The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingest and processes untrusted data from external social media feeds, making it susceptible to indirect prompt injection attacks.
Ingestion points: references/listening.md contains instructions for the agent to fetch posts and comments from Reddit, Hacker News, Bluesky, LinkedIn, and X using curl and the dev-browser tool.
Boundary markers: The skill does not specify technical delimiters for separating external data from instructions, but it implements a mandatory human-in-the-loop checkpoint: "Always ask before posting — output drafts, user approves and posts manually."
Capability inventory: The skill uses curl for network data retrieval and the dev-browser tool to navigate social media platforms.
Sanitization: The provided instructions do not include methods for sanitizing, escaping, or validating the content retrieved from social media before processing it.
[COMMAND_EXECUTION]: The references/listening.md file provides the agent with specific shell commands using curl and jq to interact with public APIs of well-known services (Reddit, Hacker News, and Bluesky) for data retrieval purposes.

social