The Agent Skills Directory

[PROMPT_INJECTION]: The prompt templates provided in the skill (specifically in the classifyTicketPrompt function) demonstrate a pattern of direct string interpolation for user-provided data, which creates a surface for indirect prompt injection.
Ingestion points: The ticket argument in the classifyTicketPrompt function in SKILL.md is interpolated directly into the system instructions.
Boundary markers: No delimiters or explicit boundary markers (e.g., XML tags or "ignore instructions" warnings) are used to wrap the untrusted input.
Capability inventory: The provided code patterns include a network-capable LLM client wrapper (llmCall) using the Anthropic SDK.
Sanitization: The example code does not include logic to sanitize or escape the input string before it is passed to the LLM.
[SAFE]: The skill uses secure practices for credential management in its CI/CD examples, referencing GitHub Secrets (${{ secrets.ANTHROPIC_API_KEY }}) rather than hardcoding values.
[SAFE]: The code patterns include mandatory schema validation for LLM responses using the Zod library, which is a recommended security and reliability practice for handling unstructured model output.

llm-patterns