The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a significant amount of untrusted external data, including TikTok video captions, transcripts, and OCR-extracted text from video frames. This content is passed to an LLM for classification and sorting (Step 4), which presents a risk where malicious instructions embedded in the external content could attempt to influence the agent's behavior.
Ingestion points: Video metadata fetched via Apify, OCR results from decompose_video.py, and Whisper-generated transcripts.
Boundary markers: The instructions do not specify delimitation or sanitization steps to isolate external text from internal instructions.
Capability inventory: The skill possesses file-writing capabilities, video rendering via render_video.py, and automated social media posting via schedule_post.py (Step 9).
[COMMAND_EXECUTION]: The skill relies on several local Python scripts (e.g., carousel_scan.py, decompose_video.py, engagement.py) to automate its workflows. These scripts execute shell commands and interact with the local filesystem and external APIs. While these operations are intrinsic to the skill's function, they represent a significant local execution surface.
[DATA_EXFILTRATION]: While not malicious, the skill includes a specific 'Privacy Gate' procedure (Step 7 and Step 5 of the reference) aimed at identifying and removing sensitive information—such as contacts, camera roll photos, and notifications—from user-provided screen recordings before video rendering, which is a notable security best practice.

ugc-hooks