The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes ffmpeg and ffprobe for video assembly, text measurement, and metadata extraction.
Commands are executed via subprocess.run using argument lists rather than shell strings, which effectively prevents shell injection vulnerabilities.
The diagnostic function in scripts/build_movie.py proactively redacts credentials, tokens, and passwords from command error output using regular expressions before they are reported.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of JSON manifests containing frame paths, captions, and narration instructions.
Ingestion points: Data enters the system via load_manifest in scripts/build_movie.py and is parsed into processing structures.
Boundary markers: The system relies on strict structured data validation (e.g., finite() checks for numbers, regular() checks for paths) rather than natural language delimiters.
Capability inventory: The skill can write media files (MP4, GIF), subtitle sidecars (SRT), and execute media processing binaries.
Sanitization: The literal function in scripts/presentation.py escapes ASS (Advanced Substation Alpha) control characters (braces and backslashes) in user-provided captions to prevent injection into the libass rendering engine.
[EXTERNAL_DOWNLOADS]: The skill depends on external binaries (ffmpeg) but maintains a high security posture regarding their acquisition.
Instructions in SKILL.md and references/assembly.md explicitly forbid the automatic installation of tools or dependencies.
The tool discovery logic in scripts/build_movie.py uses shutil.which and specific path lookups (like Homebrew prefixes) but does not attempt to download or install missing software.
[DYNAMIC_EXECUTION]: The main assembly script build_movie.py dynamically loads the presentation.py module at runtime.
This loading is restricted to a local file within the skill's own directory using importlib.util, which is a standard and safe practice for local modularity.
[DATA_EXPOSURE]: Path validation is enforced globally via the regular function in scripts/build_movie.py.
All input files are verified to be inside the manifest's run directory using is_relative_to, preventing path traversal and unauthorized file access.

tk-video