monograph-writing

monograph-writing — Academic Manuscript Batch-Writing System

English · 简体中文

Turn "writing a 12-chapter book" into an AI-assisted, script-validated, multi-author pipeline — with a hard no-fabrication guarantee for every citation.

Want to see it in action first? Browse the complete example project: examples/demo-project — every artifact from project initialization, section writing, and validation through stripped-clean-copy generation.

1. Who is this for?

Anyone writing a long-form academic manuscript: professors writing academic monographs, researchers writing technical books, engineers writing industry handbooks, graduate teams co-writing textbooks, and doctoral/master's students writing theses (blind review included, and thesis-to-book conversion) — especially if:

• the book has a dozen chapters and hundreds of thousands of words, and cannot be written in a single AI conversation;
• you use AI to write but fear it fabricating references and data;
• multiple people / multiple sessions work in relay, and you fear inconsistency and lost progress.
• you are writing a thesis and fear the AI can't tell your own contributions from citable references — misattributed innovation points, blind-review anonymization, and plagiarism-check pitfalls.

If any of these sound familiar, this skill was designed for you. No programming knowledge is required at any point — your AI assistant runs every script for you.

2. What problems does it solve?

3. Quick Start (3 steps)

Step 1: Install

If you use a command line (requires Node.js):

npx skills add newdeme/monograph-writing

No command line? Copy this sentence and paste it into your AI assistant — it will do everything:

Please install the monograph-writing skill from GitHub for me: clone the repository https://github.com/newdeme/monograph-writing, then copy only the three items the skill needs — SKILL.md, references/, and scripts/ — into a monograph-writing/ folder in your skill library (skip tests/, examples/, docs/ and .github/ — they are repository development files), confirm the skill is recognized, and tell me how to get started.

Works with any AI assistant that supports the Agent Skills open specification (Claude Code, Claude Agent SDK, Cherry Studio, …). Start a new conversation after installing.

Prefer a marketplace? The skill is also live on multiple platforms:

• Tencent WorkBuddy Skill Market — open.workbuddy.cn/skill/all (search "monograph-writing")
• SkillHub — skillhub.cn/skills/user_a72e0254/monograph-writing
• 虾评 Skill — xiaping.coze.com/skill/886eab17-7a88-43d5-af6b-6cf6573689b3
• 魔搭社区 — modelscope.cn/skills/newdeme/monograph-writing
• rednote RED Skill — skill card mounted on rednote notes; copy the install command into your local AI assistant

Step 2: Prepare your table of contents

Put your book outline into one Markdown file (your AI can format it for you; see a complete example in examples/demo-project):

# Your Book Title — Table of Contents

## Chapter 1 Title

**1.1 Section Title**
- 1.1.1 Subsection Title
- 1.1.2 Subsection Title

**1.2 Section Title**
…

**1.N Summary**

Step 3: Say one sentence to your AI assistant

"Initialize my monograph project with the monograph-writing skill; the outline is at xxx.md"

The AI runs init_project.py to scaffold the project, then confirms word-count tiers and your evidence-corpus setup with you. After that, every time you say "continue", it advances to the next batch according to the ledger. You get a report after each batch: files written, validation results, and a preview of the next batch.

4. Bring-Your-Own Evidence Corpus (signature feature)

The two biggest fears when writing with AI: fabricated references, and your materials being out of the AI's reach. This skill solves both: you put your literature into a corpus beforehand, and the AI only searches what you gave it, verifying every citation against the original before it enters the manuscript.

At initialization, the "corpus onboarding interview" offers four choices (upgradeable at any time):

The laziest path: send the files to your AI assistant and say "organize these into the project's 02_语料 folder, grouped by theme".

The boundary of your corpus = the boundary of evidence: every document you put in should support your argument. What may enter the reference list is further governed by contribution classification — published work by others and your own published work are citable (the latter with a relation note), while your own unpublished data/drafts stay as writing material by default (see the FAQ below and references/evidence-corpus.md §8). Setup steps for each option and upgrade paths: references/evidence-corpus.md.

5. The five automation scripts (your AI runs them; humans can read them)

All project parameters live in one file — 00_管理文件/书稿配置.json. Change word-count targets or add exemption tiers there; the scripts never need editing.

Note: script output and generated ledger templates are currently in Chinese (functionality is language-independent; internationalization is on the roadmap).

6. Data privacy & technical facts

• Your data never leaves your computer: all five scripts run locally — no network, no telemetry, no data collection. Your manuscript, outline, and ledgers stay on your own disk. Literature search goes through your own AI assistant and knowledge base, under your control.
• Minimal dependencies: pure Python standard library (any Python 3 ≥ 3.9); two optional dependencies as needed — python-docx for Word merging, matplotlib for data figures (your AI can install and run them).
• Platform compatibility: follows the Agent Skills open specification; works with any compliant AI assistant.
• Third-party evaluation & security: SkillHub TRACE evaluation 4.7/5 "Excellent" (Trust 5.0); platform security scanning (Tencent TIX / Cloud Ding Lab) re-runs automatically on each release — both reports on the current v1.7.0 package read "safe, no risk"; 虾评 Skill marketplace security check passed on the current v1.7.0 package (view · intent/behavior consistency verified, 0 Semgrep findings, all four risk categories low).
• Reporting issues: script bugs and security concerns go to GitHub Issues.

7. Core design principles

1. The ledger is the state: progress, conventions, numbering pointers, verified references — all recorded in project files on disk, not in conversations. Switch sessions, co-authors, or computers without losing anything.
2. Batch-by-batch progress: one section directory = one batch (all subsections + a section summary). Small deliveries; validation must reach zero errors before a batch closes, so errors never accumulate into disasters.
3. Judgment separated from execution: the author is the academic authority (arguments, evidence selection, final calls); the AI is the writing partner and process executor — it never replaces the author's judgment and never fabricates evidence.
4. Layered summaries against forgetting: section summaries (600–800 words) → chapter summaries written only from the section summaries, avoiding the context loss of re-reading entire chapters.
5. Preparation kept with the prose: every subsection file carries its argument tree and evidence list (three-part structure) for auditability during review; stripped automatically at book assembly.

8. Project structure

your-book/
├── 00_管理文件/            outline, config, progress ledger, terminology ledger, instruction list
├── 01_书稿/第X章 章标题/Y.Z 节标题/   the whole manuscript (grows batch by batch)
├── 02_语料/                your evidence corpus (PDF/Word/Excel/Markdown — you decide)
├── 03_归档素材/            external drafts (go through the repair workflow; never used directly)
├── 04_剥离版书稿/          clean copies + merged Word manuscript (script-generated)
└── 05_图表/                figures: data plots (script) / drafts (Mermaid) / finals (author)

9. Multi-author collaboration

An academic-ethics red line first: multi-author collaboration applies to monographs, textbooks, technical books, collected reports — works that can bear multiple authors. A thesis must never be co-written. In thesis projects, "multiple people" means only: (1) the same author relaying across sessions/devices via the ledger; (2) an advisor's review comments flowing through annotations. Chapter content must be the degree candidate's own; using the relay machinery to ghostwrite a thesis is academic misconduct.

• Each co-author owns several chapters; everyone follows the same method (this skill is the standard).
• Handover: the progress ledger + a handover note travel with the workspace.
• When the corpus can't travel with the files, the incoming author uses only the ledger's "verified references cache" (§4 of the ledger); new references go through the corpus owner — better to cite less than to cite wrong.

10. FAQ

Q: I have a pile of references (mixed PDF/Word/Excel/Markdown). How do I let the AI use them? Simplest: put them all into the 02_语料/ folder created at initialization (a usage note is inside). Any format, your call on content. Then tell the AI: "the corpus is in 02_语料/ — search and verify against it when writing." For large collections or semantic search, upgrade to an AI knowledge base or Zotero (see Section 4). In every mode the AI cites only corpus materials, verified one by one against the originals.

Q: Will my manuscript be uploaded or collected? No. All five scripts run locally on your computer — no network, no telemetry, no data collection. Manuscripts and ledgers stay on your disk. Literature search goes through your own AI assistant and knowledge base (see Section 6).

Q: I can't use a command line at all. Can I still use this? Yes. Installation is one copy-pasted sentence to your AI assistant; after that, all scripts are run by the assistant — you just talk ("continue", "next chapter", "set chapter 2 to 1200–1800 words").

Q: Can I use a citation style other than GB/T 7714? Yes. GB/T 7714 (the Chinese national standard, sequential numbering) is the default. Change citation_style in 书稿配置.json and set check_citation_style to false (that check is tailored to GB/T 7714 type identifiers).

Q: Why is citation checking so strict? In real monograph writing, out-of-order citation numbers and reference lists that don't match the text are the most frequent errors — manually auditing a 300,000-word book is nearly impossible. Script-enforced checks on every batch are what make the final manuscript trustworthy.

Q: My platform isn't Claude — will it work? The skill follows the Agent Skills open specification (SKILL.md + scripts + reference docs), so any compliant AI assistant can use it. The scripts are pure Python standard library (except python-docx for Word merging) and can also be used standalone.

Q: What if I need to change the outline mid-writing? The frozen outline is a serious contract: changes require author approval and a version bump (v2, v3…), recorded in the ledger. Completed chapters are never silently rewritten — contradictions are logged for adjudication. See the references/ docs.

Q: A script printed an error I don't understand. What now? You don't need to understand it. Every script error follows the format what happened → common causes → how to fix; do what the "how to fix" line says and re-run. Common errors (broken config JSON, wrong file encoding, filename drift, missing python-docx) have a symptom→cause→fix lookup table in references/troubleshooting.md. The easiest route: paste the full output to your AI assistant and say "fix this for me".

Q: My "book" is really a training handbook / collected report — does it still apply? Yes. Any multi-chapter long manuscript fits. Word-count tiers live in 书稿配置.json and are fully adjustable — global tiers, per-chapter overrides, and special-case exemptions are all supported, so defaults never box you in. For single papers or short documents the batch machinery adds little value and isn't recommended.

Q: Does it work on Windows? What about non-Chinese manuscripts? The scripts are pure Python standard library (only Word merging needs python-docx) and run on macOS/Linux/Windows; quote paths containing spaces or CJK characters. Chinese-language writing is the best supported (word counts are measured in Chinese characters; GB/T 7714 type-identifier checking included). Other languages run, but counting and punctuation rules may be imprecise — multilingual support is on the roadmap.

Q: Will the AI cite my own unpublished data or drafts as references? No — that is the "contribution classification" red line: published work by others is cited normally; your own published work is cited with a relation note ("this chapter extends the author's published work [n]"); your own unpublished data stays as writing material (labeled "author's experimental data" in text and figures; freeze it into a registered snapshot to cite it formally as [DS/OL]); unpublished drafts never enter the reference list — they are rewritten, not cited. Discipline differences are explicit too (unpublished archives in the humanities are cited per disciplinary convention, e.g. [A]). See references/evidence-corpus.md §8.

Q: Does it work for a thesis going to blind review? Yes, with a dedicated guide: declaring the thesis genre creates an innovations-and-outputs register (innovation ↔ supporting chapters ↔ published papers), ready-made relation-note phrasing for your own published work, and a blind-review anonymization checklist (acknowledgements removed, self-citations de-identified or removed per your university's rule, full name-sweep before export). Citation renumbering after anonymization still goes through the validator until it passes. Plagiarism-check handling and the dual-submission risk of concurrently-submitted drafts are flagged in the ledger. See references/thesis-guide.md.

Q: Can the AI draw figures for me? Two kinds: data figures (statistical charts) — yes. Put your data files (CSV/JSON) in the corpus, register an evidence card in ledger §3b (figure number + data file + checksum + chart recipe), and generate_figures.py renders bar/line/scatter/box/histogram charts from the data; re-run after data changes; data-file drift is flagged by the validator. Concept diagrams (schematics/architecture) — drafts only. The AI produces a Mermaid text draft (version-controlled under 05_图表/草稿/) for you to redraw; the manuscript keeps a "Figure X-Y (to be finalized)" placeholder, and the publication-grade artwork is yours — conceptual structure is your scholarly judgment. See references/figure-guide.md.

Q: GB/T 7714 got a 2025 revision — does the skill follow it? Yes. The new standard (effective July 2026) adds preprint [PP], dataset [DS], and archive [A] rules — the type-identifier check accepts both old and new codes, and frozen datasets are cited as [DS/OL] per the new standard. If your institution or journal pins a specific version, theirs wins.

11. Citing this project

If this skill helps your research, teaching, or work, a citation is the best way to support it. Click the "Cite this repository" button on the repository homepage (driven by CITATION.cff) to get a ready-to-use citation, or use the BibTeX entry below:

@software{chang2026monograph,
  author  = {Chang, Sheng},
  title   = {monograph-writing: Academic Manuscript Batch-Writing System},
  year    = {2026},
  url     = {https://github.com/newdeme/monograph-writing},
  version = {1.7.0},
  license = {Apache-2.0}
}

12. Contributing & license

• This skill distills the complete writing practice of a real academic monograph — the methodology is battle-tested over 200+ section batches.
• Feedback and improvements welcome via Issues / Pull Requests (batch experiences, new validation rules, adaptations to other citation styles). For commercial cooperation, private-deployment guidance, or custom development, please reach the maintainer via GitHub Issues.
• Released under the Apache License 2.0, © 2026 newdeme — includes patent grant, safe for academic and commercial use.
• Ideas adopted from prior Agent Skills work via a design-space survey: see ACKNOWLEDGMENTS.md.

Co-developed by an AI writing assistant and a monograph author. Methodology details: SKILL.md and the references/ directory.