The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates and scripts that interpolate untrusted data directly into LLM prompts, creating a surface for indirect injection attacks.\n
Ingestion points: The PromptTemplate class in references/prompt-templates.md and the PromptOptimizer class in scripts/optimize-prompt.py accept arbitrary key-value pairs via kwargs and TestCase.input for rendering into final prompt strings.\n
Boundary markers: The provided template rendering logic (including the ConditionalTemplate) does not implement or encourage the use of clear delimiters or instructions to ignore embedded commands within user-supplied data.\n
Capability inventory: The PromptOptimizer, PromptABTest, and TreeOfThought implementations are designed to automatically execute the generated prompts against an LLM using methods like client.complete() or openai.ChatCompletion.create().\n
Sanitization: No sanitization, validation, or escaping logic is present to filter or neutralize potentially malicious content within user-supplied variables before they are processed by the LLM.\n- [EXTERNAL_DOWNLOADS]: The skill incorporates external model loading from well-known services to support semantic analysis and example selection.\n
The SemanticExampleSelector and DiversityExampleSelector in references/few-shot-learning.md use the sentence_transformers library to download and load the all-MiniLM-L6-v2 model from Hugging Face's official repository.

prompt-engineering-patterns