Research craft
The deliberate management of knowledge work. Shelf: hamming1986 (career as a control system), alon2009 (problem choice as a procedure), keshav2007 (reading as budgeted passes), luhmann1981 (the note network as a communication partner), and karpathy2026 (knowledge as a compiled, maintained artifact), olah2017 (distillation as communal infrastructure), and matuschak2019 (making as a route to new insight), griffiths2024 (note method separated from publishing posture), with favaro2026 as frontier-lab evidence that execution cheapens before problem choice and judgment.
Craft is scheduled, not trait-based
The shared move across the procedural works: turn a virtue into a mechanism with a slot on the calendar. Hamming’s problem selection is not taste but ritual — Friday “Great Thoughts Time,” 10% of the week reserved for what the field is becoming, a warm queue of 10–20 attackable important problems. Keshav’s reading skill is not diligence but a pass structure with explicit exit decisions after each. Karpathy’s knowledge base is not discipline but standing operations (ingest, query, lint) encoded in a schema document. alon2009 supplies the selection procedure Hamming’s ritual gestures at: rank candidates on feasibility × interest, keep only the Pareto front, weight by career stage — and enforce a three-month no-commitment rule, because projects take years and weeks spent choosing repay months. Favaro and Clark make that allocation pressure concrete inside an AI lab: when code and fixed-goal experiments become cheap in human time, deciding which experiments matter, which results to trust, and when to abandon a direction becomes the remaining bottleneck. Their selected next-step probe is not a measure of research taste—it begins from 129 moments where the human had room to improve and uses a model judge—but it shows why problem-selection procedure matters more, not less, as execution accelerates. In each case the practitioner who “just has the knack” is replaced by a mechanism an ordinary week can execute.
The common enemy: attention spent without accumulation
Hamming’s engine is compounding — a steady extra hour of applied thought compounds into a qualitative career difference, while misapplied drive compounds nothing. Keshav’s target is the same waste at reading scale: front-to-back plowing spends depth where it earns nothing, so depth is allocated only after cheap passes justify it. Karpathy’s target is the waste at system scale: RAG-style retrieval re-derives understanding on every question, so nothing accumulates — the wiki exists to make every source read and every question asked add up. Three scales, one principle: structure work so its outputs compound, and treat un-accumulated effort as the default failure mode.
olah2017 moves that principle from the researcher’s own attention to a field’s shared infrastructure. Poor exposition, undigested ideas, bad formalisms, and noise are missing interpretive labor: one author saves effort once and every later reader repays it. Distillation reverses the multiplier by doing difficult conceptual and design work upstream. It is not the polish after research; finding the notation, visualization, or abstraction can change the idea itself. Keshav’s pass structure limits how much debt a reader services, while a maintained library makes the recovered understanding compound. Neither substitutes for better public sources.
Making can be an instrument of discovery
matuschak2019 adds a second feedback loop. Transformative media do not follow from applying known cognitive science to an interface; building the medium must expose new facts about memory or the target domain, and those insights must drive the next build. Their mnemonic-medium prototype is early and its retention data do not establish learning transfer, but the insight-through-making criterion is a useful test for research tooling: what did the artifact make newly observable, and how did that observation change the artifact?
The essay also explains why this craft is underprovided. Interface ideas can be expensive to discover, cheap to copy, and more valuable when widely shared, so ordinary product incentives may reward distribution and lock-in over deep tool research. Read with Research Debt, two forms of upstream work—distilling ideas and inventing their media—create diffuse future value while their cost is paid locally. A durable research practice needs institutions that can fund that mismatch, not only better individual habits.
Reframing and the write-side inversion
Hamming’s repeated career move was problem transformation — never solve an isolated problem except as the characteristic of a class; generalizing often simplifies. Alon gives the mid-project version: real research meanders through “the cloud” (assumptions down, morale down) and often surfaces a problem C better on both axes than the one you set out for — holding a schema that licenses the switch is itself craft. Selling is half the work: work unread is work undone. Keshav inverts his own reading method into writing advice: readers give one pass by default, so a paper must survive pass 1 (headings, abstract) to earn a second. Karpathy’s writeback (answers filed into the wiki as new pages) is the same inversion internally — outputs shaped so the future reader, including future-you and your agent, can build on them. The craft is symmetric: read with exit decisions, write for the reader’s first pass.
Externalization closes Hamming’s gap
Hamming manages accumulation in the researcher — subconscious loading, ambiguity tolerance (log every misfit, as Darwin did), fault inventories. luhmann1981 is the historical proof the externalized version works: twenty-six years of a note network run as a communication partner, with fixed addresses, explicit links and backlinks, no privileged notes, and a critical-mass threshold below which the box is “a mere container” and above which queries return more than was ever stored. Luhmann also names what the artifact is for — surprise; a system that only returns what you filed is not communicating. Karpathy completes the lineage — the wiki holds the cross-references and flagged contradictions so they survive between sessions; the LLM absorbs the maintenance burden Luhmann paid in person and that made everyone else’s wikis die. griffiths2024 prevents that lineage from collapsing storage method into publishing form. Evergreen content and a remixable note format are different achievements; single- versus multi-author and private versus public are independent axes. Forester can therefore operate as a Zettelkasten backstage and as a wiki in public without either role defining the other. Keshav’s pass 3 supplies the quality bar for what the external store should contain: understanding good enough to reconstruct the work from memory — his virtual re-implementation diff is what separates synthesis from summary.
Local instantiation
This library runs these mechanisms: pass-1’s five Cs are the stated bar for a notes file’s frontmatter-plus-opening (keshav2007 is cited in the ingest skills’ quality bar); transformation-not- compression is the pass-3 standard; the shadow/notes/wiki tiers are karpathy2026‘s raw/wiki layers; and /evolve is Great Thoughts Time made structural — scheduled meta-work on how the work itself is done. olah2017 supplies the criterion for whether the result is worth preserving: synthesis must lower future interpretive cost, not merely add another summary. matuschak2019 supplies the stronger design question: can changing the representation reveal relations the old medium concealed? Open questions for future ingestions: Hamming’s exposure (open doors, the important-problems lunch) has no personal-system equivalent yet — what does ambient awareness look like for a solo practitioner with agents? None of the works addresses when to stop maintaining an artifact — the retirement half of compounding. And Luhmann’s surprise criterion is an open test for this repo’s own wiki layer: does querying it ever return relations that weren’t deliberately filed?