The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a 'social listening' feature that ingests untrusted data from external sources (Reddit, Hacker News, Bluesky) via public API endpoints. This data is processed to generate draft comments and engagement strategies. Although this represents a vulnerability surface for malicious instructions embedded in social media posts, the risk is mitigated by an explicit instruction to the agent: 'Always ask before posting — output drafts, user approves and posts manually.'
Ingestion Points: External data enters the context through API responses fetched via shell commands in references/listening.md.
Boundary Markers: The skill relies on a human-in-the-loop boundary where the agent is prohibited from auto-posting.
Capability Inventory: The skill uses shell commands (curl, jq) to fetch data but does not have automated write access to the platforms.
Sanitization: No explicit sanitization of the API data is mentioned, placing the burden of validation on the human reviewer.
[COMMAND_EXECUTION]: The file references/listening.md provides functional shell command 'recipes' using curl and jq. These commands are intended to be executed by the agent to retrieve real-time social data. The commands target legitimate public endpoints of well-known services.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch JSON data from well-known services including reddit.com, hn.algolia.com (Algolia search for Hacker News), and bsky.app (Bluesky). These operations are core to the skill's social listening functionality and do not involve executable code downloads.

social