The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external documents for verification and span selection, which presents an inherent surface for indirect prompt injection.
Ingestion points: The skill ingests text from "authorized source" documents and "candidate spans" defined in the request JSON.
Boundary markers: Instructions explicitly mandate copying "the original span selected by ID" and forbid synthesizing extracted fields or fabricating quotations.
Capability inventory: The skill utilizes the jev-decide command to perform network API calls to LLM providers (OpenRouter or TypeSafe).
Sanitization: It requires preserving "source links for human checking" and reporting unsupported statements, ensuring results are verified by a human user.
[COMMAND_EXECUTION]: The skill uses the jev-decide CLI tool to validate and submit document processing requests. It provides specific instructions for dry-run validation and specifies that API keys must be inherited from the environment rather than placed in command arguments or prompt text.
[EXTERNAL_DOWNLOADS]: The skill identifies a dependency on the jev-skill package and references the author's GitHub repository (github.com/wuyoscar/jev-skill) for setup instructions. It specifically instructs the agent not to install software silently and to explain missing dependencies to the user instead.

jev-documents