Our summary
A coding assistant makes choices as well as writing code. Steve Kinney explains how Jev, an AI model that judges supplied information, can help assess a proposed command or choose useful instructions for a task. Start with one frequent, reversible decision rather than handing over every step at once.
First record what Jev would decide while the existing system stays in charge. Compare answers with examples whose correct answers are known. Decide which answers are strong enough for each action and what happens when requests fail. Ask independent questions together when they use the same evidence, and keep that evidence focused.
A correctly shaped answer can still be wrong, and hostile text can influence the model. Uncertain or failed Jev answers should leave existing decisions unchanged. Firm permission checks must still block forbidden actions: Jev may raise concern but never lower a safety requirement. This lesson offers a rollout method, not a tested implementation.
Key takeaways
- Choose one reversible decision and watch its answers before letting them affect the work. Check mistakes against known answers and keep checking after the change.
- Give each option a distinct meaning and offer an other answer. Remove secrets and irrelevant material before sending evidence to the service.
- Choose how strong an answer must be based on what a wrong action would cost. Measure cost and errors per completed task, not just per request.