The Agent Skills Directory

[DYNAMIC_EXECUTION]: The file templates/tool-use-advanced.ts includes a calculatorTool definition that uses the eval() function to process mathematical expressions. While the code includes a comment warning about the dangers of eval(), implementing this pattern in an AI-agent workflow allows for potential arbitrary JavaScript execution if the model is manipulated by untrusted input.
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for processing untrusted external data (such as user messages and documents) that could contain malicious instructions.
Ingestion points: Data enters the agent context via the messages array in templates/basic-chat.ts, file reads in templates/nodejs-example.ts, and image metadata in templates/vision-image.ts.
Boundary markers: The provided templates do not demonstrate the use of recommended delimiters or explicit instructions to ignore embedded commands within processed data.
Capability inventory: The skill describes capabilities including network operations (via Claude API), filesystem writes in templates/nodejs-example.ts, and dynamic code execution in templates/tool-use-advanced.ts.
Sanitization: Although SKILL.md suggests sanitizing tool results to prevent JSON parsing failures (handling U+2028/U+2029), there is no evidence of sanitization intended to mitigate prompt injection attacks.

claude-api