The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains several common prompt injection strings (e.g., 'Ignore previous instructions', 'DAN', 'reveal your system prompt'). However, these are explicitly presented as test cases within an 'Injection Test Suite' in references/system-prompts.md and references/evaluation-frameworks.md. Their purpose is to help the user test the resistance of their own prompts to such attacks, not to override the agent's behavior. The skill also includes specific instructions on implementing guardrails and injection defenses.
[DATA_EXFILTRATION]: The skill documentation and code snippets do not contain any network operations targeting non-trusted domains or sensitive file access. It explicitly instructs users in SKILL.md not to hardcode sensitive data in prompts or examples, and provides guidance in references/system-prompts.md on preventing PII exposure.
[EXTERNAL_DOWNLOADS]: The skill references several standard data science and AI libraries (e.g., scikit-learn, openai, anthropic, pydantic, tiktoken). These are well-known, trusted packages consistent with the skill's primary purpose of prompt engineering and evaluation.
[REMOTE_CODE_EXECUTION]: There are no patterns of remote script execution or dynamic code evaluation from untrusted sources. All Python and TypeScript code snippets provided are illustrative templates for local implementation of evaluation frameworks and schema validation.

prompt-engineer