The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests potentially untrusted data from both external transcript files and user-supplied text areas during the review process.
Ingestion points: Source transcripts are read from the library/*.md directory using the read_script tool (Step 4), and user-modified transcript revisions are collected via the reviewedSourceTranscript form field (Step 6).
Boundary markers: The skill relies on XML tags for its widget-based review system, defining clear start and end points for user-controlled data.
Capability inventory: The skill possesses the capability to submit data to the submit_video_translation tool, which triggers backend media processing and credit-based billing.
Sanitization: The instructions explicitly mandate the use of XML attribute escaping when placing recognized text into the widget's default field. It also enforces a human-in-the-loop review process where the agent must wait for explicit user submission via the widget before proceeding to the final translation step.

video-translation