The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains examples of malicious injection patterns (such as instructions to ignore prior rules or reveal system prompts) within references/evaluation-frameworks.md and references/system-prompts.md.- Evidence: These patterns are encapsulated within JSON test case objects and Python list variables for diagnostic purposes.
Context: They are provided to teach developers how to test for vulnerabilities and implement effective guardrails, not to execute against the current agent environment.
Surface Analysis: The skill processes user data through prompt templates. However, it explicitly mitigates this risk by documenting the use of XML delimiters and boundary markers (e.g., in references/context-management.md) and by providing defensive patterns in references/system-prompts.md.
[EXTERNAL_DOWNLOADS]: Reference files include Python snippets that utilize well-known external libraries and APIs for metrics calculation and structured data handling.
Evidence: References to openai, anthropic, HuggingFace's evaluate, and OpenAI's tiktoken.
Context: These are documented as standard implementation patterns for prompt engineering workflows and do not involve unauthorized background downloads or remote code execution.
[COMMAND_EXECUTION]: The reference files contain Python code for data processing and evaluation metrics (e.g., using scipy and scikit-learn).
Evidence: Implementation examples for F1 scores, BLEU metrics, and A/B testing logic.
Context: These are provided as static reference implementations for the user to review and adapt, rather than as executable scripts that run during the skill's operation.

prompt-engineer