instruction-layer-auditor
Audit and de-conflict complex agent instruction stacks to fix inconsistent behavior and logic bloat.
skill install https://www.promptspace.in/skills/instruction-layer-auditorDebug Your AI's Logic Stack
When an AI agent behaves inconsistently, it’s rarely a single prompt issue—it's usually a layer conflict. Modern AI development involves stacking system prompts, developer instructions, repository-level rules (like AGENTS.md), skill-specific logic, and persistent memory. These layers often collide, leading to "behavior drift" where the agent becomes over-cautious, ignores instructions, or hallucinates constraints.
What it does
The Instruction Layer Auditor acts as a debugger for your agent’s "operating system." It maps out every instruction layer, extracts operational rules, and identifies precise points of failure. Rather than just adding more instructions to fix a bug, this skill helps you prune and normalize your prompt stack for maximum reliability.
- Layer Mapping: Identifies conflicts between system prompts and local repo rules.
- Conflict Detection: Flags direct contradictions and hidden logic loops.
- Normalization: Provides a strategic rewrite plan to move rules to their high-precedence layers.
- Risk Assessment: Predicts when a stack will cause over-verbosity or tool-use failures.
Why use this skill?
Manually auditing thousands of lines of stacked prompts is error-prone. This skill uses a structured framework to find "misplaced" rules—like a global formatting constraint buried in a local memory file—and tells you exactly where they should live to avoid logic bloat.
Use cases
- Identify conflicting rules across system prompts and repository-level docs.
- Reduce prompt token bloat by identifying redundant instructions across layers.
- Fix "over-cautious" behavior caused by overlapping safety or logic constraints.
- Normalize where rules live to ensure predictable agent behavior across a team.
- Clean up stale instructions in durable memory or long-term context windows.
Example
Prompt
Output
LAYER MAP: System (Global), AGENTS.md (Repo), Memory (Context). CONFLICTS: System says 'Be concise'; AGENTS.md says 'Explain every step'. RISK: Agent will likely pause for confirmation unnecessarily. PROPOSAL: Delete 'verbose' rule in AGENTS.md; move 'step-by-step' logic to the 'Code-Review' skill.