Prompt Engineering 2.0: install a semantic firewall, not more hacks

Most of us have broken the requests here that feels happening by chance: the model hallucinated quotes, the "Style Guide" collapses halfway, multiple instructions take nonsense, or calling up the right document, but the wrong section. I thought that would only be quirks ... until I started to map them, it turned out that they were not at all by accident. They are reproducible, diagnosed and fixable. Check the acceptance goals so that the system rejected an unstable output why it is important. This changes in the engineering of "Try it until it works" → diagnose once "forever". Instead of pursuing hacks after the model has failed, install a semantic firewall before the generation. If the semantic state is unstable, the system loops or reset. 70–85% + reproducible correctness. Add to the map you saw?