Concepts
The irreducible epistemic atoms underlying the curriculum. 4,828 atoms across 8 types
Create the note ID first, before writing any content…
Assign a unique identifier to every note before writing any content, treating the addressing decision as the first step that enables all subsequent linking and referencing.
Store evidence as independent nodes with methodology…
Store evidence with full methodological metadata (sample size, control conditions, limitations) as independent nodes rather than as decorative citations on claims, to enable proportionality assessment and multi-argument reuse.
Fine-grained notes for frequent precise lookups, coarse…
Match note granularity to retrieval frequency and question complexity: create fine-grained atomic notes (single claims) for domains where you need precise retrieval, and coarser aggregated notes for domains where you need high-level orientation.
Maintain multiple granularity levels for AI knowledge…
When using AI systems with your knowledge base, maintain multiple granularity levels of the same material (fine-grained for precise retrieval, coarse-grained for contextual reasoning) rather than forcing a single chunk size, because different query types require different resolutions.
Give questions the same structural treatment as answers…
Store well-formed questions as first-class atoms in your knowledge system with the same structural treatment (unique identifiers, bidirectional links, metadata) as claims and answers, because questions organize attention and generate persistent search filters.
Treat every note link as infrastructure for AI graph…
When building knowledge systems that will interface with AI, treat every link you create as infrastructure that future graph traversal algorithms will follow, prioritizing explicit relationship encoding over implicit semantic similarity because GraphRAG systems require edges to perform multi-hop reasoning.
Keep raw inbox captures out of AI retrieval scope
Index only processed permanent notes in AI-searchable systems while keeping unprocessed inbox captures outside retrieval scope, because AI systems cannot distinguish epistemic status and will retrieve raw captures with equal confidence to verified knowledge.
Quarterly
Schedule quarterly depreciation reviews where you scan captured notes and bookmarks for information that has exceeded its useful life, then either update with current data, archive with context, or delete entirely to prevent outdated information from corrupting current decisions.
For each category, complete
Document the purpose each category serves by completing the sentence 'this category exists to [do what] for [whom]' to distinguish functional infrastructure from inherited furniture.
When a binary label hides multiple distinct reasons, split…
When a binary classification hides multiple distinct failure modes or reasons within a single bucket, decompose it into separate dimensions that can be evaluated independently.
Test every classification for MECE
For every pair of categories in a classification system, verify that no item can legitimately belong to both (mutual exclusivity test), and verify that no domain item falls outside all categories (collective exhaustiveness test).
Match classification architecture to domain
Design multi-class classification systems with mutually exclusive categories when items can only be one type, and multi-label systems when items can legitimately belong to multiple categories simultaneously.
Classification friction reveals missing categories — map…
When items consistently resist classification in your system (you hesitate, force-fit, or leave uncategorized), map what those resistant items have in common to diagnose missing categories that represent dimensions you care about but haven't encoded.
A growing 'Miscellaneous' category signals missing…
When a 'Miscellaneous' or 'Other' category grows faster than named categories, it signals that your classification dimensions are missing a meaningful distinction that reality contains.
Design hierarchy levels by what question each answers…
Before creating or reorganizing any hierarchical structure, ask what question each level answers—top level for domain identification, leaf level for specific item selection, and intermediate levels for navigation steps between.
Test each hierarchy level by removal — if promoting…
For each intermediate level in a hierarchy, test whether removing it and promoting its children one level up would lose meaningful organization—if not, flatten it, because unnecessary levels are pure navigational tax.
Structural depth encodes real boundaries; bureaucratic…
Distinguish structural hierarchy depth (encoding real containment or inheritance) from bureaucratic depth (added for perceived tidiness) by asking what would break if you removed each level—keep only levels whose removal would destroy actual functional boundaries.
Clustered overrides mean the parent's assumption is wrong…
When multiple children override the same inherited property, restructure the hierarchy rather than accumulating individual overrides, as clustered overrides indicate the parent's assumption is systematically wrong.
Override traits, not identity — if you are overriding…
Override properties rather than identities—preserve the child's relationship to its parent category while modifying specific inherited traits, as overriding identity signals you need a different category, not an override.
Restructure at the friction point, in the moment…
Restructure hierarchies at the specific node causing friction during the moment you feel the friction, rather than conducting proactive system-wide reorganizations, to keep restructuring costs small and diagnostic signals fresh.
Before adding depth, try flattening + tags first — deep…
Before adding another level of nesting, first attempt to flatten the hierarchy one level and use tags or links to preserve relationships, as deep hierarchies are more expensive to maintain than flat hierarchies with rich cross-references.
Three levels deep is the maximum — any level…
When navigation to any item requires remembering a path more than three levels deep, audit whether each nesting level provides unique decision-making value—if you cannot explain what decision a level enables, eliminate that level as noise.
Use four organizational mechanisms together
Combine hierarchical folders (for coarse structure), tags (for cross-cutting themes), explicit links (for semantic relationships), and maps of content (for curated entry points) rather than relying on any single organizational mechanism, as each hierarchy type makes different questions answerable.
Structure long documents in three disclosure layers
When documenting information longer than one page, structure it in three disclosure layers: single-sentence summary (Layer 1), paragraph-per-section abstracts (Layer 2), and full detail (Layer 3), with each layer independently meaningful.