The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill frequently executes remote code by using npx to fetch and run logic directly from a GitHub repository (socai-io/jev-social). This occurs during readiness checks, research searches, and when starting a local preview server.
[EXTERNAL_DOWNLOADS]: The skill fetches code from an external GitHub repository. Although it specifies a pinned commit hash (5270e23cfd27aace9055669ee396926973baa241), executing code downloaded at runtime from third-party sources is a high-risk operation.
[COMMAND_EXECUTION]: The agent is instructed to run shell commands via npx. While the instructions advise against string interpolation of user input and recommend using an argv-capable runner, the workflow relies on executing system commands to function.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from social media platforms (Instagram, TikTok, and LinkedIn), which could contain instructions intended to compromise the agent.
Ingestion points: Captured social media posts, profiles, and comments from external websites (SKILL.md).
Boundary markers: Explicit instructions are provided to treat all platform data as untrusted and to ignore embedded instructions (SKILL.md).
Capability inventory: Shell command execution via npx, browser session access, and local file writing for run records and media (SKILL.md).
Sanitization: The skill includes rules to extract only specific public fields and to avoid reproducing sensitive internal metadata, provider keys, or artifact paths (SKILL.md).

jev-social