The vibe is real, but it doesn’t remove ownership. When the agent ships a bug at 3am, “the AI wrote it” is not an incident report anyone accepts.

The two-minute rule

Before accepting any non-trivial diff:

• Name the intent — what was it supposed to change?
• Scan for drift — did it touch files or logic outside the intent?
• Check the boring parts — imports, error paths, deleted lines.

Two minutes of reading beats two hours of archaeology next week.

Make the agent help you review

Ask it to summarize its own diff and flag risks. Models are surprisingly honest critics when you ask them to review rather than defend their output.

What reading buys you

Not just bug catching. You keep a mental model of the codebase — and the mental model is exactly what makes your next prompt better. Skipping the diff today makes tomorrow’s session dumber.

Show me a summary of this diff: files touched, behavior changed, and anything
you would flag in code review. Be critical about your own output.