AI can help you move faster. It can also help you ship the wrong thing with remarkable confidence. A better prompt will not fix that. You need an engineering workflow around the tool.

This talk follows a realistic piece of work from a vague request to a reviewed change. We use AI to plan, write, refactor, test, and document the work. We also stop when the evidence does not support its answer. The examples include invented APIs, stale assumptions, overloaded context, and tests that merely agree with the generated code.

We turn those mistakes into a workflow a team can use: work in small steps, provide only useful context, verify important claims independently, and keep clear human checkpoints. If you already use AI and want something more dependable than a fast demo, this review loop lets the agent move quickly without handing over your judgment.