The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides architectures for interpolating data into prompt templates, which is a core function for prompt engineering but inherently creates a surface for indirect prompt injection if inputs are untrusted.
Ingestion points: Data entering through kwargs in the PromptTemplate.render method (references/prompt-templates.md) and test_case.input within the optimization script (scripts/optimize-prompt.py).
Boundary markers: Most templates use direct string interpolation; however, the skill documentation explicitly advises on using system prompts to set constraints and safety guidelines.
Capability inventory: The skill's logic is centered on LLM completions and metric calculations; no dangerous subprocess or file-system operations are performed on interpolated data.
Sanitization: The provided utility scripts do not perform automated sanitization, as the skill is intended for developer-led prompt construction and refinement.
[EXTERNAL_DOWNLOADS]: The skill's reference documentation and scripts mention standard, well-known libraries and services including OpenAI, Sentence-Transformers (Hugging Face), NumPy, SciPy, and Scikit-Learn. These are industry-standard tools for the domain of prompt optimization and natural language processing.
[SAFE]: No evidence of malicious behavior, credential exfiltration, persistence mechanisms, or obfuscated content was found across the skill's scripts and documentation.

prompt-engineering-patterns