From an idea
"A 30-second launch video for our cold-brew can."
The agent writes the Brief, makes a creative plan you can approve, and only then asks to spend.
How briefs work →A brief or a reference video in. A checked, finished short film out — casting, voice, word-level captions, motion and mix on one timeline.
The whole production chain: generation, speech timing, composition and a delivery check. Not only the composition step.
npx skills add EmergenAI/video-recut -g
One take
A brief arrives. Words, shots, captions and sound land on one timeline. The cut plays, the stamp lands on its word, and the checks pass before anything ships.
The gap
If you have tried to make a short video from generated material, you have probably hit these. The right-hand column shows what the Skill takes care of.
One set of reference images goes into every request for that character, and the chain of references is tracked.
guides/3-materials/reference-chains.mdEvents are bound to the spoken word, not to a typed second. New words or a new take move them along.
timing/timeline.json · guides/4-compose/timing.mdPictures, captions and motion live in an HTML page. Changing them costs a re-render, which costs only time. Generated material is reused through the manifest.
manifest.json · bibei.mjs statusVoice is cast like a role: chosen once, sampled, then kept for every segment the speaker performs.
guides/3-materials/voices.mdThe final render is checked against the plan, and the agent watches and listens to the encoded file before calling it done.
plan.json · scripts/check.mjsCompared with
| Video generator + editing app | Agent that writes scripts & prompts | Video Recut | |
|---|---|---|---|
| What you get | Clips. You assemble them. | A script and prompts. You generate and cut. | A rendered film that passed its checks, plus the editable sources |
| Where a reveal or cut sits | At a second on the track | Not decided | On the word it answers |
| Cost of a change | Regenerate, or re-edit by hand | Regenerate | Edit the page and re-render. Paid material is reused |
| Who owns the visual system | The model, then your timeline | No one | An HTML composition page you can read and edit |
| Before delivery | Your eyes | Nothing | plan.json checks, then the agent reviews the final file |
Video Recut does not replace the generator. It drives one: images, video, voice and word timing all run on your own Bibei account.
Ways in
"A 30-second launch video for our cold-brew can."
The agent writes the Brief, makes a creative plan you can approve, and only then asks to spend.
How briefs work →"Like this one, but with our product and a different host."
It reads the reference end to end — why each cut, picture and sound is there — then rebuilds that for your words.
How adapting works →"Here are my takes from yesterday."
Each file gets a job: a reference, a performance to keep, a standalone picture, or a sound source.
How footage is used →How it works
Nothing lives only in a chat. Every stage writes its result to disk, so another session — or you — can pick up the production where it stopped.
Worked example
In an editor, the stamp sits at a fixed second, and drifts the moment the line changes. Here the page asks the timeline when the word is spoken, and puts the stamp there.
// composition/index.html var MOMENTS = { soldOut: find("s4", "gone").start }; tl.from("#stamp", { scale: 1.9, opacity: 0, duration: 0.28 }, MOMENTS.soldOut);
Reword the line. The stamp moves with it.
Worked example
Every render except a draft runs the delivery checks against plan.json first. If a check fails, nothing renders. If the finished length is off, the file is set aside as .REJECTED. A film is delivered only with its report.
The checks: the plan exists and is complete, every planned shot is on the page, no shot is longer than 15 seconds, every sound is registered with its source, all speech comes from the script, no reference audio slipped in, and the length matches.
$ node render.mjs render launch renders/launch.mp4ok plan-present: plan.json foundok plan-shape: 4 shots, target 24sok scene-count: all 4 planned shots are realizedok shot-length: every generated shot is at most 15sok audio-registered: 3 audible file(s) registeredok speech-script: 5 speech segment(s) from script.jsonok reference-audio: no audible file matches the reference's sound (1 reference file(s) compared)rendered renders/launch.mp4 (18.4 MB, 30 fps, standard) in 212.6sok duration: film is 24.08s; planned 24sdelivery report: renders/launch.mp4.report.json
Formats
Each format has a playbook: how that kind of video holds attention, and the craft problems that come up in it. Formats can be combined.
One person speaks to camera; their attitude carries the film.
A question, a reaction, a payoff — the relationship drives it.
A stranger with a life of their own, and an answer worth watching.
A board keeps earlier verdicts on view while the next one is argued.
What characters want and do to each other moves the story.
A presenter talks; demos and motion graphics carry the argument.
A voice-over over products, hands, screens and typography.
Install and prepare
Everything else, including the install itself, can be left to your AI agent: it checks what this computer is missing and installs it. Before a large download or anything that changes your system, it tells you what it is and how big, and waits for your OK.
Recommended: let your AI agent install it. Copy the sentence below and send it to the AI agent you use (for example Claude Code, Codex or any agent that can run commands on your computer). If your system asks for confirmation during the install, click Yes.
Please install the Video Recut skill for me: check that Node.js 22 or newer is on this computer and install it if it is missing, then run npx -y skills add EmergenAI/video-recut -g -y and tell me when it is done.
Comfortable with a terminal? Run it yourself. The command needs Node.js, so install it first: download the LTS installer from nodejs.org and click through; then open a new terminal and type node -v. A version of v22 or higher means it is ready. Then run:
npx skills add EmergenAI/video-recut -g
The first time, if it asks whether to install the skills tool (Ok to proceed?), type y and press Enter.
Once it is installed, start a new conversation and you are ready.
Bibei (bibei.cn) is an AI generation platform; Video Recut uses its open platform to generate pictures and sound. Only the items on the left use Bibei credits. Everything else runs on your own computer at no charge.
Generated on Bibei · uses credits
Done on your computer · no credits
npx skills add EmergenAI/video-recut -g