Our summary
Ted wants to spot useful blog comments and emails that need attention. He tests Jev, an AI tool that answers questions about supplied text rather than writing replies. His trial uses 15 comments, 50 emails and a separate security question on 12 account-related emails.
The first questions miss a pasted bot instruction and account alerts. Ted rewrites them to describe visible events, such as a new sign-in, rather than broad categories. Reported answers improve on those same examples. That shows the effect of wording, not accuracy on fresh messages.
His scripts keep uncertain messages visible. If the service fails, messages go through without tags and ordinary security checks still flag suspicious subjects. The comment watcher suppresses only bot answers at 85% confidence or higher, saving them in a log. Neither wording nor a confidence number guarantees safety.
Key takeaways
- Describe events a reader can find in the message, rather than asking whether it fits a vague category.
- Treat improved answers on examples used to tune the questions as a small personal trial, not proof that new messages will be sorted correctly.
- Keep ordinary alerts working if the AI call fails. Show uncertain messages and retain a checkable log of suppressed bot comments.