gpt5.4 pushed me over the edge when I started using it to help with Unity projects. The writing of high quality mono behavior scripts was not the surprising part. It's the part where it once did a direct edit to a 500kb scene file (~yaml content) and came out the other side clean. The realization that apply_patch would work on any reasonably-structured plaintext format punched me in the gut. I had wasted a lot of time with tools that target specific content types and elaborate APIs over those files. I should have zoomed out a bit. These lessons keep piling on as the models become more capable.

Another "oh shit" moment was when I realized I can leave the system prompt entirely null. A properly organized agent can find its way into tool docs and iteratively work through an understanding of the environment relative to the user's prompt. The tools being more important than the prompt has actually been a massive relief for me. Magical string literals are so odious.