The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided topics and angles to generate social media content, which creates a natural surface for potential instruction injection in the input data.
Ingestion points: Topic, angle, and media details supplied by the user in SKILL.md (Step 1).
Boundary markers: The skill generates an "Approval card" (Step 9) that separates the content generation phase from the execution phase, ensuring the user reviews the final text.
Capability inventory: The skill uses lib.publish (Step 10) to transmit data to Instagram and read local media files for upload.
Sanitization: Content is scrubbed during the "Humanizer pass" (Step 7) and validated against a specific checklist in references/caption-checklist.md before being presented for approval.
[DATA_EXFILTRATION]: The skill reads local media files and reference configurations, but only for the purpose of publishing content to a well-known service.
Evidence: Accesses ../../references/voice-profile.md for context and reads user-specified local files for the media parameter in lib.publish.
Context: These file operations are core to the skill's utility and are restricted to files explicitly needed for the Instagram post, with the final transmission requiring user consent.

ig-caption-writer