browser-use/video-use
video-use lets shell-capable coding agents edit from word-level transcripts and targeted visual composites, then render with FFmpeg. Its approval and self-check loop is a useful pattern for media agents.
The open-source **video-use** skill turns word-level transcripts into an approximately **12KB text surface**, requests visual composites only at decision points, and renders edits through FFmpeg. It supports cuts, subtitles, grading, overlays, and session memory.
For media-agent workflows, keep the reasoning representation compact and make execution explicit: inventory sources, propose a strategy, wait for approval, produce an edit decision list, render, inspect every cut boundary, and persist project context.
The open-source **video-use** skill turns word-level transcripts into an approximately **12KB text surface**, requests visual composites only at decision points, and renders edits through FFmpeg. It supports cuts, subtitles, grading, overlays, and session memory. For media-agent workflows, keep the reasoning representation compact and make execution explicit: inventory sources, propose a strategy, wait for approval, produce an edit decision list, render, inspect every cut boundary, and persist project context. The approach depends on an ElevenLabs transcription key, and the model does not continuously inspect the source video. Self-evaluation retries fixes at most **3 times**, so visual or editorial errors may still require human review.
video-use makes compact context and staged execution concrete for media agents: transcripts carry most reasoning, visual inspection is requested selectively, and edits become an explicit, reviewable render pipeline. It reinforces minimum-necessary processing and reusable workflow state, while narrowing reliability claims because sparse viewing, external transcription, capped retries, and cut-by-cut inspection still leave human approval essential.