Structured Files, Not Prompts, Are the Real Way to Govern AI Systems

Software engineer Derek Wang argues that relying on chat prompts to control AI model behavior is fundamentally flawed, since prompts exist only in a context window and can be ignored or forgotten across sessions. He proposes a 'constraint pyramid' — a five-layer file-based architecture where rules are written into project files the model must read before generating any output. The pyramid spans from a high-level README or constitution at the top down to constrained generated code at the bottom, with each layer narrowing the space the next can operate in. Constraints are further divided into three tiers: hard coded gates enforced by automated scripts, architecture decisions requiring formal revision, and loose execution suggestions the model may adapt but must document. Wang's central premise is that shifting governance from words in chat to rules in files transforms AI behavior from a hope into a verifiable, auditable system.