cortex-prompt-audit

To audit Cortex agent prompts, apply an eight-item framework validating explicit JSON output directives, exact JSON shapes, and injection risk patterns. This ensures deterministic, parser-friendly behavior and prevents prompt injection before deployment.

Prompt injection prevention in Cortex templates involves scanning for data leakage patterns and harmful content using structured safety checks. It enforces explicit output constraints and unambiguous keys to ensure safe, predictable agent behavior.

Validate JSON output requirements by checking for explicit 'Return ONLY valid JSON' directives, verifying exact expected JSON shapes, and ensuring no ambiguous keys exist. This prevents parse errors and missing keys during agent execution.

Run a prompt safety audit during model updates and agent development to catch parse errors, missing keys, and injection risks before deployment. Avoid using it when operational guidance explicitly indicates the audit does not apply.

Yes, the Cortex prompt audit enforces scoring criteria including dollar-quoting and doc_search conventions. These constraints ensure parser-friendly prompts and deterministic behavior across development and production environments.

Cortex agent prompts cause parse errors when they lack explicit JSON output instructions, contain ambiguous keys, or miss dollar-quoting conventions. An eight-item audit framework detects these issues and provides actionable remediation notes.