How this works, and where it stops.

Start a project by uploading a file or pasting a link. A link import fetches the video first, so it takes longer to start than an upload of the same length.

The project then walks a fixed sequence of stages, and the stage it is on is shown while you wait: IMPORT probes the file, extracts the audio and transcribes it. CURATE reads the transcript and picks the moments worth keeping, scoring each one. REFINE decides each clip's framing and places any B-roll. RENDER encodes each clip and UPLOAD stores it. COMPLETE means the project is done; a clip that failed on its own says so on its card and can be rendered again.

A project that fails lands on STALLED with the reason attached. The two common ones are a video with no audio track, which leaves nothing to transcribe, and a video with no speech in it at all. When the transcription service is busy, the project waits and retries by itself instead of failing.

Every clip is rated out of 99, built from four parts scored with it rather than from an opinion on its own. Hook, out of 25, is how strongly the first three seconds grab attention. Flow, out of 25, is whether the clip stands alone with a beginning and an end. Value, out of 25, is how entertaining, useful or affecting it is by itself. Trend, out of 24, is how well it matches formats currently working in Nepali short-form.

Open any clip to see the breakdown and the one-sentence reason the model gave for the score it assigned. The reason is worth reading before you disagree with the number: it usually names the specific thing it rewarded.

The score ranks clips against each other within one video. It is not a prediction of views, and nothing in this product measures views, so treat it as an ordering rather than a forecast.

Captions are burnt into the rendered video rather than attached as a track, because no short-form platform displays an uploaded caption track the way you styled it.

Devanagari conjuncts compose correctly both in the editor and in the burnt-in output. This is the part most tools get wrong, and it is worth checking on your own footage early: a font that renders Nepali as separate letters with visible halants will do it consistently, so you will see it on the first clip.

Caption styling is chosen when you start a project, and any clip can be restyled in the editor's Captions tab. A clip keeps the styling it was rendered with until you save a change, which renders it again.

A 16:9 recording becomes 9:16 by cropping to where the speaker is, so no head is cut off. When no face can be found and a crop would lose the picture, the whole frame is fitted over a blurred copy of itself instead.

By default the crop follows the speaker as they move, easing like a camera operator rather than snapping, and re-aims when the source cuts to another camera. Where there is nothing to follow it holds one framing for the whole clip. Following can be turned off for any clip in the editor's Framing tab.

Framing, aspect ratio and resolution are set per clip in the editor. 4K export is limited to plans that include it.

A clip is not a single start and end. It owns a list of time ranges, and the render joins them. That is what lets you cut a stumble or a long pause out of the middle and still get one continuous file rather than two.

Word-level editing works off the transcript: remove words and the matching audio goes with them. Filler-word removal does the same thing in bulk.

Profanity censoring mutes the chosen words, or plays a bleep over them if you would rather the cut be heard as deliberate. The burnt-in captions show those words starred out.

The glossary is a list of names, places and product words handed to the transcriber before it runs. It is for words the model mishears - an unusual name it has no reason to know. Supplying it up front is more reliable than correcting it afterwards.

Replacements are applied after transcription, and are for the opposite case: the model heard correctly and wrote a spelling you do not use. A romanised name where you want Devanagari is the usual example. No hint would fix that, because nothing was misheard.

A replacement rewrites both the transcript and the word timings behind the captions, so the panel and the video agree. Phrases work as well as single words.

You can queue a clip for a specific time, see it on the calendar and edit the copy that goes with it. Per-platform copy is generated separately for each destination, because the caption that works on TikTok is not the one that works on YouTube.

What is not connected is the final step. Publishing to each platform needs an OAuth application registered and reviewed by that platform, which is an approval process rather than a piece of code. Until then a scheduled post that reaches its time is marked failed with that reason, and editing it puts it back in the queue.

The calendar deliberately has no drag-to-reschedule. A drag that rewrites a publish time is too easy to trigger by accident for something that cannot be taken back.

Analytics reports what this product can observe: how scores are distributed, which genres score highest, which facet the model rewards most in your footage, how many clips you get per source minute, and how many you produced per day.

It shows no view counts, likes or watch time, and it says so on the page rather than showing an empty chart. Those numbers live inside each platform's API, behind the same approvals that block publishing.

An API key is displayed exactly once, at the moment you create it, and only its hash is stored. If you lose it, revoke it and make another - there is no way to recover the original, by design.

Webhooks fire when a project completes or fails, when a clip finishes rendering, and when a scheduled post publishes or fails. Each delivery is signed, so your endpoint can verify the payload came from here, and one that fails on a network or server error is retried three more times. The API page shows each webhook's last delivery and its status, so a broken endpoint is visible rather than silent; after ten failures in a row a webhook is switched off.

Sending a test delivery queues it rather than sending it inline, so it can take a few seconds to appear in the list.

Each plan carries a monthly allowance of source minutes: 60 on Free, 150 on Starter, 300 on Pro, 1200 on Business. A project is charged on its real duration, which is only known once the file has been probed, so a link import can be rejected after it has already downloaded.

Several things Opus Clip advertises are not here because the models are not reachable from this deployment, not because they were skipped. Voice cloning and dubbing need a Nepali voice-cloning model, which does not exist. Generative video B-roll needs models this deployment cannot reach, though stock B-roll works. Upscaling needs a dedicated model rather than a language model.

Prompt-to-video has its own page explaining the same thing in more detail.