The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides numerous templates and scripts that interpolate untrusted user data into prompts, creating a surface for indirect prompt injection.
Ingestion points: Data enters through variables such as {query} in SKILL.md, {problem} in references/chain-of-thought.md, and {text} in references/prompt-optimization.md.
Boundary markers: Most provided templates lack delimiters (e.g., XML tags or specific markdown headers) or instructions to ignore embedded commands within user input.
Capability inventory: The skill includes capabilities to interact with external LLM APIs and execute logic-heavy reasoning chains (e.g., Tree-of-Thought), which could be manipulated via injection.
Sanitization: No sanitization or escaping logic is implemented in the provided template rendering classes or scripts.
[SAFE]: The Python utility scripts/optimize-prompt.py is well-implemented and uses standard libraries like numpy and concurrent.futures for performance testing and metric aggregation without any malicious side effects.
[SAFE]: All external references and libraries mentioned, such as Anthropic's Claude models, OpenAI's API, and the LangChain ecosystem, are standard tools within the industry and are used appropriately for educational and functional purposes.

prompt-engineering-patterns