The Agent Skills Directory

[PROMPT_INJECTION]: Static analysis flagged patterns like "Ignore previous instructions" in references/production-patterns.md and references/security-patterns.md. These instances are part of educational examples for building test suites or regex patterns for injection detection, rather than instructions to the agent.
[COMMAND_EXECUTION]: The skill includes uv run python commands in SKILL.md and eval/README.md. These are documentation for developers to manage the skill's lifecycle and are not executed by the agent automatically.
[DATA_EXFILTRATION]: No network exfiltration patterns were found. The skill configuration specifically disables network access (network: none) and provides guidance on isolating untrusted data.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle user-provided data. It addresses this inherent risk by providing a dedicated security guide (references/security-patterns.md) that outlines techniques for boundary enforcement and input isolation.

prompt-engineering