The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on multiple Python and Shell scripts to execute video processing tasks.
scripts/composite.py, scripts/voiceless/composite.py, and scripts/composite_final.py use subprocess.run to call ffmpeg and ffprobe with arguments derived from an external edit_plan.json file.
scripts/normalize_clip.sh uses ffprobe and bc to calculate video ratios, then interpolates these values into an ffmpeg filter string. If the duration values or input paths are manipulated, it could lead to command or filter injection.
While the Python scripts use list-based arguments to mitigate shell injection, the construction of ffmpeg filter complexes (e.g., -filter_complex, -vf) involves string concatenation of parameters from the input JSON, which remains a sophisticated injection surface.
[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection (Category 8) due to its processing of untrusted data.
Ingestion points: The skill reads edit_plan.json, which contains paths for video assets, subtitles, and music.
Boundary markers: None detected in the instructions or scripts to prevent the agent from obeying instructions embedded in the input files.
Capability inventory: The skill can execute subprocesses (ffmpeg, ffprobe), read files, and write outputs to the filesystem.
Sanitization: The scripts perform minimal validation on the content of the edit_plan.json. An attacker could specify sensitive local files (e.g., .env, ~/.ssh/id_rsa) in the subtitles or video fields, potentially causing ffmpeg to read and render those files into the output video artifact for exfiltration.
[DYNAMIC_EXECUTION]: The skill uses dynamic execution patterns (Category 10) by assembling executable command lines and complex filter graphs at runtime based on the structure of the provided edit plan. This includes calculating time offsets and durations that are then passed as parameters to system binaries.

stitch-videos-ffmpeg