We need to stop treating Prompt Engineering like "dark magic" and start treating it like software testing.

Here's the scenario. You spend two hours brainstorming and manually crafting what you think is the perfect system prompt. You explicitly say: "Output strictly in JSON. Do not include markdown formatting. Do not include 'Here is your JSON'." You hit run, and the model spits back: Here is the JSON you requested: json { ... } It’s infuriating. If you’re trying to build actual applications on top of LLMs, this unpredictability is a massive bottleneck. I call it the "AI Obedience Problem." You can’t build a reliable product if you have to cross your fingers every time you make an API call. Lately, I've realized that the issue isn't just the models—it's how we test them. We treat prompting like a dark art (tweaking a word here, adding a capitalized "DO NOT" there) instead of treating it like traditional software engineering. I’ve recently shifted my entire workflow to a structured, assertion-based testing pipeline. I’ve been using a tool called Prompt Optimizer that handles this under the ho