I started by reverse-engineering the SousChef CMS specification — not as a developer reading documentation, but as a designer mapping the content system's logic. Every field, every enum, every relational dependency became a design constraint. I identified three layers of complexity the AI agent would need to handle: structural compliance (correct field types, required fields, nesting), content quality (instructions written for a 12-year-old, warm tone, one action per step), and relational integrity (each ingredient appears exactly once, timer triggers match countdown steps, variant assignments are correct).

Rather than treating the AI as a black box that either works or doesn't, I designed the system as a pipeline with explicit checkpoints. The AI generates structured JSON from raw recipe text. Pydantic validates every field against the schema. A normalize function catches predictable errors before they reach validation. And when validation fails, a critique-and-repair loop tells the AI exactly what broke and asks it to fix only that — up to three attempts before flagging for human review.