How 24,000 notes become a curriculum
Vidhyarthi — Sanskrit for "student" — is my research system, and after a couple of years of daily tending it holds roughly 24,000 atomic concept notes on LLMs and agentic AI, organised into 427 taxonomy clusters, each fronted by a hub note. The claim I want to make from inside it: a Zettelkasten at this scale doesn't just store what I've learned — it makes curriculum design a query instead of a blank page.
The pipeline is one line: research → wiki → books, products, training. Sources go in at one end — dozens of deep-research whitepapers, YouTube transcripts, technical web series — and each gets decomposed into atomic notes: one concept per file, preserved at full depth. The system's founding rule is never summarise. Summaries are where implementation details go to die; the goal is to make knowledge navigable, not shorter. Three structural commitments hold it together. Atomicity: every concept stands alone, so "Raft leader election" and "approval escalation chain" are separate, individually linkable notes. An explicit taxonomy: a single JSON file is the source of truth for every concept ID and its cluster assignment — membership is data, not vibes. And hubs: each cluster gets a navigation note that names its concepts, states why they belong together, and — critically — prescribes a reading order.
Curriculum as query
Those three commitments are exactly the ingredients of a curriculum, which is why the conversion is mechanical. When I built training modules for engineers learning agentic AI, the procedure was literally: select the clusters in scope, order them by dependency using the cross-links between hubs, then sequence concepts within each cluster by progressive complexity — which the hubs' reading orders already encode. A week-by-week study plan falls out as a path through the graph. Teaching agent supervision? The supervision-levels hub hands me its seven concepts in teaching order, and its cross-links surface the prerequisites — trust calibration, cognitive load — that a blank-page outline would have forgotten. The books run the same query with a different traversal and more connective prose; the wiki is the single source both draw from. And because notes are atomic, one concept can appear in five curricula without being rewritten for any of them — the note is the unit of reuse.
The part that surprised me: the structure improves the research, not just the teaching. Writing new notes against an explicit taxonomy forces the question "is this actually new, or a duplicate of something in cluster X?" — which is a comprehension test disguised as filing. Accumulation without that test is how you re-read the same idea for the fourth time and call it learning.
What the graph doesn't give you
Two honest limits. First, maintenance is real work, not overhead-rounding-to-zero: at this scale I've run link-repair passes, taxonomy audits, and full cluster rebalances — splitting clusters that had bloated past 200 concepts back into teachable units. A knowledge graph rots exactly as fast as you stop gardening it. Second, dependency order is not always learning order. Good teaching sometimes needs the strategic simplification — the temporarily-wrong model you refine later — and atomic notes, faithful to full detail by design, resist that move. The graph gives me structure and coverage; pacing, motivation, and the pedagogically useful lie still have to come from me. A query can draft a syllabus; it can't decide what a Tuesday afternoon cohort can absorb.
But I'll take that division of labour. Before this system, every curriculum started at zero and quietly reflected whatever I happened to remember that month. Now accumulation compounds: every whitepaper processed makes the next course cheaper to build and harder to get wrong. Accumulation without structure is a pile; structure without accumulation is an empty syllabus. The Zettelkasten is the contract between the two.