The Agent Skills Directory

[SAFE]: The skill's primary purpose is to provide guidance and tools for prompt engineering. All instructions and scripts align with this functional goal without introducing security risks.
[EXTERNAL_DOWNLOADS]: The ab_eval.py script utilizes the official anthropic Python package. This is a well-known, trusted library for API interaction, and its use here follows standard development practices.
[INDIRECT_PROMPT_INJECTION]: The prompt evaluation script processes external data from JSONL files, representing a standard attack surface for indirect injection.
Ingestion points: Reads user-defined test cases from JSONL files in scripts/ab_eval.py.
Boundary markers: Employs XML delimiters (<input>, <expected_behavior>) in the judge prompt template to isolate untrusted content from instructions.
Capability inventory: Performs automated LLM API calls using the anthropic SDK to generate and evaluate responses.
Sanitization: Relies on structural separation (XML tags) rather than explicit content filtering, which is consistent with professional prompt engineering tooling.

prompt-optimizer