A taxonomy is not a description of a domain. It is a tool for a specific job — usually helping someone find a thing whose name they do not know.
I start from the question of which queries need to work, not from how the domain is laid out in a textbook. Everything else — hierarchy depth, number of facets, granularity — follows from the answer.
Method
Facet analysis in the tradition of Ranganathan and the Classification Research Group: breaking the domain into independent dimensions rather than one tree in which every object must have a single home. Vocabulary control per ISO 25964. Validation against real queries and boundary probes, not intuition.
Concrete examples: why a single tree breaks
The most common failure mode in taxonomy design is gluing several independent questions into one hierarchy:
- In the competency atlas (
REF.01), ontological relations (what a skill is and requires) sat in the same layer as market mappings (how often it gets mentioned). Splitting those two layers removed 474 spurious edges — the graph got smaller and for the first time became usable for inference. - In the psychological games taxonomy (
REF.04), earlier attempts grouped items by branch of psychology, leaving borderline items with two honest homes. Separating category (what the object is: cognitive task, interpersonal game, manipulation tactic) from mechanism (why it works) turned an overlapping tree into an eight-rule yes/no cascade and raised blind-test reproducibility from 20/25 to 23/25. - In the web project taxonomy (
REF.05), payments, login and AI features stopped minting separate “site types” and became binary modifiers alongside a four-gate cascade of functional outcomes (N > P > O > R).
Boundary rules instead of bare definitions
A vocabulary consisting only of labels and dictionary definitions breaks on the first borderline item. Every category therefore receives an explicit scope note and a yes/no boundary rule: what qualifies an item for this leaf, what disqualifies it, and which neighbouring leaf takes over. Before the structure ships to production, the rules are stress-tested on real corpus items with names and worked examples hidden.