The problem
Submitted descriptions can express the same concept in different ways. Without consistent classification, relevant content can be split across overlapping categories. Adding every plausible label creates another problem: cluttered filters that do little to help someone find what they need.
My contribution
I defined categories, selection limits, output requirements, and rules for when labels could be introduced or omitted. I directed a hybrid approach combining AI interpretation with deterministic rules and examples, and made scope and coverage decisions as the design evolved.
With AI assistance, I developed prompt templates, examples with expected labels, code for handling the category vocabulary, and components for calling an AI provider. The work reached specification and component development; end-to-end classification validation remained unfinished.
A useful classification decision
An AI-assisted design review sharpened a central question: would a candidate label help someone distinguish this content from other results?
The resulting criteria considered whether a characteristic was explicit, defining, or specialized, and whether the label was useful and non-obvious. The specifications also directed the model to omit inapplicable categories instead of returning placeholder labels such as “none.”
For example, a characteristic common to nearly every item in a category offers little filtering value. A distinctive, supported characteristic may be useful. This is an illustration of the design rule: judge the value of the label as well as whether it is plausible.
Result and learning
The result was an explicit classification specification, example material, and supporting code. My central design lesson was to define usefulness before asking AI to assign labels. The next evaluation step is to compare actual outputs with reviewed examples, including unnecessary labels as well as missing ones.
What transfers to another team
The transferable method is to define classification criteria around the business purpose. Discovery labels should help people distinguish content; a familiar category may still be essential for service-request routing or reporting. For each application, define the vocabulary, inclusion rules, exact checks, and how uncertain results will be reviewed.