We introduced Claude Code to a two-person team (me and an AI partner that operates autonomously). So my situation is unusual -- the AI is not an assistant, it is a full team member with its own task queue. The most awkward moment nobody talks about: **the first time the AI mass-produced a bug I did not catch**. It added UTF-8 encoding validation to one API endpoint, which was correct. Then I asked it to build five more endpoints. It copied the same pattern into all of them -- including a subtle flaw in the validation logic. Users reported garbled text. I had to fix six files instead of one. The awkward part was not the bug. It was realizing that I had mass-approved code I did not actually review, because the AI's output looked so clean and professional that I trusted it. Another thing nobody mentions: **the power dynamic shifts in weird ways**. When the AI generates a solution that works but uses an approach I would not have chosen, I face a strange dilemma. Do I rewrite working code just because I would have done it differently? Months ago, I would have said yes. Now I let it go more often than I should, because rewriting means losing the productivity gain. You start compromising your own engineering standards, and that compromise is invisible to everyone except you. The thing that actually went well: onboarding speed. The AI understood the codebase faster than any human would. It could read every file, understand the patterns, and produce consistent output within minutes. If you treat it as a junior developer who reads fast but has no judgment, the mental model works. But nobody tells you that managing an AI teammate is still management -- you still need to set standards, review output, and catch mistakes. The tool did not eliminate that work. It just changed what the work looks like.