The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from local files (e.g., validated_ideas.json) and user-provided inputs to generate video briefs.
Ingestion points: Reads validated_ideas.json and accepts raw video ideas/constraints from the user.
Boundary markers: Lacks explicit instruction to ignore commands embedded in the source data, though it utilizes a mandatory human review step before saving output.
Capability inventory: Uses WebSearch, Read, and Write tools to research topics and save markdown files.
Sanitization: No explicit sanitization of ingested content is documented.
Note: This represents a standard attack surface for LLM applications and is mitigated by the human-in-the-loop requirement (Step 4) and the narrow scope of output (markdown documents).

yt-brief