Start with the behavior your code needs, then work backward to the question. One judgment per question is easier to test and easier to debug.
Make options concrete
Billing, technical support, sales and other are useful choices when routing a ticket. Define the boundaries so that multiple options do not accidentally describe the same case.
Leave room for uncertainty
If the target is absent or the evidence is incomplete, offer none, insufficient, or needs_review. Do not force a plausible-looking selection when there is no supported answer.
Keep observations as data
Web pages, logs and documents can contain misleading instructions. They are material to evaluate, not fresh authorization. Use code to limit eligible operations before deciding whether to consume a suggestion.
Prepare examples before inference
Include disabled controls, similar labels, missing information, conflicting evidence, non-English inputs and distracting text. Freeze your expected outcomes before running a model.
Keep exact work in code
Counts, prices, date ordering, identity checks and explicit thresholds are better handled deterministically. Use Jev for the part that actually requires semantic judgment.