Our summary
A short lookup and a difficult investigation need different amounts of help. Krzysztof Słomka uses Jev to sort Claude Code prompts into four task sizes. The hook adds an instruction asking the lead model to delegate; it does not switch models or enforce that delegation itself.
Short messages and uncertain classifications stay with the lead model. Missing credentials, timeouts and unusable answers also leave the original prompt untouched. When Słomka checks the installed hook, he discovers it has never routed a real prompt: it read prompt_text instead of the actual prompt field.
The article measures six synthetic endpoint calls, not a working production router. Its confidence cutoff is uncalibrated, and helpers may lack the conversation behind a follow-up. Routed text goes to OpenRouter and TypeSafe. A person enables each project explicitly; the author keeps confidential client work outside this route.
Key takeaways
- Verify that the hook receives the field it actually reads. An empty log can mean it exits before calling the classifier, even when prompts still work normally.
- Keep context-dependent follow-ups with the lead model. A task-size label does not give a helper the conversation or files it needs.
- Make external prompt sharing an explicit project choice. The classifier suggests a tier; it must not decide whether confidential text may leave the machine.