The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/vd_frames.py executes the ffmpeg command-line utility via subprocess.run to extract frames from video files.
Evidence: The script calls subprocess.run(["ffmpeg", "-y", ...]) to process video files at specific timestamps derived from input metadata.
While the use of a list for arguments mitigates shell injection, the parameters (timestamps and filenames) are derived from the manifest.json and session markdown files within the input directory, meaning the command's behavior is influenced by the content of the processed items.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from video files and markdown highlights, which are then passed to a Vision LLM for slide selection and verification.
Ingestion points: The skill reads highlight text from {item_dir}/sessions/*.md and video content from {item_dir}/source.mp4 via extracted frames.
Boundary markers: The prompts in references/select-prompt.md and references/verify-prompt.md lack explicit delimiters or instructions for the LLM to disregard potentially malicious instructions embedded within the images (OCR injection) or the highlight text.
Capability inventory: The skill has the ability to execute shell commands (ffmpeg) and write files to the local file system (frames, montages, and selection JSON files).
Sanitization: The parse_highlights function in scripts/vd_frames.py performs minimal regex-based parsing to extract timestamps and text but does not sanitize the text for instruction-bypass patterns.

video-digest-slides