The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'social listening' workflow in references/listening.md that fetches and processes content from external, attacker-controllable platforms such as Reddit, Hacker News, and Bluesky. This creates a surface for indirect prompt injection where malicious text within a social media post could attempt to override the agent's instructions during the scoring or drafting phase.
Ingestion points: Content is ingested from public JSON APIs using curl and from browser sessions using the dev-browser tool.
Boundary markers: The instructions do not specify the use of strict delimiters to separate external data from system instructions, although the skill includes a strong mitigation by requiring human approval ('Always ask before posting — output drafts, user approves and posts manually').
Capability inventory: The agent is instructed to use curl, jq, and dev-browser for data retrieval. There are no capabilities present for writing to local files or executing remote scripts.
Sanitization: No explicit sanitization or filtering of the retrieved content is described in the provided references.
[COMMAND_EXECUTION]: The skill provides the agent with functional shell command templates ('curl recipes') for interacting with public web APIs to perform social listening tasks. While these involve the execution of shell commands, they are restricted to data retrieval from well-known services and do not involve privilege escalation or sensitive system access.

social