Recently ran into pitfalls with AI coding. Originally wanted to use tests to reduce acceptance costs, defined a goal in the flow, wrote test scripts first, and then ran standard tests after making changes. Ended up introducing an extremely bad loop. Tweaking the prompts didn’t help, the goal was too complex, and the AI got lazy and downgraded to just bluffing. The flow rewarded passing tests, so the AI kept endlessly bluffing me, just doing surface-level work. A bug on GPT-5.5 high couldn’t be solved in 3 days—drove me crazy.