kajisho5/ffmpeg-skill v2.22.5 has been released. According to the GitHub release notes, this version now rejects playlist-format inputs and rejects output names that begin with a dash. Files created alongside the output are also protected in the same way as the output. The technical documentation states that text read from files is treated as data rather than instructions. This change is an update that clearly draws the boundary between input formats and output file handling.
This project is a set of FFmpeg tools built so that coding agents such as Claude Code, Cursor, and Codex can perform video editing locally. The repository description states that it provides 42 FFmpeg tools through a machine-readable contract and an MCP server. Supported tasks include video trimming, subtitle processing, screen layout reconstruction, volume adjustment, and per-platform export. It is configured for use without a cloud service or API key, and can be used when an agent connects video editing capabilities.
The verifiable scope extends to the tool composition listed in the repository description and the input/output protection changes described in the release. The available material contains no measurements showing the performance of each feature or the actual scope of blocking. So this version is more accurately seen as an update that tightened the stated file handling boundaries, rather than read broadly as a general quality improvement across editing.