The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (LinkedIn posts, blogs, YouTube scripts, etc.) which could contain malicious instructions designed to influence the agent's behavior during the repurposing phase.
Ingestion points: As described in Step 1 of SKILL.md, the skill accepts source material in "Any format: a post, a paragraph, a script, a caption, a transcript, a bullet list, a link to read."
Boundary markers: The instructions lack explicit directives for the agent to treat this input strictly as data or to use specific delimiters to prevent the execution of embedded instructions.
Capability inventory: The skill has the ability to read local reference files (e.g., voice-profile.md) and perform network write operations via the lib.publish tool to send content to Instagram.
Sanitization: There is no evidence of sanitization, filtering, or validation of the ingested source content before it is interpolated into the prompt for post generation.
[DATA_EXFILTRATION]: The skill mentions the requirement of a "Publora key" for its publishing workflow. While it does not hardcode credentials, a successful indirect prompt injection could potentially be used to trick the agent into exfiltrating this key or other sensitive reference data (like the user's voice profile) to an external target, although the primary publishing target is the well-known service Instagram.

ig-repurposer