The Instruction Layer Auditor is a skill designed for developers who build and maintain AI agents with complex, multi-layer instruction stacks. Modern agents commonly combine system prompts, developer instructions, repository-level rule files such as AGENTS.md, skill-specific logic, and persistent memory. When these layers collide, the result is behavior drift: agents that become over-cautious, ignore instructions, or hallucinate constraints they were never given.
This skill acts as a structured debugger for that layered configuration. It maps out every instruction layer present in a given stack and extracts the operational rules each layer defines. From there it identifies direct contradictions between layers—for example, a system prompt permitting an action that a repo-level rule prohibits—and flags hidden logic loops that cause unpredictable behavior.
Beyond detection, the skill produces a normalization plan. This plan specifies which rules are misplaced and recommends moving them to the highest-precedence layer where they belong, reducing logic bloat without requiring the developer to manually trace thousands of lines of stacked prompts. It also performs a risk assessment, predicting whether a given stack configuration is likely to cause over-verbosity or tool-use failures before those symptoms appear in production.
This skill is appropriate when an agent’s behavior has degraded after iterative prompt additions, when multiple contributors have modified an instruction stack independently, or when onboarding a new agent configuration that was inherited from another team.