Our summary
This is a short post on X from Nathan Flurry, founder of Rivet, published on 18 November 2026. He says his team first asked Jev one broad question: give the severity of an issue, with a list of what each level means. He reports that this gave mediocre results, and that it treated Jev like a model that reasons its way to a conclusion.
After reading the docs, the team split the job into tiny questions: did the service go down, is the customer blocked with no workaround, is something broken. Ordinary code then picks the severity from the answers. He says this works well, and that when a ticket is weighted badly they have a language model add it to a test suite, adjust the logic and re-check old issues.
It is one team's account in a social post, with a screenshot JevMade could not read and no shared code, so the results cannot be checked. The post lists how many Jev questions their internal support classifiers ask, for example 13 for support classification and 6 for severity. Treat those counts as one example, not a target.
Key takeaways
- Ask Jev small, specific questions and let your own code reach the conclusion, because Jev is not built to reason to one.
- Turn a badly weighted ticket into a test case, adjust the decision logic and re-check old tickets to improve it.
- Expect to ask many questions per classifier; Rivet reports 13 for support classification and 6 for severity.
