Reading and steering AI output

You do not need to understand every line the agent writes, but you should always test what it built like a user would: click through the flow, try wrong inputs, resize the window. Behavior is your review surface. When something is off, describe the symptom precisely instead of guessing at the cause. "The save button does nothing after I edit the second field" gets a better fix than "the form is broken". Paste exact error messages verbatim — they are the agent's best debugging input. Steer with small corrections, one thing at a time. Five focused follow-ups beat one giant do-over prompt, because each fix stays reviewable and reversible.