The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted user input which is then interpolated into prompts for downstream agents and tools. Ingestion points: The skill takes direct user input from the initial message and the review gate. Boundary markers: The skill uses xml_escape in several LLM prompts but lacks explicit delimiters to separate user data from instructions in others. Capability inventory: The agent has access to filesystem-write via the write_file tool and process-control via sub-skills that invoke shell-based media tools like ffmpeg. Sanitization: Input is processed using slugify, truncate, and xml_escape before being used in file paths or prompts.
[COMMAND_EXECUTION]: The skill requires process-control capabilities to orchestrate media assembly, stitching, and subtitle burning using system-level tools like ffmpeg via the video-merger and subtitle-burner skills to generate the final MP4 deliverable.
[EXTERNAL_DOWNLOADS]: The skill utilizes external image and video generation models (such as Gemini and Seedance) to produce visual assets based on the user-approved script. These calls are mediated by authorized sub-skills and are only executed after user confirmation.

meta-short-drama