The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for processing external data, such as user questions and source documents, by interpolating them directly into LLM prompts which creates a potential surface for injection.
Ingestion points: The question, answer, and sourceContext variables in functions like judgeAnswer and verifyClaims in SKILL.md.
Boundary markers: The prompts use text-based labels (e.g., Question:, Answer:) but lack explicit structural delimiters (like XML tags or multi-line separators) or specific instructions to ignore embedded commands within the processed data.
Capability inventory: The provided code examples interact with the OpenAI API and include functionality for checking local module contents.
Sanitization: No sanitization or escaping of the external data is shown in the prompt construction examples.
[DYNAMIC_EXECUTION]: The skill includes a utility function that employs a dynamic loading pattern to verify the existence of symbols in modules.
Evidence: const mod = require(modulePath); in the verifyApiExists function within SKILL.md.
The function uses the Node.js require system to load a module based on a computed path string. If an agent attempts to verify code that contains malicious module paths, this could lead to the unintended execution of code during the loading process of local or built-in modules.

agent-self-evaluation-patterns