Zettelkasten for Academics: A Practical Version

The Zettelkasten that survives contact with academic research is a three-layer pipeline: literature notes from your reading, permanent notes carrying one claim each in your own words, and maps of content that assemble those claims into paper outlines. Everything else in the method — the numbering schemes, the atomicity litigation, the note-type taxonomies with six categories — is either a historical artifact of paper filing or a hobby, and treating it as load-bearing is why most academic Zettelkasten attempts are dead by month three.

This article is the honest version: what Niklas Luhmann actually did versus what the productivity internet says he did, the two failure modes that kill academic implementations, the stripped-down pipeline that works, and a worked example running one real paper all the way through to where its notes would land in a chapter. If you haven’t yet got literature notes working — the input to everything below — start with our guide to literature notes first; a Zettelkasten built on highlights is a cathedral built on sand.

What Luhmann actually did

The facts, since they’re routinely embellished. Luhmann kept two slip boxes over roughly forty years: a first collection begun in the early 1950s while he was a legal civil servant, and a second, started around 1963, that accompanied his entire sociology career at Bielefeld. Together they hold about 90,000 handwritten A6 slips — the Bielefeld archive has been digitizing and transcribing them since 2015, and you can browse the cards yourself at the Niklas Luhmann Archive. Against that input he produced an output that justifies the legend: around fifty books and several hundred articles, including a completed grand theory of society.

Three details matter for what you should copy, and they cut against the internet version.

First, the arithmetic is unglamorous. Ninety thousand notes over forty-plus years works out to roughly six notes a day. The system’s power was compounding and consistency, not volume. Luhmann was not “capturing everything”; he was writing a handful of considered notes daily, forever.

Second, his cards were not atomic in the way the modern movement means it. Many slips are dense abbreviated prose; plenty of thoughts run across sequences of cards; some cards are basically small essays in miniature handwriting. The “one idea per note, ruthlessly” rule is a modern reconstruction — a useful one, as we’ll see, but Luhmann was not enforcing a formalism. He was writing to think.

Third, his index was sparse on purpose. The keyword register pointed to only one or two entry cards per topic, because the expectation was that you’d enter the box at one point and navigate by the links and branches from there. The intelligence lived in the connections between notes, not in a taxonomy above them. That’s the actual insight of the method, and it’s the one thing every layer of the modern productization has tended to bury under templates.

What the productivity internet did — largely downstream of Sönke Ahrens’ How to Take Smart Notes (2017), which is a good book that spawned bad imitations — was extract a procedure from a practice. Fleeting notes, literature notes, permanent notes, hub notes, structure notes, index notes; a folder for each; a checklist for promoting a note from one caste to the next. Luhmann had a box, a pen, and a reason to write: he was building a theory. The reason is not optional equipment.

Why academic Zettelkasten attempts die

Watch academics adopt this method — I’ve watched many, and been one — and the failures cluster into two patterns.

Over-proceduralism. The system becomes the project. You spend the first fortnight designing templates, choosing between five note-type taxonomies, arguing with yourself about whether a note is “literature” or “permanent” or “fleeting-pending-promotion,” and installing plugins to enforce the distinctions. Every processing session begins with a small bureaucratic tribunal about categorization before any thinking happens. The tell is easy to spot: your vault has more notes about your note system than notes about your research. The method was supposed to reduce the friction between reading and writing; a sufficiently elaborate implementation increases it, and increased friction plus a grant deadline equals abandonment.

Atomicity fundamentalism. “One idea per note” is a genuinely useful discipline — it’s what makes notes recombinable across papers. Taken as dogma, it becomes a paralysis engine. You finish reading a paper, sit down to write permanent notes, and stall on ontology: is “prospect theory’s value function is steeper for losses” one idea or three? Should the reference-point claim be split from the loss-aversion claim? You split, then the fragments are too thin to mean anything alone, so you write context into each one, and now they overlap and you’re deduplicating your own thoughts. Meanwhile the actual question — what does this paper change about my argument? — sits unanswered. The workable rule is looser: one claim per note, where a claim is something you could imagine defending or attacking in a paragraph of a paper. If it would be one paragraph in your chapter, it’s one note. Calibrate to your writing, not to a metaphysics of ideas.

Both failure modes share a root: they optimize the system against an aesthetic standard instead of against the only metric that matters for an academic, which is whether the notes make drafts arrive faster.

The pared-down academic implementation

Three layers, three jobs. This slots directly into the full researcher note-taking system if you want the surrounding capture pipeline; here’s the Zettelkasten core.

Layer 1: literature notes. One note per source, in your own words: claim, evidence with numbers, method, limitation, connections. This layer is fully specified in the literature notes guide, so I won’t repeat it — the only thing to add is its role here: the literature note is about the source, bound to the source’s perspective and citekey. It’s the quarry, not the building.

Layer 2: permanent notes. One claim per note, stated in your own words as a proposition, with the supporting sources cited and — non-negotiably — at least one link to another permanent note. The shift from layer 1 to layer 2 is a shift of ownership: a literature note records what Kahneman and Tversky claim; a permanent note records what you now hold to be true, or worth arguing, having read them. Title the note as the claim itself (“Losses loom larger than equivalent gains”) so that linking to it mid-sentence reads naturally in other notes. Write three to ten sentences: the claim, why you believe it, what would change your mind, who says otherwise. Most papers you read generate zero permanent notes, and that’s correct — a permanent note is a decision that an idea has recurring value to your research program, not a tax levied on every PDF.

Layer 3: maps of content (MOCs). A map of content is a plain note that lists and arranges links to permanent notes on a theme, with connective prose between them. This is where the method stops being note-keeping and starts being pre-writing, because an MOC that orders claims into a sequence — with your sentences explaining why claim two follows claim one — is a paper outline that wrote itself incrementally over months. When I say the Zettelkasten is a paper-outline generator, this is the mechanism: you never face a blank outline, because the outline accreted every time you filed a permanent note into a map and asked “where does this sit relative to what’s already here?” A dissertation chapter is an MOC that got long enough to defend.

The daily practice, concretely: read and annotate; process into a literature note within 48 hours; while the paper is fresh, ask “does anything here deserve to become mine?” — write zero to three permanent notes; link each new permanent note into at least one existing note and one MOC. Twenty minutes past the literature note, most days. That’s the entire method.

A worked chain: Vaswani et al. (2017)

Here’s one highlight from “Attention Is All You Need” traveling the whole pipeline — the same paper annotated in our Zotero-Obsidian workflow guide, where you can see the upstream half of this chain in screenshots.

The highlight. Yellow, page 2: the observation that recurrent models compute hidden states sequentially — each position waits for the previous one — which “precludes parallelization within training examples.” A marginal comment typed at reading time: “the whole paper exists to delete this constraint.”

The literature note (vaswani2017, excerpted). Claim: attention is sufficient — a sequence architecture with no recurrence or convolution, using self-attention alone, matches and beats recurrent models on translation while training in a fraction of the time. Evidence: 28.4 BLEU on WMT14 English-German, new state of the art at publication, at a training cost of days on eight GPUs rather than weeks. Limitation: self-attention is quadratic in sequence length; the paper’s wins are on sentence-length sequences. Connection: the motivation section is really about hardware, not language — flag for my methods-history section.

Permanent note 1: “Recurrence in sequence models was a serialization bottleneck, not a modeling necessity.” The claim in my words: RNNs conflated two things — modeling order and computing in order — and the Transformer showed the second was never required for the first. Self-attention connects any two positions in a constant number of steps and computes them all at once. Cites vaswani2017; links to my existing note on inductive biases, and gets attacked by a later note on quadratic scaling.

Permanent note 2: “Architectural winners are selected by hardware fit as much as by expressiveness.” A generalization the paper licenses but doesn’t state: the Transformer won substantially because it saturates parallel accelerators, and the field’s architecture history reads differently once you track hardware fit as a selection pressure. Cites vaswani2017 alongside two other sources already in the box. This note has nothing to do with translation — which is the point. It’s detached from its source’s context and available to any argument that needs it.

Where they land. In a dissertation with a chapter on why particular methods came to dominate, note 1 files into the MOC section on the recurrent-to-attention transition — it will become a paragraph of related work. Note 2 files into a different map entirely, “compute as an epistemic constraint,” where it sits with notes derived from entirely different literatures and will anchor a discussion-section argument. One reading session, two claims, two different chapters served. Multiply by a few hundred papers over a PhD and you see what the machine is for.

The folgezettel question, dispatched

Luhmann’s cards carried branching alphanumeric addresses — 21/3d7 sits behind 21/3d, which branched off 21/3 — and a corner of the community holds that this Folgezettel positioning, each note filed behind the note it continues, is the method’s secret engine. In Obsidian the debate is moot: the numbering was a physical addressing system for a cabinet with no search, no backlinks, and no graph — the only way a paper slip could “link” was by naming another card’s address, and the only way it could have a neighborhood was by physically standing in one. Links, backlinks, and unlinked-mention search do everything the numbers did, with less ceremony and no renumbering anxiety, and if you want the “what does this continue?” prompt that folgezettel enforced, a follows:: link at the top of a note buys it for free. People who enjoy the numbering should enjoy it; nobody’s stalled dissertation was ever unstalled by an addressing scheme. (Vault structure and the plugins worth having are covered in the Obsidian for academic research guide.)

When not to bother

The Zettelkasten’s payoff function is back-loaded: it compounds across projects and years, which means it’s a bad fit for anything short and self-contained. Skip it if you’re writing a one-off paper in a field you’re passing through — literature notes alone will carry you, and the permanent-note layer would be effort invested in connections you’ll never revisit. Skip it during coursework; seminar reading serves the seminar, and forced “permanent” notes on material outside your research program are compost, not capital. Skip the permanent-note layer if your work is primarily experimental and your bottleneck is the lab, not the literature — a good literature-note practice plus a project log may be your whole system, honestly held. And skip it entirely for the next four months if a deadline is that close: mid-crisis is the worst possible moment to adopt infrastructure, and the method’s first weeks are all cost.

Adopt it when the shape of your problem matches the shape of the tool: a multi-year research program, arguments that will be prosecuted across several papers, a literature too big to hold in your head. That’s most PhDs and every research career — which is presumably why a Bielefeld sociologist with a filing cabinet needed it, and why the stripped-down version above, minus the cabinet, still earns its twenty minutes a day.