The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates indirect prompt injection by providing template systems and optimization logic that interpolate external data directly into prompts.
Ingestion points: In scripts/optimize-prompt.py, the PromptOptimizer.evaluate_prompt method renders templates using test_case.input. Similarly, references/prompt-templates.md demonstrates various render methods that accept arbitrary keyword arguments.
Boundary markers: Although some templates in assets/prompt-template-library.md use descriptive labels like 'Context:' or 'Question:', they do not implement robust delimiters (e.g., XML tags or unique markers) nor do they include 'ignore embedded instructions' warnings for the LLM.
Capability inventory: The skill's scripts utilize ThreadPoolExecutor for parallel execution and are designed to interface with external LLM APIs (e.g., OpenAI), creating a pathway for malicious data to influence model actions.
Sanitization: There is no evidence of input validation or sanitization within the provided template rendering logic to detect or neutralize adversarial prompt injection attempts.

prompt-engineering-patterns