video-to-srt

Generate timecoded SRT subtitles from local video or audio files. Use when a user wants a local low-cost subtitle workflow, asks to transcribe local media in...

Security Scan

The skill is internally consistent with its stated purpose (converting local media to SRT); it uses a local Python script and a virtualenv and does not request external credentials or hidden endpoints.

Purpose & Capability

The name/description match the included scripts: the Python transcriber and wrapper create a local venv, install faster-whisper, run transcription, and write an SRT. Minor mismatches: the registry metadata lists no required binaries, but the wrapper assumes a system python3 is available (and in practice transcription will often require system media tools like ffmpeg); the default language is zh, which is an unusual default but explainable for a China-focused workflow.

Instruction Scope

SKILL.md instructs the agent to locate and operate on the user's local media file and to run the provided wrapper script. That is appropriate for the stated purpose. SKILL.md asks to request permission before installing packages (good). One important runtime effect not explicitly called out: the WhisperModel/faster-whisper runtime will download model weights from the Hugging Face Hub (network I/O, potentially large downloads) and will cache them under HF_HOME/XDG_CACHE_HOME; this is expected but should be disclosed to users and controlled by the environment.

Install Mechanism

No explicit install spec in registry; the wrapper script creates a local virtualenv and pip-installs faster-whisper from PyPI via requirements.txt. This is a standard approach (moderate risk): pip packages are third-party code and may pull native wheels or additional dependencies. There are no obscure download URLs or extracted archives in the skill itself.

Credentials

The skill requests no external credentials or config paths. It does set local cache and venv locations inside the skill folder and allows overriding HF_HOME/XDG_CACHE_HOME/VENV_DIR for reuse — these are reasonable and proportional to the task.

Persistence & Privilege

always is false and the skill does not request persistent system-wide changes. The wrapper creates a venv and caches models in the skill folder by default (scope-limited). Autonomous invocation is allowed (platform default) but not combined with other concerning factors.

Assessment

This skill appears to do what it says: it runs a local Python transcription tool (faster-whisper) inside a venv and writes an SRT next to the input. Before installing or running it, consider: 1) It will pip-install faster-whisper (third-party code) into a local virtualenv — review and permit package installation. 2) The model weights will likely be downloaded from the Hugging Face Hub (network traffic and potentially many gigabytes of storage) unless you already have cached models — be prepared for bandwidth and disk usage and confirm you want that. 3) Ensure your environment has python3 (and typically an ffmpeg binary or equivalent media backend) even though the registry metadata didn't declare required binaries. 4) The skill operates on local files: confirm the agent/session has permission to read the media you plan to transcribe. 5) The default language is Chinese (zh); change to --language auto or set the language explicitly if that is not desired. If these points are acceptable, the skill is coherent and reasonable for local transcription.

Workflow

1. Create a new top-level task folder for the request in the workspace.
2. Locate the user's local media file and confirm the language choice only if it is unclear.
3. Run scripts/run_local_subtitles.sh from this skill folder.
4. Default to --language zh and --model small. Switch to --language auto for mixed language audio. Switch to --model medium only when the user wants better accuracy and accepts slower runtime.
5. Inspect the generated .srt file by checking the first and last few cues before handing it off.
6. Return the subtitle path and mention that editors which accept SRT, including Jianying desktop, can import it.

Stable Defaults

• Keep dependency installation inside a local virtual environment created by scripts/run_local_subtitles.sh.
• Keep caches local to the skill folder. This avoids macOS cache permission issues.
• Keep HF_HUB_DISABLE_XET=1 enabled. This avoids a common Hugging Face Xet download failure.
• Prefer SRT. It is the simplest subtitle format for broad editor compatibility.
• Override VENV_DIR, HF_HOME, or XDG_CACHE_HOME only when reusing an existing environment or model cache is helpful.

Commands

Use the wrapper script for the normal path:

scripts/run_local_subtitles.sh "/absolute/path/to/video.mp4" --output-dir "/absolute/path/to/task/output" --copy-next-to-input

Use these common variants:

scripts/run_local_subtitles.sh "/absolute/path/to/video.mp4" --language auto --output-dir "/absolute/path/to/task/output" --copy-next-to-input
scripts/run_local_subtitles.sh "/absolute/path/to/video.mp4" --model medium --output-dir "/absolute/path/to/task/output" --copy-next-to-input

Validation

• If the wrapper needs to install packages or download a model, request permission when required by the environment.
• Confirm that the final .srt exists.
• Preview the first and last cues with sed -n '1,24p' and tail -n 24.
• If the user wants a faster import workflow, keep a copy beside the source media with --copy-next-to-input.

Resources

• scripts/run_local_subtitles.sh: create or reuse a local virtual environment, install dependencies, configure stable cache paths, and run transcription.
• scripts/transcribe_to_srt.py: transcribe media and write timecoded SRT output.
• scripts/requirements.txt: minimal dependency list for the workflow.