Vibe Editing Agent Skills

Two Agent Skills for editing video by conversation. Raw takes become an approved edit decision list and a frame-accurate FFmpeg cut, with captions moved to the new timeline. The second skill reviews the render with evidence before it ships. They work next to the HyperFrames and Remotion skills.

What it is

“Vibe editing” is video editing through conversation. The author gives the intent and the notes, and the agent changes the project, runs the tools and shows the next version. It works when each decision is written down and can be checked.

Two Agent Skills make the agent write the cut as an edit decision list (EDL) with a reason for each clip. The agent waits for approval, keeps source time and output time apart, and says what it checked and what it did not. The skills work in any tool that supports the Agent Skills open format.

Next to HyperFrames and Remotion

HyperFrames and Remotion ship their own agent skills for titles, overlays, explainers and product videos. Install those too. These skills cover the work around them:

The cut. Takes, pauses and false starts, clips from long recordings, an EDL that the author approves, and a frame-accurate FFmpeg render of it.
The timeline maths. Each caption, insert and review note moves from source time to output time, and goes to HyperFrames or Remotion in their own format.
The review. The rendered file against the brief, the EDL and the captions, plus the checks that no script can do.

Install

npx skills add https://github.com/alexey1312/Vibe-Editing-Agent-Skill

Claude Code takes the plugin, next to HyperFrames:

/plugin marketplace add alexey1312/Vibe-Editing-Agent-Skill
/plugin marketplace add heygen-com/hyperframes
/plugin install vibe-editing-skills@vibe-editing-skills
/plugin install hyperframes@hyperframes

Then ask for a plan, for example: Use the vibe editing cut skill to plan a 40-second video from the takes in originals/.

The two skills

Skill	What it does
`vibe-editing-cut`	Raw takes, interviews, podcasts, tutorials and screen recordings into a cut: inventory, transcripts, the EDL, the draft render, then captions, sound and inserts on the new timeline
`vibe-editing-review`	A render against the brief, the EDL and the captions: black, frozen and silent parts, loudness, caption limits, one full viewing; a broken render; the handoff; a format turned into a series

Plan before cut

Report the machine. Can it render frame-accurately, burn captions, measure loudness, recognise speech locally, and run HyperFrames or Remotion? The report names the keys that are set, never their values.
Inventory. Find variable frame rates, rotation, HDR and mixed formats before they break a cut.
Plan. Each item has an ID such as CUT-03, source and output times, a reason, evidence and a way to verify it. The EDL is the sum of the approved items.
Apply only approved items, one level at a time: story, pace, text, picture, sound. Then review with a scan and one full viewing with sound.

An upload of media, a cloud transcription or an install is a change too, and it waits for approval.

Measured with FFmpeg

A probe makes synthetic media and measures each FFmpeg behaviour the skills rely on. CI runs it on each push and once a week. On FFmpeg 9.0.2 it found:

The render keeps exactly the planned frames, 156 of 156, and the audio lands where the EDL says, within 0.1 ms.
A cut with -ss and -c copy starts at the previous keyframe. A decoder that ignores the MP4 edit list shows a second that was cut.
A join without fades makes a step in the waveform 14 times steeper than the tone itself. Fades of 10 ms remove it.
The concat demuxer joins clips of different sizes without an error, and labels the result with the size of the first clip.
The default Homebrew FFmpeg has no subtitles filter, so it cannot burn captions.

How a recogniser handles your language, loudness normalisation, and meaning, rhythm or taste are not measured. Those claims cite FFmpeg’s documentation, the Netflix and BBC subtitle guidelines and EBU R 128, or they stay with the person who watches the video.

Benchmark

Nine cases, three runs each, with and without the skills. The mean score is 0.837 without the skills and 0.976 with them, and 23 of 27 runs pass each criterion against 10 of 27. The gains are where the model alone lacks knowledge or process: YouTube publishes no loudness target, a scan cannot judge meaning or rhythm, audio sent for recognition leaves the machine, variable frame rate and rotation hide in the files, and a result is checked before it is reported. Three cases score 1.00 without the skills as well.