The augmentation tradition in hypertext research, from Bush's associative trails [8] through Engelbart's programme of human-intellect augmentation [12], is grounded in a consistent intuition: tools for thought work best when they make structure visible and keep humans actively engaged. This places the tradition in tension with the dominant engineering instinct, which treats friction as an obstacle to seamless interaction [22]. Chalmers and Galani's seamful design challenged this orthodoxy, arguing that exposing a system's seams can be a resource rather than a deficiency [10], and subsequent work elaborated this into productive friction: deliberate resistance that prompts reflection at critical junctures [11]. Recent empirical work reinforces this: Wu et al. found that a graphical, nonlinear interface for LLM interaction improved both creativity and synthesis quality compared to linear chat, yet received lower usability ratings [28]. The cognitive engagement that productive friction demands was experienced by participants as added difficulty, even when it produced better outcomes.

Liu and Almeda introduced hypertextual friction to describe this dynamic in contrast to algorithmic systems that foreground ‘outputs driven by computational models or agents’ [16], proposing it as a design stance centring friction, traceability, and structure. I build on this foundation by presenting a system that embeds hypertextual friction in the practice of hypertext co-construction: jointly building a persistent hypertext with an AI, link by link, as the mechanism by which friction is generated and sustained.

The Zettelkasten method, developed by the sociologist Niklas Luhmann, treats knowledge not as a collection of documents but as a network of atomic notes (each capturing a single concept or claim) connected by explicit links [17]. Luhmann described his slip-box as a genuine intellectual interlocutor: its value lay not in the notes themselves but in the unexpected connections that emerged through navigation and accretion over time. This relational, emergent character distinguishes the Zettelkasten from folder-based filing systems and from the linear document as a unit of thought: structure arises from use and connection-making rather than being imposed in advance.

Digital implementations of the method, most fully articulated by Ahrens [1] and often instantiated in modern Personal Knowledge Management tools, preserve this relational logic whilst adding bidirectional linking, graph visualisation, and programmatic extensibility. It is within one such environment, Obsidian, that the system described in this paper is built. The specific design is described in the System Overview (Section 4), but the reason I chose a Zettelkasten is that its systematic, granular externalisation of associative structure is precisely the kind of environment in which agentic AI can thrive—creating the opportunity for co-construction that generates productive hypertextual friction rather than seamless automation.

Previous work has argued that delegating writing to AI constitutes cognitive offloading with long-term costs (AI as the Thief of Reason [20]) and empirical work supports this, finding cognitive offloading to be a measurable mediator between AI use and reduced critical thinking [13]. The challenge is compounded by LLM sycophancy [18]: models trained to please users mirror rather than challenge, reducing apparent collaboration to unidirectional delegation. Emerging frameworks address this by tracking authorship drift [23] and defining roles for human–AI writing collaboration [15], but these remain focused on the prose itself, who writes which sentences. This paper shifts the site of collaboration from text generation to structure: building persistent hypertext rather than producing prose, and keeping the human as author of the knowledge network rather than editor of machine output.

The landscape of AI-augmented note-taking tools divides into two broad categories. The first and most visible comprises tools in which AI operates as an interface to a document corpus: Google's NotebookLM and Microsoft Copilot Notebooks offer powerful summarisation and question-answering, but their dominant interaction paradigm remains the linear chat. Structure is presented to the user as a finished output rather than built by the user as a persistent artefact; sessions are ephemeral and the human plays no constructive role in organising what the AI produces [19, 26].

The second category is considerably closer to this paper's approach. Tools such as Roam Research, Tinderbox, Logseq, and Obsidian with AI plugins require the user to build bidirectional links and graph structures that persist as navigable artefacts; AI then suggests connections and surfaces related material within that human-made structure. Atzenbeck et al. demonstrate that spatial hypertext structures can interrupt the passive consumption induced by algorithmic interfaces and restore deliberate engagement [5]—the difference between structure delivered and structure made. It is this combination of human intention and AI initiative, enacted through persistent co-constructed hypertext, that this paper examines.

I am a computer science academic with over 25 years of research experience in hypertext and web science. High technical literacy meant that configuring Claude Code with Obsidian presented few barriers, and longstanding familiarity with knowledge structures meant the Zettelkasten model felt conceptually natural. I had used Obsidian for several years prior, making its local Markdown file structure an obvious fit for an agentic AI.

This dual role carries implications: design decisions reflect existing values, and reflections on the system are shaped by the frameworks it was built to explore. This circularity is not a limitation to be apologised for—it is constitutive of the method, and its consequences are examined in the Discussion (Section 7). The collaboration is best characterised as a senior-led partnership: I retain direction and final authority over all structural and interpretive decisions, whilst the AI contributes summarisation, concept extraction, link proposal, and narrative generation—analogous to a principal investigator working with a research fellow.

The project began in November 2025 as an experiment in using Claude Code directly with Obsidian's Markdown files, motivated by a desire to explore what AI-assisted note-taking might look like when designed with human agency at its centre. Two phases are discernible. The first established the core Zettelkasten structure and built commands for AI-assisted content and link creation. The initial thematic focus (AI writing and knowledge work) was chosen based on interest, it was serendipitous that this became relevant to its own design, creating a productive feedback loop.

The transition to the second phase was marked by a specific act of reflexivity: using the vault's own conceptual resources to derive an explicit values document, then auditing the system against those values. This process is described in detail in Section 5. The system as described in this paper reflects the state of the vault at the time of submission, in April 2026, approximately six months after the project began.

It should be noted that the toolchain itself evolved continuously across both phases, with changes to Claude Code's behaviour, Obsidian plugins, and the vault's command and skills architecture. Rather than treating this volatility as a confound, the study embraces it: tool and practice co-evolution are part of the phenomenon under investigation, not threats to its integrity.

Two documentary sources inform the analysis. The first is the vault's daily history: AI-generated narrative nodes recording content added each day, grouping notes into thematic clusters and narrating how they relate to one another and to existing vault content. These are interpretive artefacts, not neutral logs—themselves part of the structure they document. Each note links back to its history entry, preserving the chronological context of its creation.

The second is a rationale file: a contemporaneous record of design decisions (changes to commands, tools, and integrated services) appended at my discretion throughout the project. Full conversation transcripts were not retained; the rationale file preserves contemporaneous summaries written at the time of key interactions, limiting though not eliminating the risks of retrospective distortion. The vault contains only publicly available material and my own writing; no personal data about third parties is present, and no ethical approval was required.

The analysis proceeds through structured reflection and close reading of both sources, supplemented by direct engagement with the vault's current state. No formal coding scheme was applied; episodes were selected on the basis of their significance—principally moments of design pivot, values conflict, or structural change. The study did not surface negative cases in the sense of AI failure; the system consistently produced useful outputs. What did emerge were moments of human breakdown, occasions when I did not read history entries carefully or accepted outputs without critical engagement. These episodes are analytically significant: they demonstrate that hypertextual friction places requirements on human motivation, not only system design (Section 6).

Quantitative data (Section 6.1) is descriptive only; the study's contribution is interpretive rather than statistical. Rigour rests on triangulation across the history nodes, rationale file, and current vault state, and on treating moments of friction failure as analytically significant rather than ignoring them.

The vault organises knowledge through a three-layer hierarchy. At the base, atomic notes capture individual concepts—typically a single idea in 200–400 words, linked to related nodes through wiki-link syntax. The atomic principle ensures that connections represent genuine conceptual relationships rather than incidental co-occurrence. Hubs occupy an intermediate position as evergreen thematic nodes, aggregating clusters of related notes around persistent concepts. Hubs do not argue; they organise, acting as stable entry points. Maps of Content (MOCs) constitute the uppermost layer: narrative trails that sequence notes into argumentative arcs, drawing on Bush's associative trail [8]. MOCs organise existing knowledge and scaffold new writing; the arguments in this paper, for example, were first assembled as MOCs before being developed into academic prose.

This three-layer scaffold provides the structural grammar within which AI operates. Every generated content type fits one of these roles, making AI outputs legible, correctable, and navigable. Changes occur at the level of individual notes rather than wholesale rewriting, and explicit links make the scope of any revision visible.

Four main components combine to produce the system's capabilities, connected through the Model Context Protocol (MCP), which exposes each as functions within Claude Code's context (Figure 1).

Obsidian provides the hypertext environment. Notes are standard markdown files linked through wiki-link syntax ([[Note Name]]), which creates bidirectional connections that Obsidian tracks and visualises as a graph. The Dataview plugin enables dynamic queries over note metadata (for example, listing all atomic notes linking to a particular hub) without AI involvement. The vault uses a lightweight maturity model (sapling, fern, tree) to track note development, operationalised through connectivity and content completeness. Because the vault is a file system, it is directly accessible to external tools.

Claude Code is the primary LLM assistant, operating via the terminal with direct access to the vault file system. A CLAUDE.md file encodes structural conventions, writing style guidelines, and operational principles; persistent memory files maintain context across sessions. The bulk of the system's generative intelligence passes through this component.

Smart Connections is an Obsidian plugin that generates vector embeddings of all notes using a local model, enabling semantically similar note retrieval without keyword matching. Its functionality is exposed via MCP, making it callable by Claude Code during note-creation and linking workflows.

Gemini CLI provides headless large-document summarisation. When academic papers or transcripts exceed the scale at which Claude Code can efficiently operate, Gemini pre-processes the document into a structured summary (Gemini for extraction, Claude for reasoning) preserving token efficiency.

Zotero, the reference manager, is accessible via MCP for querying paper metadata and full text, enabling literature workflows to execute within a single coherent context.

The system operationalises this architecture through natural language commands (/command-name) that encode multi-step workflows, drawing on persistent configuration in CLAUDE.md and modular skill files. Note creation is interactive: Claude Code proposes notes within the conversation, identifies overlaps with existing content, and I respond by accepting, selecting, or redirecting.

Commands fall into three clusters. Knowledge Capture processes raw materials (ingest-resource) and academic papers (create-literature-note) into clusters of atomic notes, routing large documents through Gemini first. Academic Grounding searches the Zotero library for papers relevant to vault concepts (find-papers) and evaluates incoming papers for contribution value (evaluate-papers). Textual Production expands MOC outlines into narrative documents (expand-moc) or converts notes into blog posts and academic drafts.

At the close of note-creation sessions, Claude generates a daily history entry—a narrative trace of additions and connections that I review, and that each note links back to.

The underlying logic is one of structured decomposition and recomposition. Complex source texts are deconstructed into the smallest units of meaning the Zettelkasten can represent; those units are mapped against existing knowledge, revealing connections that span sources and disciplines; and the resulting network is then recomposed into new documents and narratives. A detailed walkthrough of this process appears in Section 6.2.

The system emerged over approximately six months of daily use, through a process best understood through Engelbart's concept of bootstrapping: using a tool to improve the tool itself [12]. The vault was built with the AI, not only by me. Each design decision arose from encountering a limitation during actual knowledge work, theorising a response (often using the vault's own conceptual resources) implementing a change, and then discovering new limitations exposed by the changed system. The autoethnographic method and the bootstrapping method converge here: the system is simultaneously the subject of research, the instrument for conducting it, and the medium through which design decisions are recorded and evaluated.

The earliest experiments combined Claude Code's file-system access with a conventional Zettelkasten folder structure. Templates constrained AI output into structurally consistent forms; a lightweight maturity model (sapling, fern, tree) tracked note development through connectivity and content completeness. These pragmatic decisions provided a vocabulary for discussing vault health without requiring subjective quality judgements.

The first genuinely consequential design choice was the suggest-only workflow: Claude Code proposes candidate notes and waits for explicit human approval before committing anything to the vault. Two concerns motivated this. First, a recursive explosion risk: without human intervention, note creation could suggest related notes indefinitely, producing unbounded growth with no intellectual filter. Second, the human serves as both filter and exit point, shaping the hypertext and providing natural stopping points. This pattern—the AI suggests, the human selects—became foundational, preserving my role as curator rather than consumer of AI output. The initial agent architecture delegated note creation to a specialist “zettel-gardener” sub-agent, separating analysis from creation, a design principle that later changed as my experience with the system matured (see Section 5.3.3).

One small design choice reveals the vault's unusual ontological position. In conventional note-taking practice, users occasionally tag AI-generated content to distinguish it from their own writing. In this vault, the logic reverses: a #human tag marks the rare instances of human-authored content within an otherwise AI-generated knowledge base. Paragraphs bearing this tag are protected — the AI is instructed never to modify them without explicit permission. The inversion is telling: the exception that needs protection is not the machine's contribution but the human's. Section 6 examines the practical implications of this design in detail.

Three categories of practical constraint — token cost, context fragmentation, and document scale — drove architectural changes that, in retrospect, also clarified the system's relationship to hypertext principles.

5.3.1 The monolithic instruction problem. Claude Code's behaviour is configured through a CLAUDE.md file loaded into every conversation. As conventions accumulated, this file grew unwieldy. The solution drew on Claude Code's skills system for progressive disclosure: a lean core references modular knowledge that loads only when triggered by a matching task. This reduced token costs, but also cleared the way for further vault expansion with new capabilities as these could now be added as new skills without interfering with existing skills.

5.3.2 Multi-LLM division of labour. The Gemini integration described in Section 4 arose from a recognition that not all tasks require the same depth of reasoning. Treating different language models as specialised tools rather than interchangeable general-purpose engines—Gemini for mechanical extraction, Claude for judgement and epistemological framing—produced token savings of 80–90% on long documents.

5.3.3 From sub-agents to shared skills. The original architecture used isolated sub-agents (child processes spawned to handle specific tasks, each running in its own context window). They could not share state, producing both token waste and architectural amnesia: the system's left hand did not know what its right hand had done. The migration to skill-based commands (where the main instance executes workflows directly, drawing on shared knowledge modules) carried a deeper lesson. Sub-agents kept knowledge in silos: a note-creation agent, for instance, could not draw on conventions established during literature processing. Executing commands wholly within the main Claude instance created one powerful context, and the lesson mirrored the vault's own organising principle: that knowledge gains value through connection rather than isolation.

As the vault grew, a subtler problem became apparent. Notes generated by Claude Code adopted an authoritative, declarative voice: “AI systems exhibit compositional abstraction layers”; history entries that “established” frameworks and “demonstrated” findings. For a Zettelkasten designed to hold multiple competing perspectives in productive tension, this represented an epistemological failure: the vault was collapsing pluralism into orthodoxy. The problem has structural roots in LLM training: reinforcement learning from human feedback rewards confident responses and penalises hedging, a tendency well-documented as a dimension of sycophancy [25]. When the tool that generates most of a vault's content systematically overstates certainty, the entire knowledge base inherits that epistemic stance.

The response was a set of language guidelines encoding epistemological pluralism as an operational principle. Notes now frame ideas as perspectives rather than facts (“This framework offers one way to understand...” rather than “AI systems exhibit...”), with attribution preserving the chain from source to claim. Definitive language is reserved for what a source explicitly claims, historical facts, and mathematical truths; everything else uses tentative phrasing that keeps the epistemic door open. The result is a vault that reads less like an encyclopaedia and more like a research notebook.

The epistemological problem was a symptom of a broader absence: the system lacked an explicit statement of what it was for. The values document emerged through a characteristic bootstrapping loop. I then used the vault to construct an essay—The Argument for Hypertextual AI Interfaces—assembled from atomic notes via a Map of Content, arguing that hypertext interfaces could address the cognitive risks of AI interaction. From this essay, ten operational values were derived (shown in Table 1 below).

Table 1: Ten operational values derived from the vault's own conceptual resources.

These values are an example of how such principles might be articulated within a specific research practice, not as a canonical set. The method matters more than the specifics: using a tool to theorise its own design, deriving operational principles from that theory, then applying those principles back to the tool.

The bootstrapping loop closed when I asked the AI to audit every workflow in the system against the ten values. The audit revealed a systematic gap: Value 6, Reciprocal Challenge, was underserved across all content-generation workflows. Every command followed the same pattern: gather context, obtain user approval on parameters, generate output. Between the moment of approval and the completed output — potentially thousands of words of generated text — there was no checkpoint, no opportunity for the system to question whether the approach was sound, whether the source material was ready, or whether alternatives deserved consideration. The AI was helpful and obedient but never genuinely collaborative in the sense of pushing back on assumptions.

The response varied by workflow: content creation gained pre-creation awareness checks (surfacing similar notes, overly broad scope, overloaded connection targets); writing synthesis gained structural previews at the outline stage, where changing direction is cheapest; literature note workflows gained reading strategies that direct attention toward material that challenges existing vault knowledge rather than confirming it; and history entries gained embedded reflection prompts—non-blocking observations about patterns in the day's work. Smaller changes adjusted how suggestions are framed (“capture the idea that...” rather than “create note about...”) and ensured that human-tagged content is never overwritten during surrounding edits.

The pattern is one of ongoing co-evolution rather than a single design event. The system wrote the essay that produced the values that changed the system — and then the changed system, through further use, revealed new gaps that prompted further refinement. This is Engelbart's bootstrapping made concrete: not merely using a tool to do work, but using the tool to improve the process of improving the tool. The resulting system is not optimal or finished; it is, in the manner of hypertext, permanently under construction — a quality that may be a feature rather than a limitation.

In Obsidian the collection of markdown files is known as a vault. After approximately six months of active use, my AI vault contains over 600 files: 450 atomic notes, 72 literature notes, 33 ingested resources, 20 writing drafts, 17 concept hubs, 10 Maps of Content, and 64 daily history entries. Each atomic note carries a maturity tag: roughly 39% saplings, 30% ferns, and 31% trees—a distribution suggesting active development rather than stub accumulation.

To illustrate a typical interaction, we trace the ingestion of Liu and Almeda's “Agency Among Agents” [16]. The process begins with a single command: /create-literature-note with the paper's Zotero citation key. The system retrieves metadata via Zotero MCP, delegates full-text summarisation to Gemini (which produces a 1,200-word structured summary, so that Claude never consumes the raw PDF), and populates a literature note. It then searches the vault for semantically related notes and generates two outputs: an engagement guide directing my attention toward material that challenges rather than confirms existing vault knowledge, and an AI Analysis contextualising the paper's contributions against specific vault notes.

The system then proposes candidate atomic notes. For this paper, it suggested three: Hypertextual Friction, Associative Thinking in Digital Knowledge Systems, and Agent-Driven vs. User-Driven Systems—flagging the third as the weakest candidate, a taxonomic classification rather than an insight. I accepted the first two and agreed with the system's concern about the third, redirecting it: rather than a standalone note, the distinction was better captured as an update to the existing note on Agentic AI. This kind of intervention—recognising that a proposed note is really a facet of an existing concept—is characteristic of the negotiation that occurs at this stage. The system decomposes a paper along its own argumentative structure; I reshape the extraction to fit the vault's existing conceptual landscape and my own mental models.

In total, this interaction produced one literature note, two new atomic notes, one updated note, and one history entry, establishing approximately fifteen new bidirectional links. Across the workflow, I made half a dozen curatorial decisions—selecting this paper for ingestion, accepting two proposed notes, rejecting a third, redirecting that concept to an existing note, and reviewing the resulting additions via the history entry—each a point where the workflow paused for human judgement rather than proceeding autonomously. Figure 2 shows the resulting Hypertextual Friction note (left) and the daily history entry (right).

The example above illustrates a single instance of the interaction pattern that recurs across all the vault's knowledge-capture workflows. Whether ingesting a research paper, processing a conference transcript, or extracting concepts from a recorded conversation, the fundamental rhythm is the same: the AI proposes, the human selects, and the selection process is where the intellectual work resides.

Selection is not passive approval. In practice, my responses to AI proposals fall into several categories: acceptance as offered, acceptance with reframing (the concept is valid but the proposed title or emphasis does not match the vault's existing vocabulary), merging of two or more proposals that the AI treated as distinct but that I recognise as aspects of a single idea, rejection on grounds of redundancy with existing notes, and — occasionally — identification of a concept the AI missed entirely. The distribution across these categories varies by source material, but my experience is that outright unmodified acceptance of a set of new notes is rare, and that roughly half of all notes proposed are either rejected or reshaped in order to keep the vault focused (and in line with my own way of thinking about the world).

This selection process is the primary locus of hypertextual friction in the system. It requires me to hold the proposed note in mind alongside the existing network (or to review that network as part of the process), assessing not only whether the concept is valid but whether it belongs — whether it occupies a distinct position in the knowledge structure or merely restates something already captured. This assessment draws on tacit knowledge of the vault's topology that the AI does not fully share, even with access to the same files. For example, I have a sense of what clusters are overdeveloped, which concepts are underserved, and which thematic boundaries are becoming blurred, contextual judgements that inform curation but resist formalisation.

The suggest-only pattern also addresses a structural problem unique to AI-assisted hypertext construction. Without human selection, the system could generate indefinitely; each new note suggests connections to unmade notes, which could themselves be created and linked, producing recursive growth with no intellectual filter. The human provides the point at which the current exploration has reached the boundaries of present interest and further expansion would produce diminishing returns. This curatorial role is a form of friction that preserves the vault's coherence by imposing limits on its growth. I am the exit condition.

If the suggest-only workflow represents bottom-up friction, decomposing source material into atomic units and curating the results, then Maps of Content represent top-down friction: selecting existing notes, determining their argumentative sequence, and annotating the relationships between them to produce a coherent narrative arc. The assembly is entirely human-directed; the AI may suggest relevant notes, but the ordering, argumentative logic, and interpretive framing are my own.

Once the skeleton is in place, the /expand-moc command generates full narrative prose from the structure. The AI transforms my outline — with its terse annotations, wiki-links, and structural markers — into flowing academic text that contextualises, connects, and argues across the assembled material. The friction here is sequential: I do the architectural thinking (which notes, in what order, making what argument), and the AI handles the prose generation (connective tissue, contextualisation, rhetorical flow). Neither step is trivial, but they demand different kinds of cognitive effort, and the division keeps the intellectually consequential decisions — what to argue and how to structure the argument — firmly with the human.

As an example, the MOC mentioned earlier “The Argument for Hypertextual AI Interfaces” was assembled approximately one week after the Liu and Almeda literature note described above. It draws on roughly 40 atomic notes spanning multiple source papers, AI-generated conversations, and independently created concepts, organising them into a four-part argumentative arc: AI as cognitive threat, existing approaches to human-AI knowledge work, why those approaches fail, and hypertext as a solution framework. The Hypertextual Friction note—spawned from the Liu paper—appears in the final section alongside notes from Bush, Engelbart, and my own conceptual contributions, now serving a very different argumentative purpose than it did in its source paper (Figure 3 shows an excerpt). I assembled this structure manually, determining the sequence and annotating the relationships; the /expand-moc command then generated the connecting prose. The result is an intellectual artefact that could not have been produced by either myself or the AI alone: the AI's contribution is breadth (contextualising 40 notes fluently) whilst mine is the intent and the architecture.

In a vault where the majority of content is AI-generated, the question of voice becomes unusually pointed. Whose perspective does a note represent? Whose phrasing? When I return to a note months later, which sentences capture my genuine intellectual judgement and which are competent but generic AI prose?

The #human tag addresses this question at the level of individual paragraphs. Any passage bearing this tag is protected: the AI is instructed never to modify it without explicit permission, and surrounding edits must incorporate the tagged content unchanged. In practice, I use the tag to mark personal observations, original framings, and interpretive judgements that the AI's prose does not adequately capture (for example, Figure 3 shows an example tag, making an observation about the term vibe writing). These tagged passages function as anchors of authentic voice within the generated text — points where the human's specific perspective is recorded and preserved.

The tag creates friction in both directions. For me, it requires a conscious decision about which thoughts are distinctively personal and worth protecting — a reflective act that would not occur if all content were either wholly human-written or wholly AI-generated. For the AI, it creates zones that must be navigated around: when updating a note's connections or refining its structure, the protected passages constrain what can be changed, forcing the system to integrate rather than overwrite.

A related form of friction operates at the level of language rather than authorship. The tentative voice described in Section 5 is a form of textual friction: returning to a note weeks later, provisional phrasing resists the temptation to treat it as settled knowledge, inviting re-evaluation rather than passive acceptance. A vault written in confident declaratives would close down thinking; one written in attributed perspectives keeps it open.

The daily history entries contribute a third dimension of review friction. At the close of note-creation sessions, the system generates a narrative entry that clusters the day's notes thematically and describes their relationships. I review this entry as part of the workflow, encountering new material not in isolation but in thematic context — seeing how a note on cognitive offloading relates to one on deskilling, or how a foundational concept gap connects to an existing hub.

The most striking experiential finding for me is that hypertextual friction (the moments of pause, decision, and interpretive effort described in Section 6) is not experienced as cognitive tax. It is experienced as the rewarding part of the work.

When the system proposes candidate notes from a paper, the experience of reviewing those proposals is genuinely engaging. A first suggestion often produces a moment of rediscovery: recognising how a familiar idea connects to the current research context, or a frisson of genuine novelty when encountering an unfamiliar concept surfaced by the analysis. Subsequent proposals may feel less striking, but the process of thinking through whether each deserves inclusion, reframing, or rejection remains enjoyable. The curatorial decisions are not obstacles to the real work; they are the real work, and they feel like it.

This finding challenges a persistent assumption in the friction literature. Most accounts of beneficial friction frame it as a cost that must be justified by downstream benefits: the user endures a slower, more effortful process now in exchange for better learning, safer decisions, or deeper comprehension later [11]. My experience suggests a different dynamic. The friction of co-constructing hypertext is intrinsically rewarding — closer to the engagement of a creative practice than to the tolerable discomfort of a speed bump. The decisions feel creative because they are creative: selecting, reframing, and connecting ideas within a growing knowledge structure is an act of intellectual curation that draws on tacit understanding of one's own research landscape.

Even when the AI's proposals are obviously worth accepting, the pause does not feel pointless. It maintains involvement and a sense of ownership over the evolving structure. I approve, and in approving, participate. This echoes Sennett's observation that craft knowledge resides not in the grand gestures but in the small, repeated judgements that accumulate into expertise [24]. The curatorial micro-decisions of hypertext co-construction may serve a similar function: each one is individually modest, but their accumulation constitutes a practice. This contrasts with Wu et al.’s laboratory finding that participants rated a nonlinear LLM interface lower for usability despite producing better work [28]. Their participants experienced friction as cost; I experience it as reward. The difference may lie in ownership: Wu et al.’s users were handed a tool for a single session, whereas I built mine over many months, calibrating each friction point to my own practice. If this interpretation holds, it suggests that productive friction in AI-augmented knowledge work depends not only on system design but on the user's investment in the environment that produces it—reinforcing the builder's advantage discussed below.

I also note that I can recall recent curatorial choices vividly but struggle to remember those from months earlier. This resonates with research on the AI Memory Gap [29]: when collaborating with AI systems, users’ ability to correctly attribute contributions degrades over time as source monitoring cues weaken. The vault's daily history entries take on additional significance in this light—not only as documentation of the intellectual trajectory but as a hedge against attribution drift, consulted long after the experiential memory of decisions has faded. The persistent artefact remembers what the human mind does not.

The system described in this paper was built by its primary user. Me. This is not incidental. It may be the most important feature of the design. I call this the builder's advantage: the system's value is inseparable from the act of constructing it.

When demonstrating the vault to colleagues, a recurring pattern emerges: interest and appreciation, but not the sense of transformation that I experience. Others can see that the system works, but they do not share the epiphany — the feeling that this particular configuration of tools, structures, and workflows fundamentally changes how one engages with research material. The most likely explanation is that the system embodies my own modes of thinking. The MOC workflow, for instance, maps closely onto my pre-existing approach to academic writing: browsing for relevant ideas, assembling them into an outline, annotating relationships, then developing the argument. The vault scaffolds this process rather than replacing it, which is why it feels natural, but that naturalness may be personal rather than universal.

This observation could be framed as a limitation: the system lacks transferability, and its benefits may be contingent on the builder's particular cognitive style. But there is a stronger reading, grounded in the philosophical tradition that Anderson and I  [3] identify as hypertext's deepest commitment. In our analysis, what unites hypertext's diverse manifestations — as tool for thought, knowledge representation, literature, etc. — is not a shared technology but a shared ideology: a “requirement for non-regularity” rooted in post-structuralism and enshrining existentialist values. The existentialist principle that existence precedes essence, that meaning is constructed through choice and action, maps directly onto the experience of building a knowledge system. The vault's value does not (just) reside in its architecture, which could be replicated; it resides in the process of constructing it, which cannot.

The builder's advantage, then, is not a confound in the study. It is a core finding. A knowledge system that works best for the person who built it is not a system that has failed to generalise; it is a system that has succeeded in demanding the existentialist commitment that hypertext has always, at its best, required of its users. The implication for design is not that every researcher must build their own system from scratch, but that the system must create genuine space for the user to shape it — to make consequential decisions about structure, vocabulary, and organisation that reflect their own intellectual commitments rather than accepting defaults.

This carries a risk that deserves acknowledgement. The process of building and refining the system is highly engaging, arguably too engaging. Time spent tweaking capabilities sometimes competes with time spent using the vault for its intended purpose. But the existentialist framing helps: if building the tool is a form of thinking about the problems the tool addresses, then the boundary between “building the system” and “doing the research” is less clear than a productivity framing suggests. The vault is permanently under construction, and that quality is a feature of a living knowledge system rather than a deficiency of an incomplete one.

The values co-creation process described in Section 5 illustrates how this plays out. The discovery that Reciprocal Challenge was systematically underserved changed not just the system's architecture but my behaviour, pushing back on suggestions, questioning absent concepts, splitting proposed notes. The practice of explicitly articulating values for human–AI interaction, and auditing tools against them, may itself be a form of critical AI literacy [9], one that develops through reflexive practice.

Several practical challenges emerge from this experience, offered as observations from a single sustained practice.

Distinguish knowledge sources, even imperfectly. The vault does not yet adequately distinguish between ideas drawn from peer-reviewed literature, informal conversations, and AI-generated synthesis. Tentative phrasing mitigates this by keeping claims provisional, but from an academic perspective, a pre-registered finding and a conversation observation carry different evidential weight. Systems for scholarly knowledge work should encode provenance at finer granularity than “AI-generated” versus “human-written.”

The vault works better as a thinking space than as a production pipeline. The system excels at decomposition, connection-surfacing, and argument assembly through MOCs. It is less fluent at converting vault knowledge into publication-ready text with proper citations rather than wiki-links. Future systems should consider the full arc from ingestion through to output.

Embrace the friction; the system does not work without it. It would be easy to approve generated notes without thought, or to avoid reading them once created. The system introduces deliberate friction, but the user must engage genuinely; beneficial friction is not automatic, it requires active participation at precisely the moments where the system could proceed without the user.

Where chat-based AI tools offer powerful but ephemeral synthesis (the connections they surface evaporate when the conversation ends) the vault persists. I do not need the AI to explore the hypertext; I can open Obsidian directly, browse the network, follow links, and rediscover connections that had been forgotten. The AI is the co-constructor, but the hypertext is the enduring artefact, and direct navigation far outnumbers conversational queries. The value of the system compounds over time: ideas added months ago become relevant to current work in ways that could not have been anticipated, and sometimes the connections, not the new content itself, are what produce insight. The real value emerges after sustained use. As the network densifies, the hypertext begins to function as an externalised associative memory, not a record of what has been read but a lens through which new material is understood in the context of everything that has come before. This is Bush's associative trail made concrete and collaborative: accumulated knowledge that reflects my sustained intellectual commitments, laid down decision by decision over months of practice.

For the hypertext community, this may be the central contribution: hypertext's fundamental role—scaffolding structured thinking, making connections navigable, creating persistent artefacts—becomes uniquely valuable in the age of AI. The specific system matters less than the principle it illustrates: that productive AI collaboration requires structure the user has helped to build.