Vibe video editing: Codex edits a vlog from raw footage

My notes

Summary

A creator demos “vibe video editing”, pointing a coding agent (Codex / Claude Code) at a folder of raw vlog footage with the open-source Video Use repo as toolset, and asking it in plain English to assemble a vlog. The agent picked cuts, sequenced scenes, and produced a six-minute edit without any manual trimming. Reported as working with both Codex and Claude Code, with slightly better results on “GPT 5.5”.

Key Insight

• The interesting part is not stitching, it is cut selection. The agent decided what to keep and what to drop based on a one-line creative brief (“natural, chill, but not boring”).
• This is the same pattern as vibe coding (LLM + tool repo + plain-English brief), now applied to non-text media. Video Use functions as the MCP-like tool layer that gives the agent ffmpeg / scene-detection / transcription handles.
• Quality bar from the demo: “engaging, not boring”, good enough for personal vlogs and rough cuts, almost certainly not good enough yet for client-facing brand work where pacing and music drops matter.
• Removes the worst part of vlogging (multi-hour timeline scrubbing) and keeps the part creators actually want (record + publish). Lowers the activation energy enough that more people will try long-form video.
• For a marketing agency, short-form ad cuts from long b-roll archives is the obvious first commercial use, dozens of variants per brief, near-zero edit time.
• Watch the source carefully. A 1:37 TikTok is a hype demo, not a benchmark. Re-test with own footage before recommending to clients.