The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process video transcripts which are considered external, untrusted data.\n- Ingestion points: Data enters the workflow via the read_script tool output as described in the Workflow section of SKILL.md.\n- Boundary markers: The instructions lack explicit markers or instructions to isolate the transcript content from agent commands, creating a risk that malicious text within a transcript could influence agent behavior.\n- Capability inventory: The skill employs tools with significant modification capabilities across its scripts, including cut_narration, cut_range, and delete_shot.\n- Sanitization: There is no evidence of sanitization or validation of transcript content to prevent instructions from being interpreted as commands.

talking-head-cleanup