JEV / LESSON 07 · 10 min
Nine common Jev beginner mistakes and how to diagnose them
Troubleshoot question design, answer reading, keys, fallback, and pre-launch validation.
Mistakes 1–3: task design
1. Asking Jev for long prose; use a text model for that. 2. Combining route, urgency, and refund authority in one question; split the judgments and compose in code. 3. Overlapping Choice options with no other; clarify boundaries and add review.
Mistakes 4–6: input and probabilities
4. Missing verified facts in state; fetch them first. 5. Calling a Noul value confidence; Noul exposes noul, not confidence. 6. Treating confidence as accuracy or permission; evaluate on labels and enforce permissions in code.
Mistakes 7–9: integration
7. Putting an API key in browser code; keep it server side. 8. Handling only success; cover timeouts, rate limits, missing fields, and fallback. 9. Launching from demo examples alone; test real distributions, edge cases, languages, and model updates.
A diagnosis order
Check HTTP status and request ID, then state/model/questions, then answer IDs and fields, then your thresholds and actions. For poor decisions, inspect labels and question scope before rewriting instructions.
Practice exit criterion
Start with at least 20 de-identified, human-labeled cases from the target task and record outputs, error types, and final action. Twenty is only a learning exercise, not a statistically sufficient production sample.
Practice
Practice
A ticket reports a missing parcel and duplicate charge. What design is riskiest?
What does the support team actually need to decide?
Follow one support ticket from raw text and verified facts through three questions, an illustrative answer, edge cases, and launch evaluation.
Worked example: design a Jev workflow for a missing deliveryProgress is stored only in this browser