AI coding tools perform much better in controlled demos than in complex Unity projects due to a lack of complete project context. Successful AI integration requires providing explicit project information, including Unity versions, package details, architectural boundaries, and platform requirements. Context engineering involves organizing a project to make invisible dependencies visible to the AI, ensuring useful and reviewable changes. Small, focused tasks with clear acceptance criteria are more effective than broad feature requests. AI should never guess crucial details like Unity versions, scene ownership, execution order, or platform-specific behaviors. Automated tests are essential for validating AI-generated code, converting suggestions into verifiable evidence. A well-structured project brief should outline existing decisions, prohibited changes, and a clear definition of "done." Tasks should be broken down into verifiable units, with a plan requested before code for cross-system changes. Developers should measure accepted changes and review costs, not just lines of code. The goal is to make AI an effective assistant rather than a source of unverified experimental code.