Add sample-driven low AIGC workflow

Author: Jonnys-Li

README.md

@@ -33,6 +33,7 @@ Use $software-thesis-docx to turn my project repo into a structured thesis manif
 Use $software-thesis-docx to read my school Word template, extract a custom style preset, and build the thesis in that format.
 Use $software-thesis-docx to generate Mermaid architecture and sequence diagrams for my thesis from the repository structure.
 Use $software-thesis-docx to run an AIGC risk check on my thesis DOCX and only rewrite the flagged single-run paragraphs if I approve it.
+Use $software-thesis-docx to lower AIGC for my thesis and switch to explicit_low_aigc mode only for the paragraphs I authorize.
 ```
 
 ## What It Adds
@@ -42,7 +43,7 @@ Use $software-thesis-docx to run an AIGC risk check on my thesis DOCX and only r
 - Manifest-driven DOCX build with optional `formatting` config
 - Mermaid request contracts for architecture, sequence, ER, state, gantt, class, and mind-map diagrams
 - Optional rigorous-writing subagent mode for complex thesis tasks, default off
-- Local AIGC risk checking and conservative paragraph-level reduction workflow, default off
+- Sample-driven AIGC risk checking and dual-profile paragraph-level reduction workflow, default off and defaulting to `academic_safe`
 
 ## Compatibility
 
@@ -54,14 +55,16 @@ Use $software-thesis-docx to run an AIGC risk check on my thesis DOCX and only r
 
 - Root-level distribution files: `README`, release notes, and one-click installers
 - The actual reusable skill at `skills/software-thesis-docx/`
-- Five reusable DOCX-oriented scripts:
+- Six user-facing scripts and one shared helper:
   - `build_docx_from_manifest.py`
   - `extract_docx_style_preset.py`
   - `replace_images_by_caption.py`
   - `rewrite_paragraphs.py`
   - `check_aigc_risk.py`
+  - `rewrite_low_aigc_docx.py`
+  - `aigc_utils.py`
 - Public examples for manifests, workflow options, Mermaid requests, image maps, and paragraph rewrites
-- Reference docs for formatting presets, Mermaid planning, option intake, AIGC review, and repo-to-thesis workflow
+- Reference docs for formatting presets, Mermaid planning, option intake, AIGC review, a low-AIGC playbook, and repo-to-thesis workflow
 
 ## Repository Layout
 
@@ -95,6 +98,7 @@ Inside that folder you will find:
 ## Documentation
 
 - [Codex quick start](docs/codex-quickstart.md)
+- [v0.4.0 release notes](docs/releases/v0.4.0.md)
 - [v0.3.0 release notes](docs/releases/v0.3.0.md)
 - [v0.2.0 release notes](docs/releases/v0.2.0.md)
 - [v0.1.0 release notes](docs/releases/v0.1.0.md)
@@ -107,13 +111,14 @@ Inside that folder you will find:
 - Prefer caption-based figure replacement over position-based image replacement
 - Prefer exact paragraph rewrites when Word layout fidelity matters
 - Keep Mermaid generation and AIGC review opt-in, not forced defaults
+- Keep `explicit_low_aigc` behind an explicit user request; otherwise stay in `academic_safe`
 
 ## Limits
 
 - Mermaid support currently generates code and file outputs, not rendered images or automatic DOCX insertion
 - The paragraph rewrite tool only handles exact full-paragraph matches
 - Mixed-format, multi-run paragraphs should be inspected before automated edits
-- The AIGC checker is a local heuristic pass, not a guarantee of any third-party score
+- The AIGC checker is a sample-driven local heuristic pass, not a guarantee of any third-party score
 - Institution-specific formatting rules may still require manifest or preset extensions
 
 ## Roadmap

README.zh-CN.md

@@ -33,6 +33,7 @@ Use $software-thesis-docx to turn my project repo into a structured thesis manif
 Use $software-thesis-docx to read my school Word template, extract a custom style preset, and build the thesis in that format.
 Use $software-thesis-docx to generate Mermaid architecture and sequence diagrams for my thesis from the repository structure.
 Use $software-thesis-docx to run an AIGC risk check on my thesis DOCX and only rewrite the flagged single-run paragraphs if I approve it.
+Use $software-thesis-docx to lower AIGC for my thesis and switch to explicit_low_aigc mode only for the authorized paragraphs.
 ```
 
 ## 这次新增了什么
@@ -42,7 +43,7 @@ Use $software-thesis-docx to run an AIGC risk check on my thesis DOCX and only r
 - 支持 `formatting` 配置的 manifest 驱动 DOCX 构建
 - 面向论文插图的 Mermaid 请求契约，支持架构图、时序图、ER 图、状态图、甘特图、类图和脑图
 - 可选的严谨写作 subagent 模式，默认关闭
-- 本地 AIGC 风险检查与保守的定向降重流程，默认关闭
+- 基于真实样本抽象出的 AIGC 风险检查与双模式降重流程，默认关闭且默认走 `academic_safe`
 
 ## 兼容性
 
@@ -54,14 +55,16 @@ Use $software-thesis-docx to run an AIGC risk check on my thesis DOCX and only r
 
 - 根目录分发文件：`README`、版本说明、一键安装脚本
 - 真实可复用的 skill：`skills/software-thesis-docx/`
-- 5 个可复用脚本：
+- 6 个用户可直接运行的脚本与 1 个共享 helper：
   - `build_docx_from_manifest.py`
   - `extract_docx_style_preset.py`
   - `replace_images_by_caption.py`
   - `rewrite_paragraphs.py`
   - `check_aigc_risk.py`
+  - `rewrite_low_aigc_docx.py`
+  - `aigc_utils.py`
 - 公开示例：manifest、workflow options、Mermaid 请求、图片映射、段落改写
-- 方法论文档：格式 preset、Mermaid、交互 intake、AIGC 评估、repo-to-thesis 工作流
+- 方法论文档：格式 preset、Mermaid、交互 intake、AIGC 评估、低 AIGC playbook、repo-to-thesis 工作流
 
 ## 仓库结构
 
@@ -95,6 +98,7 @@ https://github.com/Jonnys-Li/software-thesis-docx-skill/tree/main/skills/softwar
 ## 文档
 
 - [Codex 快速上手](docs/codex-quickstart.zh-CN.md)
+- [v0.4.0 版本说明](docs/releases/v0.4.0.md)
 - [v0.3.0 版本说明](docs/releases/v0.3.0.md)
 - [v0.2.0 版本说明](docs/releases/v0.2.0.md)
 - [v0.1.0 版本说明](docs/releases/v0.1.0.md)
@@ -107,13 +111,14 @@ https://github.com/Jonnys-Li/software-thesis-docx-skill/tree/main/skills/softwar
 - 图片替换优先按图注定位，不按“第几张图”这种脆弱逻辑
 - 文本改写优先整段精确匹配，减少对 Word 格式的破坏
 - Mermaid 与 AIGC 评估都是显式开关，不默认强制开启
+- 默认只做保守 AIGC 检查，只有用户明确提出“降低 AIGC”时才启用 `explicit_low_aigc`
 
 ## 使用边界
 
 - Mermaid 当前只负责生成代码或 `.mmd` 文件，不负责自动渲染图片或自动插入 DOCX
 - 段落改写脚本只适合“整段全文精确命中”的场景
 - 多 `run`、混合格式、修订痕迹较多的段落需要先人工检查
-- AIGC 检测是本地启发式评估，不承诺对齐第三方平台分值
+- AIGC 检测是样本驱动的本地启发式评估，不承诺对齐第三方平台分值
 - 学校格式要求如果更特殊，通常需要扩展 manifest 或 preset 逻辑
 
 ## 下一步

docs/codex-quickstart.md

@@ -47,7 +47,7 @@ The important part is that `SKILL.md` stays at the root of the installed skill.
 - custom style preset extraction from `.docx` templates
 - Mermaid planning contracts
 - optional subagent rigorous-writing mode
-- optional AIGC risk checking and conservative reduction workflow
+- optional AIGC risk checking and dual-profile reduction workflow
 
 ## 4. Use The Skill In Codex
 
@@ -63,6 +63,7 @@ Typical prompts:
 - `Use $software-thesis-docx to read my Word template, extract a style preset, and build the thesis in that format.`
 - `Use $software-thesis-docx to generate Mermaid flowchart and sequenceDiagram code for my thesis based on the repo architecture.`
 - `Use $software-thesis-docx to run an AIGC risk review on my thesis DOCX and only rewrite the flagged single-run paragraphs after showing me the report.`
+- `Use $software-thesis-docx to lower AIGC for my thesis, keep academic_safe by default, and only switch to explicit_low_aigc if I explicitly ask for it.`
 
 ## 5. Optional Dependency Step
 
@@ -109,6 +110,18 @@ python3 "$HOME/.codex/skills/software-thesis-docx/scripts/check_aigc_risk.py" \
   --output /tmp/aigc-risk-report.json
 ```
 
+Rewrite authorized paragraphs for lower AIGC risk:
+
+```bash
+python3 "$HOME/.codex/skills/software-thesis-docx/scripts/rewrite_low_aigc_docx.py" \
+  --input thesis.docx \
+  --report /tmp/aigc-risk-report.json \
+  --output /tmp/thesis-low-aigc.docx \
+  --pending-output /tmp/aigc-pending-review.json \
+  --profile academic_safe \
+  --normalize-typography
+```
+
 Replace images by caption:
 
 ```bash

docs/codex-quickstart.zh-CN.md

@@ -47,7 +47,7 @@ irm https://raw.githubusercontent.com/Jonnys-Li/software-thesis-docx-skill/main/
 - 从 `.docx` 模板抽取自定义 style preset
 - Mermaid 规划契约
 - 可选的严谨写作 subagent 模式
-- 可选的 AIGC 风险检查与保守降重流程
+- 可选的 AIGC 风险检查与双模式降重流程
 
 ## 4. 在 Codex 中使用
 
@@ -63,6 +63,7 @@ Use $software-thesis-docx to build my thesis DOCX workflow from my software repo
 - `Use $software-thesis-docx to read my Word template, extract a style preset, and build the thesis in that format.`
 - `Use $software-thesis-docx to generate Mermaid flowchart and sequenceDiagram code for my thesis based on the repo architecture.`
 - `Use $software-thesis-docx to run an AIGC risk review on my thesis DOCX and only rewrite the flagged single-run paragraphs after showing me the report.`
+- `Use $software-thesis-docx to lower AIGC for my thesis, keep academic_safe by default, and only switch to explicit_low_aigc if I explicitly ask for it.`
 
 ## 5. 可选依赖安装
 
@@ -109,6 +110,18 @@ python3 "$HOME/.codex/skills/software-thesis-docx/scripts/check_aigc_risk.py" \
   --output /tmp/aigc-risk-report.json
 ```
 
+按风险报告回写低 AIGC 版本：
+
+```bash
+python3 "$HOME/.codex/skills/software-thesis-docx/scripts/rewrite_low_aigc_docx.py" \
+  --input thesis.docx \
+  --report /tmp/aigc-risk-report.json \
+  --output /tmp/thesis-low-aigc.docx \
+  --pending-output /tmp/aigc-pending-review.json \
+  --profile academic_safe \
+  --normalize-typography
+```
+
 按图注替换图片：
 
 ```bash

docs/releases/v0.4.0.md

@@ -0,0 +1,31 @@
+# v0.4.0
+
+## Summary
+
+`v0.4.0` upgrades the AIGC capability from a basic heuristic checker into a sample-driven low-AIGC workflow.
+
+This release adds a reusable low-AIGC playbook, a stronger checker aligned to real before/after thesis samples, and a new rewrite script that can auto-apply authorized paragraph rewrites for single-run paragraphs.
+
+## Highlights
+
+- Sample-driven AIGC methodology in `references/low-aigc-playbook.md`
+- Stronger `check_aigc_risk.py` output:
+  - `rewrite_recipe`
+  - `typography_flags`
+  - denser academic-template detectors
+- New rewrite script:
+  - `rewrite_low_aigc_docx.py`
+- Shared AIGC utilities:
+  - `aigc_utils.py`
+- Dual rewrite profiles:
+  - `academic_safe`
+  - `explicit_low_aigc`
+- Updated workflow options:
+  - `aigc.rewrite_profile`
+  - `aigc.normalize_typography`
+
+## Scope Notes
+
+- `academic_safe` remains the default.
+- `explicit_low_aigc` is only intended for cases where the user explicitly asks to lower AIGC.
+- The checker remains local and heuristic. It is not presented as a guarantee of any third-party detector percentage.

skills/software-thesis-docx/SKILL.md

@@ -21,7 +21,7 @@ This skill is especially useful when the user needs one or more of the following
 - Mermaid code for architecture, flow, sequence, ER, state, schedule, or mind-map diagrams
 - caption-based figure replacement without breaking Word layout
 - citation cleanup or terminology cleanup in an existing `docx`
-- optional AIGC risk review before final delivery
+- optional AIGC risk review or low-AIGC rewrite before final delivery
 
 ## 1. Ground The Repository Or The DOCX
 
@@ -66,6 +66,11 @@ Default values:
 - `subagents.enabled = false`
 - `aigc.enabled = false`
 
+For AIGC:
+
+- keep `rewrite_profile = academic_safe` by default
+- only switch to `explicit_low_aigc` when the user explicitly asks to lower AIGC
+
 If the runtime supports structured input collection such as `request_user_input`, use it.
 If not, ask the same questions in plain conversation.
 
@@ -105,6 +110,7 @@ Supporting references:
 - [references/workflow.md](references/workflow.md)
 - [references/source-conventions.md](references/source-conventions.md)
 - [references/migration-notes.md](references/migration-notes.md)
+- [references/low-aigc-playbook.md](references/low-aigc-playbook.md)
 
 ## 6. Run The Right Script
 
@@ -159,6 +165,18 @@ python3 scripts/check_aigc_risk.py \
   --output /tmp/aigc-risk-report.json
 ```
 
+Rewrite authorized paragraphs for lower AIGC risk:
+
+```bash
+python3 scripts/rewrite_low_aigc_docx.py \
+  --input thesis.docx \
+  --report /tmp/aigc-risk-report.json \
+  --output thesis-low-aigc.docx \
+  --pending-output /tmp/aigc-pending-review.json \
+  --profile academic_safe \
+  --normalize-typography
+```
+
 ## 7. Mermaid Rules
 
 Mermaid generation is an orchestration capability, not a required rendering pipeline.
@@ -204,11 +222,13 @@ Safe behavior:
 
 - scan first
 - rewrite only authorized paragraphs
+- keep `academic_safe` as the default rewrite profile
+- use `explicit_low_aigc` only when the user explicitly asks to lower AIGC
 - prefer more concrete, evidence-linked language
-- reuse `rewrite_paragraphs.py` only for single-run exact-match paragraphs
-- stop for manual confirmation on multi-run or mixed-format paragraphs
+- use `rewrite_low_aigc_docx.py` for authorized single-run paragraphs
+- stop for manual confirmation on multi-run or mixed-format paragraphs and inspect the pending JSON output
 
-Read [references/aigc.md](references/aigc.md) when the user asks for detection or auto-reduction.
+Read [references/aigc.md](references/aigc.md) and [references/low-aigc-playbook.md](references/low-aigc-playbook.md) when the user asks for detection or low-AIGC rewriting.
 
 ## 10. Quality Gates
 

skills/software-thesis-docx/agents/openai.yaml

@@ -1,4 +1,4 @@
 interface:
   display_name: "Software Thesis DOCX"
-  short_description: "把软件仓库整理成毕业论文 DOCX，并支持格式 preset、Mermaid、AIGC 与严谨写作编排"
-  default_prompt: "Use $software-thesis-docx to build or refine a software thesis DOCX workflow from my project repository, with optional style preset extraction, Mermaid diagrams, subagent writing review, and AIGC risk checks."
+  short_description: "把软件仓库整理成毕业论文 DOCX，并支持格式 preset、Mermaid、样本驱动 AIGC 检查与降重编排"
+  default_prompt: "Use $software-thesis-docx to build or refine a software thesis DOCX workflow from my project repository, with optional style preset extraction, Mermaid diagrams, subagent writing review, and sample-driven AIGC risk checks or low-AIGC rewriting."

skills/software-thesis-docx/assets/examples/thesis_workflow_options.example.json

@@ -32,6 +32,8 @@
     "check_only": true,
     "auto_reduce": false,
     "threshold": 0.58,
-    "target_scope": "generated_content"
+    "target_scope": "generated_content",
+    "rewrite_profile": "academic_safe",
+    "normalize_typography": true
   }
 }

skills/software-thesis-docx/references/aigc.md

@@ -6,6 +6,11 @@ Use AIGC risk checks when the user wants a local, heuristic pass over thesis pro
 
 This is a local risk score, not a claim that it matches any institution's external detector.
 
+The skill now supports two rewrite profiles:
+
+- `academic_safe`: default; keep academic tone and only absorb low-risk anti-template rewrites
+- `explicit_low_aigc`: only for users who explicitly ask to lower AIGC; allows stronger sentence expansion and naturalization
+
 ## Checker Script
 
 ```bash
@@ -20,34 +25,57 @@ The report includes:
 - excerpt
 - risk score
 - triggered signals
+- rewrite recipe
+- typography flags
 - whether rewrite is recommended
 
+## Rewrite Script
+
+```bash
+python3 scripts/rewrite_low_aigc_docx.py \
+  --input thesis.docx \
+  --report report.json \
+  --output thesis-low-aigc.docx \
+  --pending-output aigc-pending-review.json \
+  --profile academic_safe \
+  --normalize-typography
+```
+
 ## Current Signals
 
 - frequent transition cliches
 - low-information summary phrases
-- overly uniform sentence-length patterns
-- repeated tri-grams
+- compressed academic enumerations
+- colon/semicolon-driven template paragraphs
+- dense operation chains
+- term stacks without enough explanation
 - claims without concrete evidence markers
-- low lexical diversity
+- typography issues in Chinese thesis paragraphs
 
 ## Recommended Workflow
 
 1. Run the checker and inspect high-risk paragraphs first.
 2. Only rewrite paragraphs the user has authorized.
-3. Prefer concrete, evidence-linked rewrites over cosmetic synonym swaps.
-4. If a paragraph is a single run, reuse `rewrite_paragraphs.py` for deterministic updates.
-5. If a paragraph has mixed formatting or multiple runs, stop and ask for manual confirmation before rewriting.
+3. Use `academic_safe` unless the user explicitly asked to lower AIGC.
+4. Prefer concrete, evidence-linked rewrites over cosmetic synonym swaps.
+5. Normalize punctuation and spacing while rewriting.
+6. If a paragraph is a single run, `rewrite_low_aigc_docx.py` can auto-apply the rewrite.
+7. If a paragraph has mixed formatting or multiple runs, stop and review the pending JSON output before manual confirmation.
+
+See [low-aigc-playbook.md](low-aigc-playbook.md) for the sample-driven methodology.
 
 ## Safe Defaults
 
 - `enabled`: false
 - `check_only`: true
 - `auto_reduce`: false
 - `target_scope`: generated content or user-selected paragraphs only
+- `rewrite_profile`: `academic_safe`
+- `normalize_typography`: true
 
 ## What Not To Do
 
 - Do not silently rewrite an entire thesis.
 - Do not imply guaranteed compliance with any third-party AIGC score.
 - Do not treat the risk report as a substitute for human academic review.
+- Do not reduce citation density, weaken facts, or replace technical terms inaccurately just to chase a lower score.