The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to install external dependencies using npx skills add. It specifically references healthier-vitamins/agentic-grimoire (a vendor-related resource) and mattpocock/skills (a third-party repository). These commands download and integrate external code into the agent's environment. Additionally, the generated interactive pages may reference libraries like JSXGraph from external CDNs.
[DYNAMIC_EXECUTION]: In Step 7, the skill directs the agent to write a Node.js script that extracts and evaluates JavaScript code from the generated HTML. This process of runtime evaluation of AI-generated code is used to verify mathematical results but represents a dynamic execution pattern.
[COMMAND_EXECUTION]: The skill performs environment discovery by searching for specific directories such as ~/.claude/skills/ and ~/.agents/skills/ to find absolute paths for dependent skills. It also suggests running shell commands to install those dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data via the $ARGUMENTS field (which can contain formulas or screenshots) and uses this data to generate logic and code for an interactive page. This code is later executed for verification, creating a potential path for indirect injection.
Ingestion points: User-provided formulas or screenshots in the $ARGUMENTS field.
Boundary markers: None present; the instructions do not specify delimiters or instructions to ignore content within the formulas.
Capability inventory: The skill can write files to the local directory and execute Node.js scripts via the agent's environment.
Sanitization: There is no mention of sanitizing or validating the input formula before it is incorporated into JavaScript functions for execution.

orrery