oneshot-video

Your AI agent is now a video editor.

Raw footage in: a finished edit, captioned clips for every platform, or a promo from a brief.

Runs locally on your Mac. Hands the timeline back to Premiere Pro, Final Cut Pro and DaVinci Resolve.

_{Made by oneshot-video itself, from the author's own talk: a 24-minute phone recording from the back of a ballroom became the 15-minute edit, the clips and this trailer. Watch it with sound · its source}

> Use oneshot-video on ~/Movies/keynote/

  What should I make?
  ❯ Full edit + clips   the whole talk cut tight, then the best moments as shorts   (recommended: 3 cameras, 52 min)
    Full edit           fillers and dead air out, best angle, titles, chapters, subtitles
    Clips               standalone 9:16 and 16:9 clips with captions and graphics
    From scratch        a promo or opener from a brief, with stock footage and motion graphics

  Hand it back as?  [x] Finished MP4s  [x] Premiere Pro  [ ] Final Cut Pro (beta)  [ ] DaVinci Resolve (beta)

What it makes

Full edits and from-scratch films can come back as a timeline in your editor, on your original media, so you finish the last 10 percent by hand.

_{Big moments from five clips an agent picked, cut, reframed and finished from two 25-minute CC BY talks, with no human input: takeovers, words behind the speaker, count-ups, punch-ins, a sound under every graphic.}

Install

Claude Code (plugin):

claude plugin marketplace add kvrancic/oneshot-video
claude plugin install oneshot-video@oneshot-video

npx skills add kvrancic/oneshot-video

Or clone it and link skills/oneshot-video into ~/.agents/skills, which all of these agents read.

git clone https://github.com/kvrancic/oneshot-video && cd oneshot-video
skills/oneshot-video/scripts/install.sh    # also links the skill into ~/.claude/skills and ~/.agents/skills

The first time you use it, the agent runs doctor.sh and offers to install what is missing: ffmpeg, whisper.cpp, a Python venv, the models (about 1 GB) and Remotion. Heavy files live in ~/.oneshot-video, so updates take seconds.

Then just ask:

Use oneshot-video on ~/Movies/podcast-ep12.mp4. Four clips for TikTok and LinkedIn, and one about the part where we argue about remote work.

Edit my talk in ~/Movies/meetup/ (two cameras, a lavalier and a screen recording). Keep it tight, cut the live demo, give me a Premiere project.

Make a 45-second opener for our conference. The next one is in Lisbon. Here's our logo and a song we licensed.

How it works

flowchart LR
    A[Footage or a brief] --> M{Start menu}
    M -->|clips / full edit| I[Ingest<br/>best picture, best mic,<br/>sync and drift]
    I --> T[Transcribe<br/>Whisper + Parakeet,<br/>every word indexed]
    T --> S[Story pass<br/>the agent reads it all,<br/>picks moments or cuts]
    S --> G{Gates}
    G -->|mid-thought, dangling<br/>reference, stumble| S
    G --> P[Plan JSON<br/>word ranges, layout,<br/>captions, graphics]
    M -->|from scratch| B[Beat sheet<br/>music edit on the bar,<br/>stock from Pexels]
    B --> P
    P --> R[Render<br/>virtual camera, Remotion,<br/>mix to -14 LUFS]
    R --> Q[Look and critique<br/>stills, contact sheets,<br/>critic agents]
    Q -->|defects| P
    Q --> O[MP4s + captions + post copy]
    Q --> X[Premiere / Final Cut /<br/>Resolve timeline]

The agent never types a timestamp. Every cut is a range of word indices, so it lands on a pause or a word boundary, never mid-word, and a re-cut is an edit to a small JSON plan that rebuilds only what changed. Gates reject clips that start on "so" or "and", end mid-sentence, or point at something the viewer never saw ("as I showed you"). Then the agent looks at its own stills and contact sheets, and a fresh critic agent reviews them before anything is rendered at full quality.

Optional network calls: Pexels (stock), Wikimedia Commons (portraits of public figures), GitHub (bug reports). Nothing else leaves the machine except what your agent itself sends to its model.

Hand it back to your editor

If you open a beta export in Final Cut or Resolve, please tell us how it went. That is the fastest way to turn it green.

Made from almost nothing

_{A 54-second conference opener, made in one conversation. Watch it with sound}

This opener was the first real job for from-scratch mode. The brief: announce that next year's conference is in Boston, to be played live in a ballroom. The material: a 6-second phone clip of five students shouting at a statue, a licensed song and a logo.

What the agent did, in one conversation:

• Cut the song on the bar. It tracked the beat (143.5 BPM), spliced three sections on drum transients and pinned every lyric hit to a frame by its vocal onset.
• Wrote the story onto the lyrics. A split-flap board spells the city one letter per clue and locks the B on the sung "Bos-".
• Found the footage. 50 Pexels searches, 28 clips downloaded, 20 used, each checked for the right city (stock "Boston" includes Prague and Philadelphia).
• Built the motion graphics in Remotion: the board, a route map that dives into Massachusetts, the type, the payoff card.
• Reviewed itself. Three critic agents (a film editor, the client, an audience member) scored each version until all three signed off. v1 rendered 62 minutes after the request, before any human note.
• Mastered it and handed it back as Premiere layers: footage on V1, graphics with alpha on V2, music, effects and voice on their own tracks.

Every lesson from that session is now a rule in workflows/from-scratch.md: the unflagged AAC priming that made every hit 43 ms late, why an opener ends on its peak and not a fade, why a spinning board must never show the answer early. Every component is in the kit.

Requirements

• macOS on Apple Silicon (hardware encoding and Apple Vision). Linux support is the most wanted contribution.
• Homebrew, Node 20+, about 3 GB for tools and models, 5 GB free while rendering.
• An agent that runs skills: Claude Code, Codex CLI, Cursor, OpenCode, Gemini CLI or GitHub Copilot.

Speed on an M3 Pro: a 25-minute talk transcribes in under 2 minutes, a 60-second clip renders in about a minute per format, and a whole clips run on a 25-minute talk (transcript, 12 candidates, three clips planned, critiqued, rendered, checked and packaged) took 35 minutes with no human input. A one-minute from-scratch film renders and masters in 3 to 10 minutes.

Found a bug? Want a feature?

Ask your agent to "report a oneshot-video bug". It drafts an issue with your versions and the doctor output, scrubs your paths, user name and emails, shows you the draft and files it only when you say so.

Ideas, new caption presets, Linux support, verified Final Cut imports: see CONTRIBUTING.md. Forking it to match your own channel's look is the intended use: the agent reads the docs on every run, so a changed rule in a reference file changes the next edit.

Made something with it? Show it off. The best ones go in this README.

Credits

Demo footage: "Sheryl Sandberg Gives UC Berkeley Commencement Keynote Speech" by UC Berkeley, via Wikimedia Commons, CC BY 3.0, edited (cut, cropped, captioned). "Keynote" by Ton Roosendaal, Blender Conference 2023, Blender Foundation, CC BY 4.0, edited (cut, cropped, captioned). Stock footage in the Boston opener from Pexels.

Built on whisper.cpp (MIT), Parakeet TDT (CC BY 4.0) via onnx-asr, Silero VAD (MIT), YuNet (MIT), the RNNoise models, FFmpeg and Remotion. Fonts: Inter, Instrument Serif and Geist Mono (SIL Open Font License).

Remotion licence: Remotion is free for individuals and companies of up to three people; larger companies need a Remotion company licence to render with it.

MIT © Karlo Vrancic