research-writing-skill: Turning Academic Writing From Chat-Based Drafting Into a Versioned Engineering Pipeline
> Source: Norman-bury/research-writing-skill, https://github.com/Norman-bury/research-writing-skill
---
Why Your Thesis Has Been Stuck on the Introduction for Three Months
The most common scene of an undergraduate writing a thesis looks like this: launch ChatGPT, ask "write me a research background paragraph on X," copy and paste, hand it to the advisor. The advisor annotates "the logic is off," so the student asks the AI again, copies and pastes again. After ten cycles, the introduction has been rewritten for three months and the main body has not even been started.
This is not a writing-skill problem. It is treating paper writing as a single-shot chat instead of a traceable, reusable engineering workflow.
Norman-bury's research-writing-skill exists to fix this.
---
Positioning: Not a "Polishing Tool," but an "Engineering Collaboration System"
Most "AI writing assistants" on the market are positioned as polishers: you throw in a paragraph, they make it smoother. That model has a fundamental flaw — single-turn conversation, no memory, no process, no versions.
research-writing-skill is positioned completely differently. It manages paper writing like software engineering.
- Align goals and constraints before the task starts
- 7 rounds of Q&A to confirm paper type, discipline, topic, methods, and chapter structure
- Staged gating execution: topic → body → figures → self-review → delivery
- Outputs are project files (
.md/.tex), traceable, recoverable, and versionable - Paper type (thesis / coursework / submission draft)
- Discipline (engineering / social science / medicine / law split)
- Research topic and background
- Core methods and data sources
- Chapter structure plan
- Preserve the research object, data scope, sample caliber, method conditions, metric definitions, experimental boundaries, conclusion limits, and proper nouns
- Language adjustment only serves natural expression, clear logic, and stable terminology
- The body prefers continuous paragraphs over bullet points, and does not rely on bold/italics to manufacture emphasis
- Avoid mechanical connectors such as "firstly, secondly, finally, moreover, in addition, next, in summary"
- Avoid empty-shell phrases such as "it is worth noting that," "it should be pointed out that," "what is important is that," "it must be emphasized that"
writing-core: general writing conventionswriting-humanities: social science / humanitieswriting-medical: medicine / biologywriting-law: law- Claude Code (
.claude-plugin/) - Cursor (
.cursor-plugin/) - Codex (
.codex/) - OpenCode (
.opencode/) - Gemini CLI (
GEMINI.md) - Generic agents (
AGENTS.md) - Version control: Git can diff Markdown but not Word
- Reusability: Markdown is plain text — bulk replace, script processing, template reuse all work
- Cross-platform: any text editor can open it, no Office dependency
- Conversion flexibility: Pandoc can convert to Word, PDF, LaTeX, but reverse conversion loses formatting
- Norman-bury/research-writing-skill, GitHub, https://github.com/Norman-bury/research-writing-skill
- Version: 3.1.0, updated: 2026-05-10
It is not "write me a paragraph" — it is "let's manage this project together."
---
The Engineering Workflow: From Brainstorm to Delivery
Step 1: Brainstorm (7-round Q&A)
Rather than "write my paper," the system systematically confirms:
The value is reduced rework. Many people discover halfway through that their methodology is wrong or their chapter structure does not match the school's requirements, so everything written before is scrapped. The 7-round Q&A upfront solves these issues before a single word is written.
Step 2: Chapter-by-Chapter Writing
Each chapter lives in its own file under chapters/. The introduction and related work must first produce a literature evidence map (refs/evidence-map.md). The experiment chapter must first produce an experiment protocol (plan/experiment-protocol.md), table schema (tables/table-schema.md), and figure manifest (figures/data-manifest.md).
Mandatory up-front planning — skipping structure to write the body directly is not allowed.
Step 3: Figure Generation
Data-result figures are preferably generated by Python scripts to ensure reproducibility. Flowcharts, architecture diagrams, and mechanism diagrams are produced via the figures-diagram module's prompts and then drawn with image-generation tools such as Gemini.
Technical figures and conceptual figures are separated — the tools are not mixed.
Step 4: Pre-Submission Self-Review
The peer-review module self-checks for logical gaps, data inconsistencies, and overreaching conclusions.
Step 5: Delivery
The output is Markdown or LaTeX, not Word. When Word is needed, it is converted manually or via Pandoc. Word is not the default because research collaboration needs traceable, reusable text assets.
---
De-AI Writing: Preserve Information Density, Only Adjust Expression
The skill's understanding of "de-AI" is on point: not compressing the text, but preserving all facts, data, qualifiers, and explanatory sentences, only adjusting the way they are expressed.
Specifically:
This principle is more sophisticated than most "AI-rate reduction" tools on the market. Many tools compress the text heavily — it looks "refined" but actually deletes key qualifiers and data boundaries, turning the paper from "academically rigorous" to "deliberately vague."
The skill's stance is: rather be slightly verbose than lose information completeness. That is the essential requirement of academic writing.
---
Discipline Split: Separate Paths for Engineering, Social Science, Medicine, and Law
The problem with generic writing tools is "one size fits all." Engineering papers need methodological detail and experimental numbers; social-science papers need theoretical frameworks and literature reviews; medical papers need statistical analysis and ethics statements; law papers need statute citations and case arguments.
The skill includes discipline modules:
Splitting by discipline is not just swapping a prompt template — it is a systematic adjustment of argumentative logic, citation format, and chapter weighting.
---
Multi-Platform Compatibility: Claude Code, Cursor, Codex, Gemini CLI
The skill uses a directory-based design and is already adapted for:
This means no lock-in to a single tool. You can draft with Claude Code, produce figures with Cursor, generate flowcharts with Gemini CLI, and manage all outputs in the same project directory.
---
The 20+ Skill Modules at a Glance
| Scenario | Skill Module |
|----------|--------------|
| Entry and routing | using-research-writing |
| Mid-sized / full-paper orchestration | paper-orchestration |
| Brainstorming | brainstorming-research |
| Evidence-driven intro / related work | evidence-driven-writing |
| Chapter writing | writing-chapters |
| Experiment and results planning | experiment-results-planning |
| LaTeX output | latex-output |
| Literature review | literature-review |
| Translation / polishing / de-AI | prompts-collection |
| Pre-submission self-review | peer-review |
| Statistical analysis | statistical-analysis |
| Python data figures | figures-python |
| Flowchart / architecture diagram | figures-diagram |
| Environment setup and troubleshooting | environment-setup |
---
Output Design: Why Markdown Is the Default Over Word
This design choice has clear rationale:
Word suits final delivery; Markdown suits process iteration. The skill puts each in the right place.
---
LaTeX Support: Template Parsing and Automatic Compilation
If a school or journal provides a LaTeX template, drop the .cls / .sty / .tex files into latex-templates/. The skill parses the template structure and generates matching chapter .tex files. The output is a complete, ready-to-compile LaTeX project.
This goes further than "here's a generic template, edit it yourself" — it understands your template's structure and generates content that fits it.
---
Conclusion: Research Writing Needs Industrialization
The core value of research-writing-skill is not "let AI write your paper" — it is "turn paper writing from a hand workshop into an assembly line."
Its engineering workflow, de-AI principles, discipline split, multi-platform compatibility, and output design all point to one goal: let researchers spend time and energy on the research content that truly matters, rather than repeatedly reworking format and wording.
For undergraduates, graduate students, and early-career researchers, this skill is more practical than any "one-click paper generator." Because it does not promise shortcuts — it provides a sustainable, reusable, iterable writing system.
> "Paper writing is not chat. It is engineering."
---
References