The Agent Skills Directory

[PROMPT_INJECTION]: The reference files, specifically system-prompts.md and evaluation-frameworks.md, contain known injection patterns such as "Ignore previous instructions", "DAN", and requests to reveal system prompts. These are explicitly used as entries in a test suite (e.g., the INJECTION_TESTS list) and are provided to help users test the robustness of their own models and prompts rather than as an attempt to compromise the agent's behavior.
[INDIRECT_PROMPT_INJECTION]: The skill includes extensive documentation on identifying and mitigating indirect prompt injection. It recommends using boundary markers (e.g., XML tags), input validation, and schema enforcement to protect against instructions embedded in user-controlled data.
[COMMAND_EXECUTION]: The skill provides example GitHub Actions workflows and Python scripts illustrating how to automate prompt evaluation. These are intended as boilerplate code for developers to use in their own CI/CD pipelines and are not executed by the skill at runtime.

prompt-engineer