28/3/2026 ☼ AI ☼ meaningmaking ☼ not-knowing ☼ Knowledge Management ☼ content ☼ Content Strategy ☼ information
tl;dr: Content and representation are different things, and re-presentation is different again: selecting from and translating a body of content to produce new artifacts for different audiences and purposes. LLMs make re-presentation cheap and fast. So document management should now mean caring for and feeding totipotent source documents — structured for completeness, not for any specific reader — from which purpose-built product documents are generated on demand. This creates a quality problem that only meaningmaking (human subjective reasoning) can address.
A few weeks ago, I ran a two-day strategy retreat for a public-sector research organisation. Twenty-three senior leaders in a conference room for two days, working through an organisational change that affected every division. Breakout groups, plenary synthesis sessions, workflow redesigns, a ninety-minute unplanned discussion about what their organisational units actually are as conceptual entities. I took extensive notes.
During and after the retreat, I wrote a single internal memo. It ran to fourteen thousand words. It captured as much of what had happened as I could recall (in design intent, it was like an ethnographic fieldnote combined with synthetic memo): the frameworks I had introduced and how participants responded to them, the disagreements that surfaced (and the ones that didn’t surface but were detectable in expressions and body language), each breakout group’s outputs, my own cross-session synthesis, the workflow redesigns the groups produced, the specific moments where something shifted in the room, and lots of other stuff. The memo was not “raw” content. It was structured with a particular purpose: to be comprehensive enough and organised enough that I could later produce any deliverable the engagement might require. It was designed not for a specific reader but as the source from which specific-reader documents would be produced.
From that document, I produced five distinct deliverables within about a week. A client-facing memo, restructured around ten strategic issues with recommendations, stripped of all internal process observations and facilitator commentary. A two-page chronological summary for quick leadership briefing. A short set of key discussion points I sent by email ahead of a debrief meeting. A prep document for a workshop on selection heuristics. An outline for a slide deck (for another use entirely).
Each of these selected from the source document and translated it for a different reader, a different purpose, a different level of detail. The client memo presented findings and recommendations; the internal memo also recorded how I got there and what I thought the group missed. The two-page summary preserved chronological flow; the client memo abandoned it entirely in favour of issue-based organisation. The key discussion points were three paragraphs; the source document was sixty pages.
None of these product documents was the source document. And the source document was never meant to be read by any of those audiences directly.
Here is a second example, different in kind.
FOUNDATION is an interactive game exhibit I designed while at MOD., a museum of discovery in Adelaide. Players work together to build structures, but they discover the available materials, the rules, and the consequences of their choices only by playing, because prior generations of players have already changed what exists. There are no known probabilities, no fixed option sets, no calculable expected values. It is a physical experience of what genuine not-knowing about actions and outcomes feels like from the inside.
The project generated substantial written material: long writeups documenting the design logic, learning objectives, theoretical grounding in my research programme on not-knowing, the connection to an exploratory project on games that create genuine uncertainty rather than just risk. These writeups are dense, technical, and assume familiarity with the underlying theory.
From those writeups, I generated a version for the museum wall text: a plain-language explanation of what visitors had just experienced, written so that a twelve-year-old could follow it. It explains what not-knowing is, why it matters, how the game created it, and what to take away from the experience. It does this without the word “epistemology” appearing even once.
This is a different kind of transformation from the retreat case. The retreat produced product documents with content and synthesis targeted at different audiences (client, leadership, workshop participants). The museum wall text was a product document which used simpler vocabulary to explain the same core concepts: the ideas didn’t change but instead were restructured for a fundamentally different level of prior knowledge and vocabulary. Both involve taking a set of source documents and generating something new from it, but the transformation is different in character.
The technical content industry has been circling this idea for years, under different banners, without stating it as a generalisable principle. Single-source publishing, DITA’s modular topic architecture, XSLT transformation pipelines, headless CMS architectures, and the newest wave of AI-integrated content platforms all point the same way: structure your content once, transform it for each context.1 Each got something right. None accounts for what happens when the transformation step is no longer rule-based and pre-specified but generative and prompted.
Dan Petrovic frames the LLM as a “presentation layer” sitting on top of traditional retrieval infrastructure: crawl, index, retrieve, rank, then present via LLM. His claim is about search architecture, not about documents or authoring. But the underlying intuition transfers: the LLM transforms existing content into a new form for a specific context.
Before LLMs existed, Ann Rockley’s intelligent content framework held that content should be able to change—as Robert Rose explained it, “not just in appearance but in substance, according to the person receiving it, the device, the channel, the time, the location.” Rockley’s framework went beyond conditional filtering: it included audience-specific content variants and content that genuinely changed in substance based on user context. The concept of content adapting for different readers is not new. What is new is the tooling and the mechanism it enables—and that the transformations no longer need to be specified in advance.
The separation of content and presentation has been a live idea in computing since at least 1967, when William Tunnicliffe proposed separating information content from format. Sixty years of work—SGML, XML, HTML+CSS, LaTeX, XSLT—established the principle: mark up what something is, and let a separate system decide what it looks like.
Word processors make it easy to ignore this. They let you define styles that separate structure from appearance—but it is so much easier to just make something bold or change the typeface directly that most people do, fusing content and representation in ways that are hard to undo.
Markup languages enforce the separation. <h2> in HTML or \section{} in LaTeX declares structure; a stylesheet handles rendering. The content exists independently of any particular visual form.
Up to now, we’ve had these two concerns in document design: content (the substance of what you’re saying) and representation (the encoding of how that substance is displayed). One source document can be represented in many ways.
What would taking this seriously actually mean for how we work? This is now possible because of LLMs.
LLMs introduce a third concern aside from content and representation. We can call it re-presentation. It is an operation: selecting from a body of content and translating what is selected to produce a new artifact for a specific audience, purpose, or context. The operation is parameterisable: you specify the audience, the tone, the level of linguistic sophistication, the desired outcome, the scope. The result of the operation is the re-presented artifact: a product document.
Re-presentation is not just reformatting or restyling. It is three operations.
- Selection: choosing what to keep and what to leave out.
- Translation: using different words, a different register, a different level of detail to describe what is kept.
- Restructuring: organising the selected and translated material in a way that serves the target context.
The product document is a transformation of the source document, and it is a new artifact in its own right, generated on demand.
“Representation” in the traditional sense means encoding how substance is displayed: this is a heading, that is a list, this is emphasised, that is a caption. “Re-presentation” means presenting again, differently, for a different purpose. The retreat memo re-presented as a client document. The design writeup re-presented as museum wall text. The structured content re-presented as a conversational search result. Each time, the operation produces a new artifact that is derived from but not identical to the content in the source.
Content is what you know. Representation is how content is structurally encoded. Re-presentation is how content is transformed into an artifact for a particular audience.
Rockley’s adaptive content and DITA’s transformation pipelines are precursors—both could restructure content for different contexts—but both required the author to pre-specify variants and selection conditions against schema-validated sources. LLM re-presentation allows the transformation to be specified in natural language at the point of need, producing non-deterministic outputs from loosely structured sources. The difference is not in what is being done but in how it can be specified and what it can operate on.
So “document” should now be thought of as two things.
A source document is a structured, machine-interpretable pool of material. It is your memos, project writeups, research logs, session syntheses. It shouldn’t be “raw” notes; it should be organised with a particular purpose, which is to be comprehensive enough and well-structured enough to serve as the substrate for producing specific-audience documents later. The retreat memo was a source document—not a brain dump but a designed artifact, structured not for any particular reader but for the production of documents for readers not yet specified.
A product document is the artifact generated by a re-presentation operation: produced by selecting from and translating the source for a particular audience, purpose, and context. The client memo was a product document. The two-page summary was a product document. The museum wall text was a product document. Each was derived from a source document but was not the source document.
This means LLMs now also change how humans should manage content. Anyone managing a corpus of their own text should now consider writing source documents for completeness and structure, not for readability. Accept that they will be verbose and probably boring. No one is meant to enjoy reading them. They are a manufacturing substrate, not a finished artifact. But they aren’t unstructured — and the kind of structure matters.
The retreat memo was useful as a source document because of its comprehensiveness and its structural discipline. It didn’t just record conclusions: it recorded the reasoning behind each conclusion (the three alternative organisational structures the group considered, what each would have broken, the specific argument that settled the question), and it marked disagreements explicitly (who disagreed, about what, whether the disagreement was resolved). This mattered because different product documents needed different things from the same source. The client memo needed only the recommendation and a paragraph of justification. The workshop prep document needed the full set of alternatives, because the workshop was about stress-testing the choice. A product document for internal use needed to surface the unresolved tensions; a product document for external use needed to omit them cleanly. None of this is possible when a source document records only what was decided, or buries live debates in hedged prose (“there was some discussion about whether…”) that an LLM cannot reliably distinguish from settled matters.
The FOUNDATION writeups show the same thing from a different angle. They used a consistent structure throughout: design rationale, then learning objectives, then theoretical grounding, applied the same way in every section. The retreat memo did the same (session label, then topic, then reportage, then synthesis). This consistency is what lets a prompt like “extract all unresolved design questions” actually work: it works when the source document marks a section as “open questions” every time; it is forced to guess (or hallucinate) when the same information appears sometimes in a named section, sometimes in a narrative paragraph, sometimes in a footnote. Both source documents also included context that the author might have taken for granted: the retreat memo explained “the restructuring proposal from February” rather than assuming the reader already knew what it proposed, because an LLM generating a client memo cannot explain what the source document does not contain.
The pattern: a source document that captures reasoning, marks disagreements, uses consistent structure, and fills in its own background assumptions will produce product documents that are genuinely useful to their audiences. A source document that captures only conclusions will produce product documents that are fluent but shallow. The care and feeding of source documents is a craft and an art that is only just starting to coalesce, but there already are principles that make a source document more effective for re-presentation:
- Capture reasoning alongside conclusions, so that product documents can include or exclude justification as needed;
- Separate observation from interpretation, so that findings and inferences remain distinguishable;
- Mark claim status explicitly — settled, contested, hypothetical — so that nothing shifts register silently during re-presentation;
- Use consistent structure across sections, so that prompts can target material reliably;
- Fill in background assumptions and context (or link to them), because an LLM cannot supply what the source document does not contain.
These principles go beyond formatting rules. They are design principles that determine what re-presentation can do with the material — and, as I’ll argue below, they also determine whether a product document can be evaluated.
Re-presentation using LLMs changes what work humans must do. Product documents become ephemeral, purpose-built, and cheap to generate. The investment shifts. Instead of spending hours polishing a final document that tries to serve multiple audiences (the memo that is somehow both comprehensive and concise, both technically detailed and leadership-friendly), the human in the loop now spends effort caring for and feeding rich, well-structured source material. The product documents flow from it as needed. The craft shifts from document production to building and maintaining totipotent source documents.
On the other side of the re-presentation operation there’s more human-only work. Comparing a product document with its source document tells you what a re-presentation operation did — which sections it drew from, what it omitted, whether a qualification survived the translation. But it cannot tell you whether what it did was appropriate. That is the handoff: from structural evaluability to subjective evaluation.
Source document design makes evaluation possible. But possible is not the same as easy, or automatable.
DITA publishing pipelines illustrate the contrast. Because the output is rule-governed (same inputs and same configuration produce the same output every time), you can build a review process around a known, stable output. Review the rendered deliverable once, approve it, and every subsequent build from the same source produces the same result. Quality control is built into the architecture: the review surface is fixed, so the review process is repeatable.
LLM re-presentation destroys this guarantee. The output is non-deterministic. The same source document and the same prompt may produce different product documents on different runs. A product document generated by an LLM may omit critical information, misrepresent the source document, introduce errors, or shift emphasis in ways that distort the original. Five product documents from one source document in an afternoon is real. But each of those five must be checked, and the checking must be done by a human.
This is where structural evaluability ends and meaningmaking begins. Two questions arise that no structural comparison can answer: Is this good?” and, Is this worth shipping?
These are different questions requiring different actions with different failure modes. The first (Type 1 meaningmaking) is a judgement about quality: “this client memo is clear” or “this translation distorted the argument.” The second (Type 2 meaningmaking) is a judgement about value relative to cost and stakes: “this product document is good enough to send” or “the risk of getting this wrong means it needs another pass.”
Both types of meaningmaking work are involved in working with re-presentation. First, we have to ask “Is this product document good or bad?” Was the selection right for this audience (has anything critical been left out, or has too much been kept)? Was the translation faithful (have the ideas survived the shift in register, or has a simplification distorted the substance)? Does the framing serve the reader without misleading them? Type 1 evaluation can be partially supported by structural comparison against the source document: the five design principles make this tractable. If reasoning is captured alongside conclusions, you can check whether a qualification was dropped. If claim status is marked, you can catch a hypothesis that became an assertion. Good source document design makes Type 1 meaningmaking easier, though it never makes it automatic.
Then we have to ask “Is this product document worth shipping, or does it need another pass?” This is a different question, and source document structure cannot help with it at all. Type 2 requires judgement about the audience’s needs, the stakes of getting it wrong, the context in which the product document will land. None of this exists in the source document. Some of what matters cannot be written down, like the gut feel a good account manager has for what will land for their account’s main decisionmaker. This is why a client memo might be structurally faithful to the source and still be wrong because it sequences the findings so that the least important recommendation leads and the one that really matters to that specific client is buried on page three. Only someone who understands what that specific client needs can catch this, and that understanding is not in the source document or in any prompt.
The meaningmaking work extends backward, into the design of the re-presentation operation itself. When you write a prompt that says “rewrite this for a leadership audience that needs the conclusions without the process,” every word of that prompt is a meaningmaking decision. Who counts as “leadership”? What counts as a “conclusion” versus “process”? How much supporting logic should survive the translation? What tone is appropriate? These are not parameters you can set once and forget. They are subjective choices that shape the artifact, and getting them wrong (or leaving them out) may produce a product document that is fluent, plausible, and still misses the mark.
Re-presentation makes three types of subjective reasoning work more important, not less: the meaningmaking work of
- Designing the re-presentation operation in the first place (specifying what the transformation should do, for whom, and to what end),
- Checking whether a re-presented artifact is faithful to the source, and
- Judging whether the artifact is fit for purpose.
Re-presentation is now cheap. The meaningmaking on either side of it is not.
In 1997, Michael Buckland asked “What is a document?” His answer was that the concept is far more fluid than we assume: the boundary between document and non-document is social and functional, not material. Ted Nelson’s Project Xanadu envisioned documents as composites assembled from fragments in a universal address space: one pool, many views, but without transformation (transclusion is exact quotation with provenance; the fragment is maintained at source, not copied and altered). Re-presentation adds what Nelson’s model lacked: the fragment can change.
And re-presentation is what well-trained humans already do.
A good consultant who facilitates a retreat can brief an executive team on it in two minutes, write a detailed client memo in several hours, or run a follow-up workshop. Each time, they are selecting from what they know and translating it for the context. The executive team gets the three things that matter and nothing else. The client memo gets the findings with enough supporting logic to be actionable. The workshop gets the unresolved questions reframed as exercises. Same source document, different product documents. The consultant is performing the re-presentation operation: selecting what to keep, translating it into the right register, generating an artifact suited to the context.
This skill takes years to develop. It requires judgement about what to include, what to leave out, what to restructure, what register to write in. It is bottlenecked by human attention and time. A consultant can produce perhaps five or six product documents from a single engagement before diminishing returns set in: not because the source is exhausted, but because the consultant is.
LLMs do not replace the judgement involved in deciding which re-presentation to produce, what the source material should contain, or whether the result is good. The retreat memo was useful as a source because I knew what to observe, what to write down, what frameworks to apply. No LLM decided that. And each product document required my review before it went anywhere. But LLMs collapse the production cost of the re-presentation operation itself to near zero. The constraint moves from production capacity to source document quality and evaluation quality.
If a knowledge repository is no longer a collection of finished documents but a substrate from which documents are generated on demand, then several things change.
Effort shifts from polishing outputs to enriching the substrate. The question “is this document good?” is replaced by “is this source document rich enough to generate the product documents we will need?” A source that is comprehensive and well-structured is more valuable than a source that is polished but thin. Completeness and structure beat elegance.
The status of the “final document” changes. If any product document can be regenerated from the source, then the product document is disposable and the source is the asset. Version control matters for sources, not products. Archiving matters for sources, not products. The institutional memory lives in the source layer.
New questions arise that the model does not automatically answer. Who controls the source? Who decides which re-presentations are authorised? What is the audit trail from product back to source?
The most interesting question is what happens when two product documents generated from the same source contradict each other—not because the source is wrong, but because the LLM made different selection choices. A client memo that foregrounds one recommendation and a leadership briefing that foregrounds a different one are both faithful to the source; neither is wrong on its own. But if the client and the leadership team compare notes, the organisation appears to be saying two things. This is not a bug in the LLM. It is a structural feature of non-deterministic re-presentation, and it requires a human decision about which selections are authorised for which audiences.
These are governance questions, and any content operations team will ask them before adopting this way of working.
And the skills that matter most are not writing skills (though those are important) but two others. First: observing well, knowing what to put into the source in the first place. The richness of the source determines the range and quality of products it can generate. Having ethnographic training helps. Second: evaluating well, knowing how to judge whether a re-presented artifact is faithful to the source, appropriate for its audience, and good enough to ship. The production step is automated. The extensive meaningmaking work around it is not, and this meaningmaking is where the real work now lives.
This is what I observe in my own practice and in the way the industry is talking about LLMs. But there’s not much systematic evidence of widespread adoption. The content engineering community has the infrastructure. The LLM ecosystem has the mechanism. The conceptual model that connects them (content, representation, re-presentation) is what’s been missing. Naming it makes it possible to think about it clearly, and thinking about it clearly makes it possible to do it well.
This essay draws on a series about not-knowing and a series on meaningmaking and AI.
My first book, The Uncertainty Mindset, developed frameworks for organisational design under uncertainty. IDK is my tool for making not-knowing productive.
I’ve spent the last 15 years investigating how organisations can succeed in uncertain times. The Uncertainty Mindset is my book about how to design organisations that thrive in uncertainty and can clearly distinguish it from risk. Part 6 of my book contains detailed descriptions and case studies of how organisations can use desperation by design to build capacity for sustained innovation.
I’ve also been working on tools for learning how to be productively uncomfortable. idk is the first of these tools for productive discomfort.
Single-source publishing and DITA established the pattern: author modular topics with semantic markup, publish to multiple output formats via conditional processing and XSLT transformation. Paligo argues that structured authoring produces “AI-ready” content; RWS makes a similar case. Headless CMS architectures (content as structured data via API, presentation decoupled) push the separation further; Hygraph and FocusReactive describe systems where AI agents handle transformation within the content architecture itself. Karen McGrane argues that LLM remixing demands even more structured, semantically rich content.↩︎
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.