CutEZ — Long Video to Short Clips with one click

📡 What It Actually Does

It reads the video before it cuts. Given only a link, it works out what the source is about, picks the stretch worth clipping, decides the configuration, generates, then re-judges what came back — and it will tell you when a part it liked got rejected, instead of quietly shipping you the leftovers.

Built for talking video. Podcasts, interviews, talks, explainers, commentary — sources with continuous speech. It is not aimed at music, silent footage or fast-cut gameplay, and makes no promises there.

Any long video. A YouTube page link goes straight in. Anything else — another platform, a cloud bucket, your own recording — goes in as a direct video file link: Amazon S3, Google Cloud Storage, Azure Blob, Cloudflare R2, Backblaze B2, MinIO, DigitalOcean Spaces, or any public CDN. Source files up to 5 hours.

File links, not integrations. Those storage providers work because they serve direct file links — there is no account connection and no file browser for them.

Understands the source before spending anything. Metadata, duration and captions are read locally. If there are no captions, an optional lightweight speech model can transcribe it. If local analysis fails outright — even reading the source's own metadata — the round continues on a default window. The source is handed on, not abandoned.

Picks a continuous window of 5–20 minutes that stands on its own: it looks for a clear hook, concrete people, events, numbers or conflict, and an ending that resolves or twists. Small talk, setup, ad reads and filler get demoted.

Explains the choice in one sentence you can read. Not a machine score — a real reason, and it goes into the delivery.

Finds several candidates and does not dump a menu of chapters on you to choose from.

One window per round. It does not submit a second window until you say to.

📊 How a Round Works

Pre-analyze — duration, language, timestamped captions. Optional transcription only when needed.

Select — the window, the length, the layout, the caption style, the reason.

Generate — one request, 60–1200 seconds of source.

Review — each returned clip is judged independently: hook, completeness, whether it stands alone, filler, title match. Verdicts are ready, revise or reject.

Deliver — in batches, never one dump of links. A single passing clip is delivered as a single clip.

🧠 It Picks the Window — You Don't Have To

Ranks regions from the timeline, then expands the raw timestamps for the most promising ones before deciding.
Reads a bounded slice of transcript at a time, so a three-hour podcast never blows up its own context.
Defaults to a single strong window when the video is short, and the whole video when the source is under 20 minutes.
If you specify a range yourself, that wins.

💬 Captions: Matched to Style, Then Verified

The caption styles available upstream are listed, one is chosen to fit the content — quotable hard-talk, calm narrative, tutorials and explainers get different treatments — and the chosen style's preview is checked before it is used.

Languages with long caption lines get a style that wraps. Some styles render a single unbreakable line: on a Russian source, measured frame by frame, one style stretched to 717 of 720 pixels and severed glyphs mid-stroke, while a wrapping style cut nothing. Long-line languages are routed around it.

Nothing is called good until it is looked at. Frames are sampled across each clip, cropped to the caption band and read in one pass. Text cut off at the edges or rendered illegibly cannot be delivered as ready. The frame check is not optional for a finished clip — a clip nobody looked at cannot be marked ready.

✅ A Gate Before Anything Is Called Done

Each clip is judged on its own merits, not on the provider's score.

ready requires the caption frame check to have been performed. There is no unchecked path to a finished clip.
Clips that fail review are separated out and their problems are stated — they never get mixed into your finished set.

revise means the problem was named, not fixed. A clip that needs work is reported with what is wrong with it; nothing is silently re-cut or re-captioned and passed off as done.
Downloads are validated — a valid response and a non-empty file, not an assumed success.

📦 What You Receive

Each passing clip with its direct link, a recommended title and a note on why it was kept — plus an opening hook and hashtags where the clip supports them.

A local copy that stays put. CDN links expire; your files are kept locally and the path is given to you.
A scope line telling you which part of the source was covered — and what remains.
Delivered in batches, so a long job arrives in pieces instead of one wall of links.

⏱️ What to Expect

Waiting time. A round usually takes on the order of 10–20 minutes, and longer when the queue is busy or a model has to be downloaded first.

How many clips varies. A 20-minute window may yield several, one, or none that pass review. Five is not a promise — and a window that produces nothing is reported, not hidden.

Nothing runs behind your back. One window per round, and no automatic re-submission after a failure.

🔄 Longer Sources: One Window Per Round

The product handles up to 20 minutes per request — the source can be far longer.
Reply "continue" and it advances to the next ranked window, having re-read that range and re-decided the configuration for it.
It will not advance on its own, and it will not prompt you to continue before the first round is finished.

⚠️ Honest Limits

The upstream service only clips talking video. Its check fires per window, not per source: on a three-person interview podcast, one window was rejected at 40% while an earlier window from the same video produced clips normally. The rejected window scored 5.058 on local analysis, so this cannot be predicted in advance or avoided by choosing differently. When it happens, that window is reported with the provider's own wording and you move to another.

A direct link has to be a real file link. Dropbox and Google Drive share links do not work: they return an HTML preview page rather than the video bytes. Use object storage, a CDN, or a server you control.
A platform page link other than YouTube (Bilibili, Vimeo and so on) is not accepted directly — obtain that video's file link instead. YouTube is the one source whose page link can be handed over as-is.

Source files are capped at 5 hours upstream.
Layout is not fixed to vertical — auto, fill, fit and square are available. The "keep the whole frame" option can letterbox rather than crop to a face.
Caption appearance is limited: there is no font size, line-wrap or position control. The style is the only dial.
A failed or rejected window is always reported. You will not be told a job is complete when part of it failed.

Generated titles and hooks are suggestions, not verdicts — check them before publishing.