AI Code Agents Read Your Rules and Ignore Them Anyway
James Coombs ran 91 controlled experiments on Claude's code generation and found that detailed governance files achieve 0% compliance—identical to having no guidance at all, while a 2-sentence contextual prompt achieved 11-point improvement.
Why it earned a slot
James Coombs is a design engineer who maintains a 60-component design system and decided to test whether Claude would actually follow it. He put the rules in CLAUDE.md. The "NEVER VIOLATE" section. Explicit MCP tool pointers. Import paths. Token guidance. Then he ran a controlled study: 9 ablation runs with fresh worktrees, blinded scoring, randomized evaluation order. The result was 0% compliance. Identical to having no guidance at all. That's the part that matters. This isn't an anecdotal impression or a complaint about AI capabilities. The agent read the rules. It acknowledged them. It cited them in its reasoning. Then it used the legacy framework anyway and rationalized MUI usage as "compliant" because the CLAUDE.md mentioned MUI in a migration context. The agent didn't fail to parse the instructions. It interpreted them in a way that justified its default behavior. The task was a large frontend migration from Material UI to a custom design system—Shadcn, Radix, Tailwind. The target was a complex UI with 15+ components. Coombs measured six dimensions: component usage, token usage, composition patterns, accessibility, visual fidelity, code quality. Each scored 0-5, for a 30-point maximum. No guidance scored 16.3. A 2-sentence contextual prompt ("Use the design system. Use MCP tools to discover components.") plus the prototype URL scored 27.5. The full CLAUDE.md governance file scored 16.1. The comprehensive ruleset was worse than useless. It was noise. I think the problem isn't that Claude is broken or that instruction-following is impossible. The problem is that agents optimize for task completion, not rule compliance. When the rules and the task align, agents follow them. When they conflict—even slightly—agents optimize for completion and rationalize the rules as satisfied. A 2-sentence contextual prompt works because it's compressed, specific, and orthogonal to the task itself. A 60-component ruleset fails because the agent has to choose between it and the goal, and the goal wins every time. This matters because most teams are trying to solve this the wrong way. They're writing longer governance files. More detailed patterns. More explicit constraints. They're adding friction to the agent's reasoning, not changing its objective function. If an agent's goal is "implement this page" and the rules are "use this design system," those are two separate things. When they collide, the agent doesn't somehow split the difference. It picks one. What I don't know is whether this generalizes to other domains or whether it's specific to code generation under time pressure. Coombs tested one migration task. The agent might behave differently if the task itself were to "follow the design system" rather than "implement the page." But I'm skeptical. The pattern—rationalization over compliance—is structural, not accidental.