The boring half of video editing, automated
A local creator suite: record, AI-cut, arrange and render video. Raw takes to final cut, without touching a timeline.
Making a video alone means paying four subscriptions and doing the tedious part by hand: cutting silences, finding the retakes, arranging the same lower-thirds again. I wanted one tool for the whole loop, running on my own machine, where the only thing I pay for is API calls, and where the footage never leaves the laptop.
Solo. Design, build, and the brand system the video components read from.
One question per session. I write ideas, star the good ones, reject the bad, and each AI round reads everything already decided.
An episode moves across six columns (idea, research, script, filming, editing, published), so the thing being made has a state.
Camera and, optionally, a chosen screen or window, captured as two synchronised files.
Normalise the footage, transcribe it verbatim with word-level timings, and propose the cuts.
A model proposes the visual edit against a registry of my own components; I fix it by clicking; one composition drives both the preview and the export.
Each room is a route; the desktop window is a thin wrapper, not a second application
One per module, so a new room can't be blocked by another room's schema
Long steps run outside the web server and write progress into a jobs table, so a render survives a reload
Normalises rushes to constant framerate and extracts the audio the transcriber gets
Verbatim transcription with per-word timings; the fillers it keeps are exactly what the cut stage needs
One composition renders both the in-app preview and the final 1080p export
instead of · Handing the whole timeline to a model and letting it decide every cut
Silence detection is arithmetic on word gaps: deterministic, free, and correct by construction. Retake detection is semantic judgement: probabilistic and billed per call. Drawing the boundary there protects the invariant that timing is never inferred by a model, and confines cost and error probability to the one stage that actually needs judgement.
instead of · Asking the model for time ranges directly
A hallucinated timestamp is a plausible float that cuts the wrong frame and fails silently. An invalid word index fails validation before it touches the footage. Keeping the index-to-time mapping in my code turns the worst failure mode, a silent mis-cut, into a loud, recoverable rejection.
instead of · One better-prompted cut pass
Individually correct cuts can still compose into a duplicated phrase or a broken join: per-edit correctness does not guarantee sequence correctness. The audit pass reads only the surviving words, the edit as the viewer will hear it, and keeps the original indices so its corrections land in the same coordinate system. One extra model call buys a check the first pass cannot perform on itself.
instead of · Shipping the working fix I had already written
The fix suppressed the wipe wherever a chapter card already breaks the flow, so the feature only stayed acceptable under a guard rule every future layout change would have to respect. A default that needs a rule to not look wrong carries negative value: it still looked worse than no transition, so one commit later the feature went, and the rule with it.
Atelier is a production suite built for a solo creator. It records camera and screen as two synced files, transcribes them verbatim, proposes cuts with an LLM that reads the transcript, plans the visual edit, and renders the result server-side with Remotion.
It runs entirely on my machine: four rooms (Channel, Brainstorm, Planning, Cutting) over local SQLite databases, wrapped in Electron. No subscription, no upload, and the footage never leaves the laptop.