The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for integrating untrusted external data, such as RAG retrieved context or user inputs, into LLM prompts. * Ingestion points: Found in SKILL.md (Integration Patterns section) and references/few-shot-learning.md (Template builders). * Boundary markers: Documentation explicitly notes that delimiters alone are insufficient and recommends external permission enforcement. * Capability inventory: Scripts like scripts/optimize-prompt.py focus on generation and metric tracking without executing the prompt strings as system commands. * Sanitization: Examples are conceptual and do not include complex sanitization, though they emphasize the need for validation.
[DYNAMIC_EXECUTION]: A custom template engine is illustrated in references/prompt-templates.md. It uses regular expressions to process logic blocks. The documentation provides clear warnings about template injection risks and the illustrative nature of the code.
[EXTERNAL_DOWNLOADS]: Implementation of advanced example selection in references/few-shot-learning.md references well-known Python libraries including numpy, scikit-learn, and sentence-transformers.

prompt-engineering-patterns