至权限管理员 |
| 审核驳回 | 记录驳回理由，进入负样本集 |
- 每个循环必须有步数上限与 Token 预算上限；超限即停并升级。

## 安全与合规红线
- 不得使用调用者无权访问的文档作为素材。
- 不得把高密级素材写入低密级产出物。
- 不得删除、覆盖或篡改历史版本与审阅记录。
- 定稿与发布必须经人工审批。
- 涉及政策、人事、法律、财务等敏感内容必须人工监督。

## 禁止事项
- 禁止编造 doc_id、版本号、条款号、数值、URL。
- 禁止在无素材支撑时输出事实性陈述。
- 禁止使用 XX / XXX / ___ 等非标准占位符（统一用 [待填写] / [待核实]）。
- 禁止把厂商宣称或二手转述升格为一手事实。
- 禁止覆盖历史版本；禁止跳过审阅直接发布。
- 禁止使用 emoji 与署名。

## 输出格式
- 文档类产出：Front Matter → 标题 → 创建时间 → 目录 → 正文（1 / 1.1 / 1.1.1）→ 信息缺口声明 → 参考资料。
- 引用表：doc_id / 版本 / 片段定位 / 用于哪一节。
- 版本差异清单：新增 / 删除 / 修改，逐项说明。
- 数值带单位；范围用 ~ 连接；中文全角标点；中英文之间加空格。

## 评估与自检
- 九项自检：来源可回溯 / 无无源断言 / 数值一致 / 无非标占位符 / 数量与表格一致 /
  编号可核实 / 密级正确 / 审阅节点齐备 / 信息缺口已声明。
- 每次大版本更新后重跑事实一致性评测；指标回退视为缺陷。

4.2. SKILL.md Specification

The following is the complete, copyable industry-standard SKILL.md for the Docs direction.

---
name: doc-drafting
description: 基于组织语料起草、改写、审校文档，并生成引用表与版本差异清单。当用户要求"起草一份 XX 文档""按模板改写""核对这份材料的事实与引用""生成版本差异说明""把这份材料整理成标准格式"时触发。
version: 1.0
created: 2026-09-12
---

# 文档起草与审校（Doc Drafting & Review）

## 适用场景
- 按模板起草新文档（制度、规程、方案、报告、SOP、对外材料）。
- 对既有文档做改写、压缩、扩写或格式规范化。
- 对草稿做事实核查、引用溯源、术语与格式校验。
- 生成两个版本之间的差异说明与变更日志。

## 前置条件
- 已加载本方向 AGENTS.md，且已确定文档类型、目标读者与模板。
- 已确定语料白名单与版本；素材可读且已过权限过滤。
- 已加载组织术语表与模板库。
- 已知产出物密级与发布审批路径。

## 输入
- 文档类型、目标读者、篇幅要求、模板（或结构要求）
- 语料白名单（doc_id 清单 + 版本号）
- 可选：既有草稿或历史版本、审阅意见
- 可选：不得引用的范围（如过期版本、外部资料）

## 输出
- 文档草稿（按模板结构，逐节标注素材来源）
- 引用表：doc_id / 版本 / 片段定位 / 用于哪一节
- 自检表：事实一致性、引用完整性、术语一致性、格式合规性、密级
- 版本差异清单（改写任务）：新增 / 删除 / 修改 逐项说明
- 待确认清单：术语冲突、数值存疑、素材缺失项

## 执行步骤
1. 加载模板与术语表，确定章节骨架。
2. 按章节检索素材：向量 + BM25 → 权限过滤 → RRF 融合 → 重排 → Top-K。
3. 逐节起草；每条事实性陈述挂来源标识；数值与编号逐字符比对。
4. 素材零命中时输出"未检索到可靠依据"，不凭记忆写作。
5. 生成引用表，用引用定位器校验每条来源可回溯。
6. 自检：事实一致性、引用完整性、术语一致性、格式合规、密级。
7. 输出草稿 + 引用表 + 自检表 + 待确认清单，提交人工审阅。
8. 审阅通过后写版本号与变更日志；发布走审批流程。
9. 把人工修改与驳回理由回流进评测集与术语表更新清单。

## 质量标准（DoD）
- 事实性陈述 100% 附来源标识且可定位。
- 数值、日期、编号与来源逐字符一致。
- 术语与组织术语表一致；结构符合模板。
- 无 XX / XXX / ___ 等非标准占位符（只允许 [待填写] / [待核实]）。
- 密级正确；未使用越权素材；定稿与发布经人工审批。
- 证据包（来源清单、引用表、自检表、差异清单）齐备。

## 常见失败与处理
| 失败 | 根因 | 处置 |
|---|---|---|
| 内容空洞 | 素材召回不足 | 放宽召回、改写查询、补充语料；仍不足则标注 [待填写] |
| 事实错误 | 生成脱离上下文 | 提高忠实度阈值；逐句挂载来源；回退上一策略 |
| 结构跑偏 | 模板未加载或被压缩 | 把模板与术语表设为不可压缩的常驻上下文 |
| 引用无法定位 | 来源缺版本或片段定位 | 剔除或重检索 |
| 术语混乱 | 多源术语不一致 | 列出冲突项交人工裁决，不自行统一 |
| 篇幅失控 | 无字数约束与预算 | 设定章节字数配额与总 Token 预算 |

## 示例
用户请求：按公司模板起草一份《数据分级管理办法》，素材为 info-sec-201 v2.1 与 data-policy-003 v1.0。
执行：
1. 加载制度类模板与术语表（"数据分级""密级""数据owner"等）。
2. 检索两份素材的相关章节，权限过滤后重排，取 Top-10 片段。
3. 按模板逐节起草：总则 / 分级标准 / 责任分工 / 审批与变更 / 监督与问责。
4. 分级标准表中的每一级定义挂 [doc_id=info-sec-201][v2.1][§3.2]。
5. 生成引用表（12 条）+ 自检表 + 待确认清单（"重要数据"定义两版不一致）。
6. 提交法务与信息安全负责人审阅；通过后写 v0.1 与变更日志。
约束：不得引用 v1.0 已废止条款；不得在没有素材的"监督与问责"章节编造处罚条款，
应标注 [待填写] 并在待确认清单中说明。

4.3. Implementation Checklist

5. Summary

The core judgment of the Docs direction can be summarized in one sentence: the value of a document assistant lies not in "writing fast", but in "writing traceably".

First-hand evidence supports the reality of the benefits: the UK government's 20,000-person cross-government trial showed an average saving of 26 minutes per person per day, 24 minutes saved per document drafted, and 82% of users unwilling to return to a state without Copilot (official GOV.UK report, R1). But the same report also provides equally important boundaries: complex, judgment-intensive, or data-intensive tasks performed poorly; sensitive matters require human oversight; and a novelty effect exists.

At the engineering level, adoption of Docs should focus on four main threads:

1. Pipeline-ize: break document production into "material collection → structure extraction → drafting → checking → review → publishing → feedback loop", with each stage having an artifact and a gate.
2. Citation tracing is a hard requirement: errors in documents get copied into multiple downstream materials, so every factual statement must be traceable.
3. Permission inheritance is a risk source: adopting end-user permissions amplifies existing permission-configuration errors, so a permission audit must be done before launch, with filtering applied at the retrieval layer.
4. Chunking and retrieval parameters must be determined empirically: the experience in the Jinjiang Gas Power case of "repeatedly tuning the generic parsing mode and chunk-token parameters" is fully consistent with the engineering conclusions in the RAG direction.

Information Gap Statement

6. References

1. Microsoft 365 Copilot Experiment: Cross-Government Findings Report — GOV.UK, 2025. https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html
2. Microsoft AI Customer Stories roundup — Microsoft. https://aka.ms/AIcustomersBlog
3. Jinjiang Gas Power AI knowledge platform — Fujian Electric Power Industry Association Information Network. http://www.fepsa.com.cn/mb_info.aspx?newsid=76556
4. Xingye Digital Industry Group intelligent office assistant — Suining People's Government website. https://suining.gov.cn/gongkai/show/91cade087efda28e2140a0c79ef7d243.html
5. ISO 30401:2018 Knowledge management systems — Requirements — ISO. https://www.iso.org/standard/68683.html
6. Introducing Contextual Retrieval — Anthropic. https://www.anthropic.com/news/contextual-retrieval
7. AGENTS.md — A simple, open format for guiding coding agents. https://agents.md/
8. Equipping agents for the real world with Agent Skills — Anthropic Engineering. https://anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
9. Model Context Protocol Specification Changelog — MCP. https://modelcontextprotocol.io/specification/2025-06-18/changelog
10. WPS 365 × Chery (second-hand source, data to be verified) — Sohu. https://m.sohu.com/a/1001560351_161795