The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user data (text or JSON) through the Jev semantic judgment API to perform tasks like selection, comparison, and extraction.
Ingestion points: User-supplied materials are ingested as 'state' in evaluation requests, as described in SKILL.md and references/runner.md.
Boundary markers: The instructions recommend using named state fields and backticked paths (e.g., `path`) to delimit evidence and instructions, which helps differentiate data from commands in the remote prompt.
Capability inventory: The skill includes the scripts/jev.mjs script which utilizes fetch for network operations and node:fs for reading and writing configuration and result files. The agent is also instructed to generate additional .mjs scripts for complex workflows.
Sanitization: The jev.mjs script performs rigorous JSON structure and type validation on both the outgoing request and the incoming API response to ensure compliance with the service contract.
[DYNAMIC_EXECUTION]: The SKILL.md file explicitly permits the agent to generate and execute additional .mjs task code when preprocessing, loops, or complex compositions are necessary to fulfill the user's request. This represents a controlled dynamic code generation surface intended for task automation.
[COMMAND_EXECUTION]: The skill operates by executing the local helper script scripts/jev.mjs using the Node.js runtime to perform evaluations, manage configurations, and validate inputs.

jev