Skip to content
Blazej Mrozinski
PL

How I Use AI for Academic Writing: A Pandoc Paper Template System

How I Use AI for Academic Writing: A Pandoc Paper Template System

Every academic knows the feeling. You have 40 pages of results, six months of literature review in your head, and a co-author waiting on the methods section. You open Word. You stare at the blinking cursor. Then you spend the next two hours adjusting margins, fixing reference formatting that broke when you moved a paragraph, and debating whether your heading levels comply with APA 7th edition. The actual writing, the thinking-on-paper part, gets maybe 30 minutes.

I have been writing academic papers for nearly 20 years, almost all of that time at SWPS University in Warsaw, where I have worked since 2006 and now teach research methods and psychometrics as an Assistant Professor. In that time, the tools have barely changed. Word. EndNote or Zotero. Email for collaboration. Maybe Google Docs if you’re feeling modern. The substance of academic writing, the argumentation, the methodology, the synthesis of literature, remains one of the most intellectually demanding tasks in professional life. The tooling around it remains stuck somewhere around 2008.

I built something that finally addresses this gap. It’s an academic paper template system that treats manuscript production the way a software team treats code: structured, repeatable, version-controlled, and partially automated. This post explains what I built, why I built it this way, and what it reveals about the gap between how academics work and how they could work.

Table of Contents

The Problem with Academic Writing Tools

Academic writing has a paradox at its center. The content demands extreme rigor. Hypotheses must be stated precisely. Methods must be replicable. Statistical claims must survive peer review from people who do this for a living. Every citation must be traceable, every claim supported.

The tooling demands none of this rigor. Most academics I know write in Word documents with manual formatting. References get managed by plugins that break between versions. Collaboration happens by emailing files named manuscript_v3_final_FINAL_BM_edits.docx. The gap between the intellectual standards of the work and the engineering standards of the workflow is, in my experience, enormous.

I’ve felt this gap for years, but I always solved it the way academics do: by being more careful. Check the formatting again. Verify the references manually. Keep a mental model of which version has which changes. This works, in the same way that a surgeon could probably operate with kitchen knives if they were careful enough. It works, but it wastes a resource that matters: cognitive bandwidth that should be spent on thinking.

The trigger for building something better was specific. I started using AI as a technical collaborator across my other projects, building products, managing infrastructure, running experiments. I had developed a workflow where AI handles implementation while I direct architecture and make decisions. One day it occurred to me that the same pattern should apply to academic writing. I know how papers should be structured. I know what good methodology looks like. I know the APA style guide well enough to teach it. What I lack is patience for the mechanical work of formatting, reference management, and document assembly. That’s exactly the kind of work AI handles well.

What the Template System Actually Does

The system is a folder structure combined with a set of workflow documents that define how to go from a blank project to a submitted paper. It has four layers.

The first layer is a structured manuscript directory. The paper lives as separate sections (introduction, methods, results, discussion), each in its own file. This sounds trivial, but it changes the cognitive experience of writing. When you open the methods section, that’s all you see. No scrolling past 15 pages of literature review to find where you left off. No accidental edits to a finished section while you’re working on a different one.

Some people might argue this adds unnecessary friction, that switching between files is slower than scrolling. For short papers, that might be true. For a 40-page manuscript with co-authors, I find the isolation is worth the extra clicks.

The second layer is a knowledge wiki. This is where research lives before it becomes prose. Source summaries, concept definitions, connections between ideas, synthesis notes, and a running research log. Think of it as the research equivalent of a codebase’s documentation. When you read a paper, you write a structured summary: what the authors found, how it connects to your work, what questions it raises, what you disagree with. When it’s time to write the literature review, you’re assembling from organized materials you already have.

The third layer is APA formatting automation. A CSL citation style file handles reference formatting. A reference template document defines margins, fonts, heading styles, and page layout to APA 7th edition specifications. A build script assembles the manuscript sections and exports them to a properly formatted Word document through Pandoc. You write in plain text, focused entirely on content, and the system handles the rest.

The fourth layer is the workflow pipeline itself. A master instruction document defines the entire paper-writing process as a structured sequence: brief, wiki population, style calibration, drafting, export, revision. Each phase has clear inputs, outputs, and checkpoints. This is the document that makes AI collaboration possible, because it gives the AI a complete map of the process, the standards, and the decision points where human judgment is needed.

Building a Research Knowledge Wiki

The wiki layer deserves its own discussion, because it represents a different philosophy about how research knowledge should be managed.

In my experience, most academics treat their literature review as a one-time activity. You read papers, take notes (maybe), and write the review. The notes live in a pile somewhere. When you start a new paper on a related topic, you begin the cycle again, often re-reading papers you’ve already processed. I suspect this is true across most fields, though I can only speak with confidence about psychometrics and social science.

The wiki approach treats each source as a permanent knowledge asset. Every paper gets a structured summary with consistent fields: key findings, methodology, relevance to your work, critical evaluation, and connections to other sources. Concept definitions get their own entries, linked to the sources that inform them. Synthesis notes capture higher-order patterns across multiple sources.

This takes more time upfront. There is no way around that. Summarizing a paper in a structured format takes longer than highlighting a PDF. The investment pays off across every subsequent paper that touches the same literature. That’s the theory, at least, and it matches what I’ve seen so far. The second paper on a related topic doesn’t require re-reading. The connections between sources are already mapped. The synthesis notes may even contain the seed of your next argument.

For someone who writes across related topics in psychometrics and assessment, this changes the economics of literature review. The first paper in a domain costs more. Every subsequent paper costs significantly less. Knowledge compounds.

APA 7 Formatting with Pandoc: A Solved Problem

Here’s something that should embarrass the academic tooling ecosystem: APA formatting is a published, rule-based specification. Margins, heading levels, and citation formats are all defined in the Publication Manual, and the rules change only with a new edition, roughly once a decade. This is exactly the kind of problem that should be solved once per edition and then left alone.

Instead, academics spend hours on it per paper. Is the running head correct? Are the heading levels right? Did Zotero format that multi-author citation correctly? Is the references section in the right order? These questions consume real time and mental energy on every single manuscript.

The template system uses a small set of components to make this a non-issue. Pandoc, the document conversion tool, handles the assembly and export. A CSL file from the official Citation Style Language repository defines how in-text citations and references should appear. A reference Word document sets the page-level formatting: margins, fonts, heading styles. A small Pandoc filter builds the APA title page. The result: you write content in plain text, run a build script, and get an APA-formatted Word document, formatted the same way on every build. Formatting stops being something you fix by hand, and “did the reference plugin break again” stops being a question.

I want to be clear about what this does and doesn’t solve. It solves mechanical formatting. It doesn’t solve the intellectual challenge of writing clearly, arguing rigorously, or choosing the right statistical approach. Those remain entirely human problems. But removing the formatting tax means more cognitive bandwidth for the problems that actually matter.

What a Writing Pipeline Adds to a Paper Template

Templates are static. You download a Word template with the right margins and heading styles, and that’s helpful for about 20 minutes. Then you start writing and the template doesn’t know anything about your process. It doesn’t know that you should draft the methods section before the introduction. It doesn’t know that your literature review should be grounded in the sources you have already summarized. It doesn’t enforce any workflow.

A pipeline is different. It defines the output format and the sequence of activities that produce good output. The template system includes a paper brief template that captures the research question, hypotheses, methodology, and target journal before any writing begins. The brief works as a forcing function. If you can’t fill out the brief clearly, you’re not ready to write.

You might think this sounds like overhead, especially for a quick conference paper or a shorter empirical report. That’s fair. For straightforward projects, some of these steps can be compressed. But for multi-study papers or anything with complex methodology, the upfront structure pays for itself many times over.

The pipeline then moves through wiki population, where you build the knowledge base for this specific paper. Then drafting, with each section written in sequence with clear dependencies, starting with the methods. Then export, where Pandoc assembles everything. Then revision cycles with co-authors, with structured checkpoints.

Each phase has a defined output. At any point, you can generate a Word document of the current state. This matters for collaboration, because your co-author doesn’t need to understand the system. They get a properly formatted Word document they can read and comment on using whatever tools they prefer.

This pipeline structure is what makes AI collaboration effective. When an AI assistant has a clear specification of the process, it can handle the mechanical parts of each phase while you focus on the intellectual decisions. It can draft a methods section from a well-structured brief. It can format references from your wiki entries. It can assemble sections and run the export. What it can’t do, and shouldn’t try to do, is decide your research question, evaluate your methodology, or judge whether your argument holds up. The pipeline makes these boundaries explicit.

Voice Calibration: The Part Nobody Talks About

One component of the system surprised me by how much it mattered: the style memo. This is a template for calibrating writing voice to match a target academic style.

Academic writing is often treated as if it has one voice: formal, passive, impersonal. Anyone who reads widely in their field knows this isn’t quite right. Different research groups have distinctive voices. Some write with precise, economical prose. Others favor longer explanations that walk the reader through each logical step. Some embrace first-person reporting of results. Others maintain strict third-person throughout. The differences are subtle, but reviewers and readers notice when the voice shifts within a paper.

When you collaborate with co-authors, matching voice becomes a real challenge. I’ve worked with John Nezlek, whose writing style has a particular precision and economy that takes conscious effort to match. Before the style memo, this matching happened through intuition and revision cycles. Now it happens through a structured analysis: sentence length patterns, preferred terminology, transition styles, level of formality. The memo captures these patterns explicitly so they can be applied consistently.

This is another area where AI collaboration becomes more effective with structure. An AI can match a voice profile if you define it precisely. It cannot intuit the voice of a research group it has never read. The style memo bridges that gap. Whether the match is perfect is another question. I’d say it gets to about 80% and the remaining 20% still requires human revision. That still saves significant time.

What This Changes About Collaboration

Academic collaboration has a coordination problem. Multiple authors, different tools, different writing habits, files going back and forth with version confusion. The template system changes the shape of the problem, though parts of it remain unsolved.

Because the pipeline produces standard Word documents at any checkpoint, collaborators don’t need to adopt the system. They receive a clean document, provide feedback in their preferred format, and the lead author integrates changes back into the structured source. This is a pragmatic choice. Asking a senior professor to learn a new toolchain is a non-starter. Giving them a better Word document than they usually receive is an easy sell.

I should note that this model works best when one person owns the pipeline. For truly equal co-authorship where both parties are writing simultaneously, plain files with a single lead author may not scale. That’s a limitation I haven’t solved.

The structured source also makes it easier to track what changed and why. When sections are separate files, you can see which parts were revised and which are stable. The research log captures decisions and their rationale. If a reviewer asks “why did you choose this analytical approach,” the answer was written down when the decision was made.

What I Got Wrong (and What Still Needs Work)

The first version took a few days to build, so some decisions were fast and pragmatic. The wiki structure assumes solo or small-team use. For larger research groups, it would need access controls and concurrent editing support that plain files don’t provide. The build script is functional but minimal. It handles the common case well, but edge cases in citation formatting (obscure source types, non-English references) haven’t been tested thoroughly.

The bigger open question is longevity. Academic careers span decades. A paper template system is only valuable if it outlasts the current set of tools. Plain text and Pandoc are good bets for durability, better than any proprietary format. But the workflow documents that make the system powerful are also the most likely to need revision as my process evolves. I’m betting on the foundation being stable even if the specifics change. That bet could be wrong.

There’s also a question I can’t answer yet: how much of the productivity gain comes from the system itself versus from the process of building it. The act of making my paper-writing workflow explicit forced me to examine habits I’d held for years without questioning them. Some of those habits turned out to be load-bearing. Others turned out to be superstition. The template system captures the load-bearing ones, but the real value might be in the examination process itself. If that’s the case, the system works mainly as a catalyst.

What I’m confident about is the direction. Academic writing is too important and too intellectually demanding to have its practitioners wasting hours on solved problems like formatting and file management. The tools should handle the mechanical parts so the humans can focus on the parts that require actual thinking. Twenty years of writing papers has taught me what those parts are. For the kind of work I do, this system reflects that knowledge in a form that can be reused, shared, and improved. Whether it generalizes beyond psychometrics and social science is something I’d need to see tested.