eight reviewer prompt audit

The problem

LLM prompts often fail silently in production due to hidden vulnerabilities like prompt injection, hallucination triggers, or fragile output schemas. Developers typically rely on "vibe-based" testing, which misses critical edge cases and results in unstable agent behavior.

What it does

• Performs an 8-point audit covering framing, safety, robustness, and cost.
• Identifies prompt injection vectors and privilege separation gaps.
• Scores prompts using a weighted rubric to provide a quantitative quality verdict.
• Generates specific diffs or rewrites for every identified weakness.
• Enforces strict gating rules for security and grounding criticals.

Frameworks & tools

Works with any LLM instruction set, including system prompts, agent personas, and tool-use instructions. Supports JSON schema validation for output contracts.

Why this beats prompting it yourself

Manual review is subjective and inconsistent. This skill applies a systematic, disjoint rubric that forces the auditor to evaluate safety and robustness independently of clarity, ensuring no axis is overlooked. It replaces guesswork with a repeatable, scored process that includes immediate code-level fixes.

Use cases

• Hardening an agent against prompt injection before deployment.
• Reducing token costs by identifying redundant instructions.
• Ensuring consistent JSON output for downstream parsers.
• Stress-testing prompts against adversarial inputs and edge cases.

Known limitations

Requires access to internal rubric and schema files to execute the full scoring logic. Injection safety analysis is most accurate when the runtime context and tool definitions are provided.