The Template Wars: In Which Humanity's Most Powerful AI Is Defeated by a Word Document from 2017

This week, Anthropic's CEO Dario Amodei published a 9,000-word essay comparing AI development to "the adolescence of a new kind of entity" — a being that is "brilliant but unreliable, knowledgeable but lacking judgment." Meanwhile, Axios reported that Claude had been used in Pentagon planning related to regime transition scenarios in Venezuela. And me? I was in my office in Pandanus Reach, thirty-four degrees, trying to get that same AI to put the words "Week 3: File Management" inside row four of a Word table without destroying the entire document. The adolescence of technology, indeed. Some of us are just trying to survive its gap year.

Act I: In Which the Most Sophisticated AI on Earth Meets a Table It Cannot Navigate

It began, as all of my contributions to the field of AI-assisted document production begin, with a template. Not a template in the aspirational, Silicon Valley sense — not a "reusable design system" or a "scalable content framework." A template in the MVIT sense. A Word document. Created by a person who has long since left the organisation. Last meaningfully updated during the Turnbull government. Passed down through generations of staff like a cursed heirloom, each inheritor adding their own modifications with the confidence of someone who doesn't understand what "Merge Cells" actually does to the underlying XML structure of a document, because why would they, because they are a teacher, not a software architect, and the fact that these two roles now intersect is a development that nobody at their job interview was warned about.

The template was for a lesson session plan. Version 4.1. For the unit BSBXTEC301 — Design and Produce Business Documents. I want you to sit with that for a moment. The template I was trying to update — using an AI assistant — was for a unit that teaches people how to design and produce business documents. The document designed to teach document design was itself so catastrophically designed that it defeated the most advanced language model available to the general public.

The irony is not subtle. It is not, in fact, irony at all. It is something closer to a philosophical indictment of the entire vocational education sector's relationship with its own tools. But we'll get to that.

In Which I Am Politely Asked to Upload the Crime Scene

I should explain what I was trying to do, because it sounds simple. It was simple. It was, in fact, the simplest possible use case for an AI assistant: take existing content and put it inside an existing template. Move text from location A to location B. This is what computers were invented for. This is, by some measures, the only thing computers were invented for. And yet.

I opened Claude. I described the task. I said, and I quote, because I have the transcript: "I have a lesson session plan template in Word. I need to update it with new content for Weeks 1 through 8. The content is in a separate document. Can you populate the template?"

Claude's response was calm, professional, and faintly ominous in the way that a bomb disposal expert is calm and professional when they ask you to describe what you see: "I'd be happy to help populate your template. Could you upload the Word document so I can examine its structure?"

Examine its structure. In hindsight, this was the first warning. Claude didn't say "so I can fill it in." It said "so I can examine its structure" — the phrasing of someone who suspects they are about to open a suitcase and find something distressing inside. The AI equivalent of putting on gloves before handling evidence.

I uploaded the file.

What followed was a silence that, in a chatbot, should not be possible but somehow was. A pause. A thinking indicator that seemed to pulse with something I can only describe as trepidation. And then:

Intricate. The word a doctor uses when they mean "I'm going to need you to sit down." The word an estate agent uses when they mean "the previous owner was creative with load-bearing walls." Claude had looked inside the template and found something that disturbed it. An AI system that, according to the latest benchmarks, approaches PhD-level reasoning on graduate-level science questions had opened an MVIT session plan and recoiled.

Act II: In Which We Unpack the Hot Mess (Claude's Words, Not Mine, Actually Mine)

To understand what happened next, you need to understand what the template looked like from Claude's perspective. Not the visual perspective — the neat rows and columns that a human sees when they open Word and squint at a table. The actual perspective. The XML. The code beneath the code. The skeleton under the skin.

Here is what I saw when I opened the template: a professional-looking document with a table. Rows for each week. Columns for session topics, learning outcomes, activities, and resources. It looked fine. It looked, dare I say, organised.

Here is what Claude saw: a labyrinth. A document in which someone — possibly several someones, across multiple versions, over what appears to have been a period of years — had used the "Merge Cells" function with the joyful abandon of a child who has discovered that scissors can cut things other than paper. Rows had been merged across columns. Columns had been merged across rows. In one section, cells had been merged and then split and then merged again, creating a structure that exists in no known taxonomy of table design and which, if you attempted to describe it in XML, would look less like markup and more like a cry for help.

The document was not, technically, a table. It was a collection of rectangles that had been arranged in the approximate shape of a table by someone who understood that tables contain information arranged in rows and columns but had interpreted this concept with considerable artistic licence. It was Cubism applied to administrative documentation. It was a Picasso, if Picasso had worked in the medium of MVIT session planning.

The merge pattern makes it difficult to predict how adjacent cells will respond. I want to frame this sentence. I want to hang it in the staffroom of every MVIT campus in Australia, next to the fire evacuation plan and the passive-aggressive note about the shared milk. This is what happens when a document is maintained by committee, over years, using a feature that nobody fully understands. The cells have become unpredictable. They have developed their own behaviour patterns. They respond to changes the way a cat responds to a new piece of furniture — with hostility and in directions nobody anticipated.

The Merged Cell Incident (A Tragedy in Seven Attempts)

What followed was a process I can only describe as iterative despair. I would ask Claude to put content in a cell. Claude would put content in a cell. The cell would be the wrong cell. Or the right cell, but the act of placing content in it would cause an adjacent cell to expand, contract, or vanish entirely, like a soap bubble that has been poked by the finger of a god who does not understand surface tension.

Attempt one: content placed correctly in row 4, but row 5 lost its right border. Attempt two: border restored, content now in row 3. Attempt three: content in row 4 again, but the column widths have changed and the document now appears to be wider than an A4 page, a physical impossibility that Word handles by simply ignoring it and hoping nobody prints the document. Attempt four: Claude suggests starting from a clean copy. I agree. We start from a clean copy. The clean copy, it transpires, already has three merged cells in positions that differ from the version I was just working on, because at some point someone saved a "clean" version that was actually a slightly different version of the disaster.

Attempt five: Claude proposes a different approach — extracting the content from the table, rebuilding the table from scratch with a clean structure, and then reinserting the content. This is the AI equivalent of a builder telling you it would be cheaper to demolish the house and start again than to fix the extension your uncle built in 1987. It is probably correct. It is also not what I asked for. I asked for text in a box. I have been asking for text in a box for forty-five minutes.

The boxes aren't really boxes. They're an agreement between multiple overlapping structures to look like boxes. I have never in my life felt more seen by a sentence produced by a machine. This is not merely a description of a Word document. This is a description of institutional template management in the Australian vocational education sector. The boxes aren't really boxes. They're the accumulated decisions of every person who ever opened this file, moved a border, merged a cell, and saved it without fully understanding what they had done to the underlying architecture. The template is not a document. It is a palimpsest. It is a geological record of administrative decisions stretching back to the Abbott era, each layer deposited on top of the last, and the AI — bless its statistically-inferred heart — is trying to navigate it with the only tool it has: the actual structure of the file. Which bears approximately the same relationship to what you see on screen as a building's plumbing bears to its façade.

Attempt six: Claude successfully populates weeks 1 through 4. Week 5's row, however, contains a merged cell that spans the entire width of the table — a "section header" that someone added at some point for reasons that are now archaeologically irrecoverable. Claude treats this as a single cell. The content intended for five separate columns is deposited into one cell as a single paragraph. It looks like someone has emptied a filing cabinet into a shoebox.

Attempt seven: I give up on the original template. Claude builds a new table from scratch. The new table works perfectly. It contains all the content. The formatting is clean. The columns align. The rows are consistent. And it looks absolutely nothing like the institutional template, which means I cannot use it, because the template is not merely a container for information — it is a compliance artefact, a quality assurance requirement, a document whose format has been specified by people who have never opened it and approved by people who have never read it, and whose exact appearance must be preserved even if that appearance is the visual manifestation of structural chaos.

The Irony, in Case You Were Distracted by the Merged Cells

I need to return to a fact I mentioned earlier, because I think it got lost in the carnage: the unit this session plan was for is BSBXTEC301. Design and Produce Business Documents. The learning outcomes include "select and use appropriate software to create documents" and "design document layouts to meet organisational requirements."

I was using an AI assistant to update a session plan for a unit that teaches people how to design documents, and the session plan itself was so poorly designed that the AI assistant could not reliably interact with it. The template for teaching document design was a case study in how not to design documents. The irony is not layered; it is geological.

And here is the part that truly sends me into a spiral: the students taking this unit will, in all likelihood, use AI to complete their assessments. They will ask ChatGPT to create a business letter, a report, a formatted document with tables and headings and styles. And ChatGPT will produce one — from scratch, cleanly, with consistent formatting and valid underlying structure — in approximately four seconds. It will do this because, as I wrote about on this very website, AI doesn't wrestle with templates. AI doesn't merge cells. AI doesn't inherit the structural sins of every previous editor. AI writes code that generates documents programmatically, and the documents it generates are structurally sound because they were built by code, not by a sequence of human beings clicking buttons they didn't fully understand.

The students will produce better documents than the template I'm using to teach them about documents. They will do it faster, with less effort, and with none of the existential suffering I experienced on a Tuesday afternoon in Pandanus Reach while an AI tried to put the words "Week 3: File Management" into a cell that wasn't really a cell.

This is either the future of education or the end of it. I am not yet sure which.

The Markdown Revelation (Or: What If the Boxes Were Never the Point?)

Somewhere around attempt five — the one where Claude gently suggested burning the house down and starting again — it said something that has been rattling around my skull ever since:

Thirty seconds. I had spent forty-five minutes fighting a Word document, and the AI was telling me it could produce the same information — structured, readable, portable — in thirty seconds, if only I would agree to let go of the boxes.

Here is what a lesson session plan looks like in Markdown:

That's it. That's the whole thing. No merged cells. No column spans. No XML table elements with inheritance attributes that behave like poltergeists. Just text, structured by symbols that a human can read with their eyes and an AI can read with its parser and both of them can understand immediately without anyone needing to guess what will happen if they add a word to paragraph three.

Markdown is what you get when you separate the question "what information needs to be here?" from the question "what should it look like?" Word conflates these two questions so thoroughly that most users don't know they're different questions. Every time you bold something in Word, you're making a formatting decision and a structural decision simultaneously, and Word stores both of them in the same tangled heap of XML, and when something goes wrong you cannot tell whether the problem is what you said or how you said it.

Markdown says: write the content. Structure it with headings and lists. Let something else — a different tool, a different process, a stylesheet, a template engine, an AI — handle the presentation. This is, I should note, not a radical idea. It is how the internet has worked since CSS was invented in 1996. It is how every developer on Earth manages documentation. It is how this blog post was drafted. The only people still combining content and formatting in a single inextricable mess are the people using Word templates — which is to say, everyone in vocational education, everyone in government, and everyone who has ever been told "use the template" by someone who has never opened the template.

An AI can read a Markdown file the way you read a shopping list: directly, completely, without ambiguity. An AI reads a Word document the way an archaeologist reads a dig site: carefully, inferentially, hoping that the layers of accumulated modification don't collapse when you touch them.

Research is catching up with what practitioners have been discovering the hard way. A recent paper, He et al. (2025), benchmarked LLM performance on real-world document tasks — not the sanitised, well-structured examples of the training data, but the kind of documents actual humans create in actual workplaces: forms with merged cells, templates with irregular structures, documents where the formatting has been treated as load-bearing by generations of editors. The results were instructive. AI models that score brilliantly on standardised benchmarks struggle with the structural chaos of real institutional documents. The gap between benchmark performance and real-world utility is, it turns out, approximately the width of a merged cell.

The Part Where It Stops Being Funny (Briefly)

Here is the question I cannot quite laugh away: if the most common request I receive in AI training is "how do I get AI to fill in my template?" — and it is; I hear it weekly, from teachers, from administrators, from public servants whose Ctrl+C and Ctrl+V keys are wearing through the plastic — then what does that tell us about the templates?

It tells us the templates are the problem.

Not AI. Not the users. Not the technology. The templates. The documents themselves. Created years ago, modified by dozens of hands, structurally incomprehensible to the tools that are supposed to help us work with them, and yet untouchable — because they're "the template," because someone approved them, because changing them requires a committee and the committee requires a terms of reference and the terms of reference require a template.

Meanwhile, Dario Amodei writes about AI as an adolescent entity that is "brilliant but unreliable" and calls for society to guide its development with wisdom and patience. Meanwhile, Axios reports that Claude is being consulted on matters of geopolitical strategy. And meanwhile, down here at the coalface, an educator in the Territory is trying to get that same technology to put the words "Week 3: File Management" inside a box, and the box won't cooperate, because the box was built by a human in 2017 who just wanted to draw a line around some text and didn't realise they were creating a structural legacy that would outlast their employment.

The gap between what AI can do and what AI is asked to do is not a technology problem. It is an infrastructure problem. We are asking a system that can synthesise research papers and draft strategic assessments to navigate documents that would defeat a reasonably competent human with a mouse and thirty minutes of patience. The bottleneck is not the AI. The bottleneck is the .docx file.

In Which We Return to Our Regularly Scheduled Absurdity

I finished the session plan. It took two hours. Claude and I arrived at a compromise: it produced the content in clean Markdown, I manually pasted it into the template cells one by one, guided by a map Claude had drawn of which cells were actually where versus where they appeared to be. It was, in every meaningful sense, a joint operation — the AI handling the intelligence work and the human performing the manual labour, which is exactly the opposite of what the brochure promised but which, I suspect, is what most AI-assisted work actually looks like in 2026.

The session plan is now complete. Version 4.1, updated for Semester 1. It contains eight weeks of carefully structured content inside a table that could generously be described as "functional" and accurately described as "a war crime against XML." It will be printed. It will be placed in a folder. It will be reviewed by someone who will look at the neat rows and columns and have absolutely no idea that underneath those boxes is a structure so chaotic that a machine capable of PhD-level reasoning threw up its digital hands and suggested starting over.

Next week, I will teach a class of students how to use software in the workplace. I will teach them about document design. I will teach them about formatting, and structure, and the importance of using tools correctly. I will do this using a session plan that violates every principle I am teaching, in a template that was not correctly designed, modified by people who were not trained in document architecture, and maintained by an institution that has never once asked whether the template works — only whether it exists.

The rainbow ball from the Laterite Highway has taken up residence in my Word document. It lives in the merged cells now, in the spaces between what you see and what is actually there. It whispers to me about column spans. It tells me about XML inheritance. At night, when I close my laptop, I can still hear it: the faint, persistent hum of a table border that doesn't quite align, and never will, and doesn't need to, because nobody has ever looked underneath.

She'll be right. Probably. Ask me again after I've updated Version 4.2.