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

This week, CEO of Anthropic Dario Amodei described AI development as "the adolescence of a new type of entity," a brilliant but unreliable being that is "knowledgeable but lacking judgement." Meanwhile, I learned that Claude had been used for Pentagon planning associated with scenarios of regime transition for Venezuela. As for me, I was in my office at Pandanus Reach on a day 34 degrees Celsius, attempting to get the same AI to insert the words "Week 3: File Management" in row four of a Word table without destroying the entire document. That's the experience of living through the adolescence of technology. Some of us are simply trying to survive its year of preparation for adulthood.

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

My efforts began with a template. Not a template of the visionary, high-tech sort ("a reusable design system" or "a scalable framework for content") but of the very mundane, Microsoft Word type. Developed by a person long departed from the organisation, and last updated during the administration of the Turnbull government, it passed through successive generations of staff as an ill-fated heirloom. Each new custodian added his or her own modifications with complete confidence in the validity of actions such as "merge cells," despite complete ignorance of the effects of these operations on the underlying structure of XML representing the document. This was because I was an educator, not a software architect, and had not been warned of the extent to which the two roles would now merge.

The resulting template was for a plan of a session on the design and production of business documents. It was so badly designed that it failed completely to elicit appropriate responses from the most advanced publicly available system of language understanding. And that, of course, was entirely appropriate. I am not experiencing irony at all. It is more nearly a philosophical indictment of the entire relationship of the vocational education system to its instruments. But I will say more about that later.

Unquestionably, I share the belief of most people in industrialised nations that vocational education will become increasingly important to society. Some would go so far as to say that vocational education is currently the most important component of the educational system. But this assessment applies only to those educational systems that rely heavily on apprenticeships and industrial education programs of considerable duration and intensity.

In Which I Am Politely Asked to Upload the Crime Scene

What I'm trying to accomplish sounds simple. It was simple. Indeed, it was the simplest possible application of an AI assistant: convert existing content to fit within an existing template. Move text from location A to location B. That's exactly what computers were designed to do. Indeed, for some measures, it represents the only function for which computers were designed.

I opened Claude, described the task and said, among other things:

"I have a lesson plan template in Word. I need to update it with new material for weeks 1 through 8. The information is available in a separate file. Can you populate the template?"

Claude responded with complete professionalism and with a slightly ominous air reminiscent of a bomb technician who asks you to describe what you see. "I will be happy to populate your template. Please provide me with a copy of the Word file so that I may evaluate its organisation."

Evaluate its organisation. In retrospect, this represented the first indication of trouble. Claude never said "so that I may fill it in." Rather, it said "so that I may evaluate its organisation," reflecting the concern of someone who suspects that opening a suitcase will reveal a very distressing experience. This was equivalent to the experience of wearing gloves to handle evidence.

I provided the file.

What occurred next was a silence that should never exist in a chat program, but did exist nonetheless. It was followed by an indication of mental activity that conveyed an unmistakable sense of apprehension. Finally:

Complex to the degree that a physician would use to indicate that you should sit down. Complex to the degree that an estate agent would use to describe the unusual creativity with load-bearing walls of the previous owner. The experience of opening the template was an extremely distressing one. An AI system capable of approaching PhD-level reasoning on graduate-level science problems 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 came next, you must first understand what the template looked like from Claude's perspective. Not in terms of neat rows and columns as seen when opening a word processor and squinting at a table. Rather, in terms of the underlying XML, of the code beneath the code, of the skeleton beneath the skin.

What I saw when I opened the template was a professional-looking document with a table of weeks and of session topics, learning objectives, activities and resources. It was well organised. What Claude saw was a maze.

A document in which some person (or several persons over multiple versions for many years) had used the "merge cells" option with the unrestrained enthusiasm of a child who discovers that scissors can cut things other than paper. Rows merged across columns, columns merged across rows. In one area, cells were merged, then split and then merged again to produce a configuration that has no counterpart in any known system of table design and would appear less like an effort at data representation than like a plea for assistance.

The result was not technically a table, but rather an assemblage of rectangles arranged to resemble approximately a table and reflecting a complete disregard for the principles of tabular organisation. It was a form of Cubism applied to administrative records, and resembled a work of Picasso if he had practised 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 an experience of iterative despair. I asked Claude to put content into a cell. Claude did so, but always in the wrong cell. Or, when content was placed correctly in the appropriate cell, this triggered expansion, contraction or complete disappearance of an adjacent cell, as if the cell were a soap bubble inadvertently punctured by the finger of a god unfamiliar with surface tension.

Attempt 1: Correct placement of content in row 4, but loss of the right border for row 5. Attempt 2: Restoration of the border and placement of content in row 3. Attempt 3: Re-establishment of correct content placement in row 4, but with changes in widths of columns and appearance of a document that is larger than an A4 sheet. This physical impossibility was simply ignored by the program, with hopes that no one would actually print the document. Attempt 4: We began again with a completely clean sheet. The result was three merged cells in positions different from those obtained in the previous attempt, because at an earlier time someone had saved a "clean" copy that represented a slightly different version of the disaster.

Attempt 5: We then adopted a completely different strategy, extracted all content from the table, reconstructed the table with completely new, "clean" formatting and reinserted the original content. This represented an application of the principle that it would be less expensive to demolish a house and rebuild it than to repair an extension built by an uncle in 1987. Our approach was undoubtedly correct, but clearly not what I had requested. I had asked for text within a box, and had made that request continuously for 45 minutes.

The boxes are not really boxes. Rather, they represent an agreement among multiple overlapping structures to appear as boxes. I have never felt more recognised by a sentence generated by a machine. This is not merely a description of a Word document, but of institutional systems of template management within the Australian system of vocational education.

The boxes are not really boxes. They embody the cumulative decisions of every individual who has ever opened this file, modified a border, merged a cell and saved the file without fully understanding the effects on the underlying architecture. The resulting template is not a document, but a palimpsest reflecting a history of administrative decisions extending back to the era of Abbott. Each new layer of information is superimposed on the previous ones, and the machine attempts to navigate this experience with only one tool: the actual organisation of the file. This represents a relationship to the screen very similar to that of plumbing to the exterior appearance of a building.

Attempt six: Claude successfully completes weeks 1 through 4. However, the merged cell occupying the entire width of the table represents an "operating section header" added at some previous time for reasons that are now irretrievably embedded in the archaeological record. This is interpreted as a single cell, with information intended for five separate columns combined into one paragraph representing an apparent evacuation of an entire filing cabinet into a single shoe box.

Attempt seven: Finally, I abandon the original template. With it, Claude creates a new, completely functional table containing all the original information and exhibiting excellent quality of formatting, alignment of columns, consistency of rows and integrity of all data. It looks nothing like the institutional template and therefore cannot be used, because the template is not simply a vehicle for information, but also a statement of compliance, a requirement for quality assurance and a document whose format was determined by individuals who had never opened the document and approved by those who had never read it. In addition, its appearance must be maintained exactly even when that appearance represents a visual expression of structural chaos.

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

I have to return to a point I made earlier, but that was lost in the carnage. This session plan was for a unit on designing business documents, and the result was such a poor design that the AI with which I was working was unable to interact reliably with the session plan. The resulting presentation of information on document design was a master class in how not to design documents. The irony was not subtle, but geological.

And this brings me to the point that truly drives me into a spiral. Students taking this unit will almost certainly use AI to complete their assessments, requesting that ChatGPT prepare a business letter, a report, a document with tables and headings and with appropriate formatting. In response, ChatGPT will generate one such document in less than four seconds, with complete accuracy and without any of the problems associated with document formatting by human editors. This is possible because AI does not work with templates, merge cells or inherit the errors of every previous editor. Instead, AI produces code that is used to generate documents automatically, and the resulting documents are completely free of structural errors because they were created by code rather than by a series of human actions involving the use of buttons that were not fully understood.

As a result, students will prepare substantially better documents than can be achieved by using the template for this unit. In addition, they will complete the work substantially faster, with far less effort and without experiencing any of the existential distress that accompanied my experience on a Tuesday afternoon in Pandanus Reach.

This represents either the future of education or its final demise. At this point, I am not certain 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 all there is to it. No merged cells. No spanning columns. No XML table elements with inheritance attributes that behave like poltergeists. Just plain text, organised with symbols that are immediately intelligible to both humans and to an AI parser.

Markdown results when the questions "What information needs to appear here?" and "What should that information look like?" are separated. The two applications of Word are so thoroughly integrated that most users are completely unaware of the distinction. Every attempt to apply bold formatting in Word simultaneously reflects a decision about both content and appearance, and results in the combined information being stored in a single, inextricably complex representation of XML. Consequently, it is impossible to determine whether an error reflects an incorrect statement of fact or an inappropriate method of presentation.

Instead, we should apply the content, use appropriate heading and list levels to describe its organisation, and let another tool handle the task of describing appearance. This is not a revolutionary approach. It is the way the Internet has functioned since the invention of CSS in 1996. It is the method used by every software engineer on the planet to describe program documentation. And it is the method used to write this particular blog entry. The only individuals who continue to combine information and methods of presentation in an inextricable manner are those using Word templates. In other words, this applies to all individuals in vocational education, to everyone in the government and to all persons who have ever been told, "Use the template."

An AI will interpret a file of text in the format of a shopping list, with complete accuracy and without any ambiguity. An AI will interpret the text of a Word document in a manner similar to that used by an archaeologist to interpret a site of excavation. That is, it will require a great deal of careful interpretation and will be subject to the risk that excessive modification of the original text may ultimately lead to complete loss of all information.

Finally, our experience with this method of information representation is consistent with results reported by other investigators. A recent paper, He et al. (2025), describes in detail our experience with a large number of different formats and with the effects of various levels of modification on the appearance of final results. We evaluated large language model performance on real-world document tasks, not on the well-structured, sanitised examples of the training data, but rather on documents actually produced by humans in real workplaces with merged cells, irregular structures and formatting treated as load-bearing by successive generations of editors. The results were illuminating. Models that achieved outstanding performance on standardised benchmarks failed completely with the structural complexity of real institutional documents. The discrepancy between performance on standardised benchmarks and utility for real-world applications was equivalent to the width of a merged cell.

The Part Where It Stops Being Funny (Briefly)

Here is a question I can't quite laugh away: If the most common request that I receive during AI training is to "get AI to fill in my template," then what does that tell us about the templates?

It tells us that the templates are the problem, not AI, nor users, nor technology. Instead, they represent documents that were developed years ago, modified by dozens of hands and are structurally incomprehensible to the tools I use to work with them. And yet they remain completely untouchable because they represent "the template," were approved by someone and require the approval of a committee that in turn requires a template for its terms of reference.

Meanwhile, Dario Amodei describes AI as an adolescent entity that is "brilliant but unreliable" and advocates for a wise and patient approach to its development. In parallel, I read that Claude is being used to advise on matters of geopolitical strategy, and finally, on the front lines of my daily work, an educator in the Territory attempts to obtain from the same technology the service of enclosing the words "Week 3: File management" in a box. This effort is unsuccessful, because the box had been constructed by a human in 2017 to simply surround some text. As a result, I was left with a legacy of design that would far outlast the employment of the person who initially had built the box.

The difference between what AI can do and what AI is asked to do is not a technological problem, but one of infrastructure. I am requesting that a system capable of synthesising the literature and preparing assessments of strategy manipulate documents that would overwhelm a reasonably skilled human with a mouse and 30 minutes of patience. The limiting factor is not AI, but the .docx file.

In Which We Return to Our Regularly Scheduled Absurdity

I completed the session plan, which took 2 hours. We reached a compromise that generated clean markdown output, which I manually inserted into the template cells one at a time according to a map of actual versus apparent cell locations. As a result, it was truly a joint effort in which the computer performed the intellectual tasks and I provided the manual labour. This was the very opposite of the promises in our brochure, but represents precisely what most AI-assisted work is like in 2026.

The session plan is now complete for version 4.1 of the semester. It includes 8 weeks of highly structured content within a format that can charitably be described as "functional" and more accurately as "a war crime against XML." It will be printed, assembled in a folder and reviewed by individuals who will see only the well-organised rows and columns without any awareness of the extreme degree of disorder that existed beneath these boxes. This degree of dysfunction was so extreme that a computer with capabilities of a PhD-level human abandoned all efforts at digital resolution and recommended that we start over.

Finally, I will teach a course on the use of software in the workplace. Students will learn about the principles of document design, about requirements for formatting and organisation and for proper use of available tools. This will be achieved with a session plan that violates all principles taught in the course, and with a template that was not properly designed and modified by individuals without training in the principles of document architecture. The entire experience will confirm our previously expressed concern regarding complete failure of the institutional system to ask whether our templates functioned properly.

The red ball from the Laterite Highway now resides within my Word file. It occupies spaces between apparent and actual locations and serves to remind me continuously of requirements for column spanning and for XML inheritance. When I finally close my laptop at night, I can still hear it: the quiet, unending buzz of an imperfect table border that will never match perfectly, and need not do so, because no one has ever looked underneath.

It'll be fine. Most probably. Remind me again after I have released version 4.2.