The Agent Skills Directory

[COMMAND_EXECUTION]: The skill bundle supports a "Tier 2" extensibility feature where users can define a custom publishing command via the FB_SKILLS_CUSTOM_POSTER environment variable. The lib/backend_selector.py script uses subprocess.run() to execute this command, passing the post data via stdin. While documented for advanced users to build custom posters on the Facebook Graph API, this mechanism allows for arbitrary command execution on the host if the environment variable is manipulated or misconfigured.
[INDIRECT_PROMPT_INJECTION]: Several skills (e.g., fb-audience-insights, fb-hook-extractor, fb-engagement-drafter) ingest untrusted data from external Facebook Pages and posts, including profile descriptions and user comments. This content is then processed by the agent to generate summaries, hooks, or replies.
Ingestion points: Untrusted data enters the agent context through Apify API calls and manual text pasting as described in the SKILL.md files.
Boundary markers: The instructions do not define explicit delimiters or instructions to ignore embedded commands within the ingested data.
Capability inventory: The skill has capabilities including local file system reads (.env, references/voice-profile.md), network requests to Publora, Apify, and Pixfaro APIs, and optional command execution via the FB_SKILLS_CUSTOM_POSTER variable.
Sanitization: The skill relies on standard LLM processing for data interpretation. The "approve-before-publish" pattern mandated in the SKILL.md files acts as a critical human-in-the-loop control to mitigate accidental execution of injected instructions.
[DYNAMIC_EXECUTION]: The lib/backend_selector.py script performs late imports of library modules (publora_client, pixfaro_client) based on the active configuration tier. Additionally, the aforementioned "Tier 2" custom poster feature involves executing a command string provided dynamically via environment variables.

facebook-marketing