Skill or MCP Server? Ask What Stops the LLM From Disobeying

When building LLM applications, it is crucial to distinguish between methods that inform an LLM of a rule and those that enforce it through code. This distinction is explored through the problem of preventing a coding agent from seeing test acceptance criteria to ensure valid, independent testing. While tools like Skills provide instructions, they are ultimately advice that the LLM can choose to ignore. In contrast, mechanisms like context assembly (programmatically cutting information from the prompt) or MCP servers (API-based tools with hardcoded logic) enforce rules by removing the LLM's ability to disobey. The central question for developers should be what mechanism truly stops the LLM from breaking a rule, rather than simply which tool to use.