Marketing / playbook

Stop editing every video by hand.

An open-source pipeline turns a take into transcription, word-timed captions, cuts, and a rendered video — with an approval gate before anything spends.

KKeith Rumjahn·9 min read·Free

THE EDIT PIPELINE · ILLUSTRATIVE

An empty timeline becomes a cut you can review.

The mechanical work runs; the creative call stays yours

THE STARTING POINT

A one-minute take, then an evening spent transcribing, timing captions, trimming pauses, and exporting.

A MORE USEFUL RESULT

A captioned, trimmed, level-corrected cut to review — produced by commands you approved, with the failures reported rather than hidden.

Transcribe: Every word gets a timestamp, locally and offline when you choose that route.
Cut: Measured speech boundaries decide where the dead air goes, instead of guesswork.
Caption: Captions and motion graphics are built from the real per-word timings.
Gate: Design, lint, verification, contrast, and audio checks run before a render is accepted.
Approve: Anything that spends credits quotes the cost first and waits for your explicit approval.

An illustrative description of the project’s documented pipeline, not a screenshot of a finished edit or a claim about how a video will perform. Available steps and costs depend on the current repository and the route you choose.

You can record a one-minute video in a single take and still lose an evening to the edit. The slow part — transcribing, timing captions to each word, cutting the dead air, rendering — is mechanical, and mechanical work is exactly what a pipeline can take. The judgment stays yours.

YOU’LL WALK AWAY WITH

The open-source project, an honest account of what it costs to run, and a starter prompt that sets it up read-only first.

macOS or Windows · Node.js · a coding assistant · FFmpeg · an account for the hosted steps

01 / Automate the repetition, not the judgment

The slow part of short-form editing is rarely the creative decision. It is transcribing the take, timing captions to individual words, finding where speech starts and stops, cutting the pauses, and rendering. Each step is mechanical, and each one can be described as a command.

That is what an agent-driven pipeline does. You point a coding assistant at a command-line tool and it runs the stages in order: transcribe, prepare the beats, build the document, run the checks, render, then lay the audio back on. What it does not do is decide whether the video is any good.

Transcribe the take, with a timestamp for every word.
Measure where speech starts and stops, then cut the dead air.
Build captions and motion graphics from the real word timings.
Render the cut, then mix and normalize the audio.

02 / Know what is free and what is billed

This matters more than the feature list. Transcription can run locally and offline at no cost, or you can use the hosted provider and spend transcription credits. Generating footage and a voice from a script spends two separate allowances. Re-lipsyncing a clip bills a third-party key that you supply yourself.

The project is open source under Apache-2.0, but the render engine it downloads is licensed separately, and the hosted steps need an account login. Treat "open source" as a statement about the source code, not a promise that every step is free. Decide which route you want before you start, so nothing bills you by surprise.

Free and local: offline transcription, FFmpeg work, the gate checks.
Account credits: hosted transcription, generating video and voice from a script.
Your own third-party key: the faster background-removal and lipsync routes.

03 / Let it stop before it spends

The safeguard worth copying is the approval gate on cost. Generating is deliberately two commands rather than one: the first quotes what it will spend and records your approval, the second spends exactly what was approved and nothing more. An abandoned run can be cleared rather than silently re-charged.

There is a second chain of checks before a render is accepted, covering design, lint, verification, contrast, recording, probing, and the audio mux. A render that fails a gate does not quietly ship. This is the honest version of letting the machine clear the repetitive work: the stop button is real, and it sits in front of the expensive steps.

READY TO MAKE IT YOURS?

Get the exact prompt.
Make it your own.

You’ve seen how it works. Unlock the copyable prompt and the remaining steps to try it yourself.

Open the project and copy the setup prompt
Prove one short video before trusting the pipeline

New here? Create a free account