Upload the recording of a webinar, a customer call, an all-hands or a course session. The AI reads the whole thing, pulls the exchanges that stand on their own, crops to whoever is talking, burns in captions for silent autoplay, and scores each clip 0-100 so you are not guessing which one to send.
Real ClipSpeedAI output — every clip below was scored, cut, reframed to 9:16 and captioned automatically from the video above it.
Source video
15:27
The whole job is deciding what to keep. That is the part being automated.
Local recordings already sit in a folder on your machine as an MP4 — that file is what you want. Cloud recordings need downloading from the Zoom web portal first, since a share link is a player page rather than a video file. Upload it here, or if the session was published to YouTube, paste that link instead and skip the download.
It works through the transcript and audio looking for exchanges that survive on their own: a clean answer to a question from the chat, the objection a prospect raised, the two minutes where a concept finally landed. Most of a call is scheduling, small talk and someone finding the right tab, and none of that competes for a place in the shortlist.
Clips come back cropped to the speaker, captioned and graded. Pick the ones that are genuinely useful, rewrite the titles in your own language, then download them for Slack, a follow-up email or a knowledge base — or publish straight to TikTok, Reels and Shorts if the recording was a public webinar.
Call recordings fail in their own particular ways. These are the parts aimed at those failures rather than at podcasts.
A Zoom frame is mostly furniture, ceiling and a slightly off-centre head. Face tracking crops in on the actual speaker so a vertical clip is a person rather than a wide shot of a spare room, and it follows them when they lean out to reach for something, which everyone does.
Half of every meeting recording is a deck or a product demo with a small speaker tile in the corner. The screenshare layout stacks the shared content above the speaker instead of cropping to one and discarding the other, which is what makes a demo clip legible on a phone.
Waiting for the last person to join, someone unmuting, a pause while a file loads, and the ums that meetings are made of are all removed automatically. On call audio this is the single biggest quality lever — a two-minute answer often tightens to eighty seconds without losing a word of substance.
Internal clips get watched with the sound off in an open-plan office, and public ones autoplay muted in a feed. Word-by-word captions burned into the frame handle both. Eleven styles are available, and the plainest ones are usually right for business content.
A ninety-minute session might yield fourteen candidate clips, and nobody in your team wants to review fourteen. The grade ranks them on how they open and how they pace, so you look at the top four and ignore the rest without watching the source again.
Cuts land on sentence boundaries. That matters more for business content than it does elsewhere, because a clip you are sending to a customer or a board member cannot open on the second half of a word without looking careless.
Three and four-way conversations land on the 3-up arrangement so the people who are talking are all visible, and 4-up is one click away when a fourth participant genuinely earns a pane. There is a limit to this and it is worth stating plainly: a twelve-tile gallery view will never make a good vertical clip, whatever tool you use.
Webinar and demo clips usually need to look like they came from the company rather than from a laptop. Fonts, colours and a logo apply across a whole batch, so twenty clips from one session are visually consistent.
If you run a weekly webinar the same steps happen every week, which is exactly what an API is for. There is a developer API and an MCP connector so Claude can process a recording conversationally — both are in the API documentation.
One hour-long webinar becomes a month of promotion for the next one. The registration page for the following session is far easier to sell with three clips than with a replay link.
The moment a customer explains why they bought is worth more than any case study you could write. Clip it from the call and use it, with their permission.
A live teaching session contains several self-contained explanations that work as standalone lessons or as free samples of the paid programme.
Nobody watches a fifty-minute all-hands recording. A ninety-second clip of the one announcement that affects them gets opened, which is the entire difference.
Plenty of shows are recorded on video calls. If that is you, the podcast clip generator covers the show-specific side of the workflow.
A recorded walkthrough answering a common question turns into a help-centre clip, which is cheaper to make and easier to keep current than a written article with screenshots.
Two choices you make before the call ever starts determine most of the quality of the clips afterwards, and almost nobody is told about them. The first is view mode. If your recording captures active-speaker view, the person talking fills the frame and vertical cropping is straightforward. If it captures gallery view, you get a grid of small tiles, and there is no crop of a twelve-tile grid that reads well on a phone.
The second is whether you record participants to separate files. Zoom can do this, and if it does you end up with several videos rather than one. That is excellent for editing and useless as a direct input here, because the pipeline expects one finished file. Either record the composite as well, or combine them before uploading.
Audio is the third thing and the one people underrate. Laptop microphones in echoey rooms produce transcripts with holes in them, and a caption with a hole in it is worse than no caption. Anyone likely to be clipped should be on a headset. This is cheap advice and it improves the output more than any setting inside the tool.
Almost every team records their webinars and almost none of them do anything with the recordings. The replay link goes in a follow-up email, gets a handful of clicks, and the file lives out its life in a cloud folder. It is not laziness. Watching back an hour to find the four good minutes is a genuinely unpleasant task, and it competes with work that has a deadline.
So the recording is not really an asset until somebody has cut it up, and the cost of cutting it up is why it never becomes one. Removing that cost changes the calculation for the whole format: if a webinar reliably produces eight usable clips for a few minutes of review, running a webinar is worth substantially more than it was.
The scoring is what makes review fast enough to actually happen. You are not watching fourteen clips to find four. You are watching the top four and either approving them or not, which is a five-minute task somebody will genuinely do on a Friday afternoon.
Clipping the wrong thing is the biggest one. Teams instinctively reach for the part where the presenter states the value proposition, and that part is always the least interesting ninety seconds of the session. The material that travels is the unscripted answer, the objection somebody raised, the aside that was not on the slide.
Including the question but not the answer, or the answer without the question, breaks a surprising number of otherwise good clips. Meeting speech is conversational, so a self-contained moment often spans two speakers. Check that the clip makes sense to someone who was not on the call, because that is the only audience it has.
Posting a clip that is visually a slide is a third trap. If the frame is static for ninety seconds it will not hold anyone regardless of how good the words are. Prefer the parts where a person is on camera, and use the screen-share layout when the deck genuinely matters to the point being made.
And then there is the internal-jargon problem. A clip full of product codenames and acronyms reads as competent to your team and as gibberish to everyone else. If you cannot explain the clip to a stranger in one line of title text, it belongs in the internal library rather than on a public feed.
Set the recording up correctly once and it stops being a decision. Active-speaker view, everyone on headsets, one composite file, saved somewhere the marketing person can actually reach without asking. Most of the quality problems teams blame on software are decided in this step.
Run the recording through clipping the same afternoon, while the session is still fresh enough for somebody to sanity-check the shortlist quickly. Approve four clips maximum. The temptation is to approve twelve because they exist, and twelve unposted clips is the same as none.
Use them in three places rather than one: the follow-up email to people who registered and did not attend, the promotion for the next session, and the sales team, who will use a good customer clip more readily than any piece of collateral you write for them.
Then close the loop by keeping a shared folder of clips that got used. Over a couple of months it tells you what kind of moment your audience responds to, and it quietly changes what your presenters do on the next call, which is a bigger win than any individual clip.
Four symptoms account for almost every disappointing first run on meeting footage, and none of them are fixed by changing a setting inside the tool. Work down them in order, because the earlier ones make the later ones impossible to judge honestly.
Every clip is a wide shot of an empty room. This is decided at the recording, not at the crop. A gallery-view capture hands the reframer a grid of tiles rather than a person, so there is nothing in the frame for it to move toward and it holds still — which is the correct behaviour and still the wrong-looking clip. Re-record in active-speaker view where you can. Where you cannot, switch the affected clips to the screenshare or grid arrangements yourself instead of expecting a vertical crop to rescue a tiled source.
The captions have holes in them. Transcription is doing what the audio allows and no more. A laptop microphone in a hard-surfaced room produces word-level output with gaps in it, and those gaps do not stay in the captions. Here's why that costs more than it looks: the same transcript is what the moment picker reads when it decides which ninety seconds of a call are worth keeping, so a passage the transcriber half-heard is a passage that cannot be selected at all. Bad audio is a selection problem wearing a captions costume.
Everything in the shortlist came from the first twenty minutes. The transcript handed to the picker is capped in length, and a genuinely talkative two-hour session can run past that ceiling before the closing stretch is ever read. Splitting a long session into two uploads is the fix, and it puts the second half in front of the picker on equal terms rather than leaving it invisible.
A clip stops two words into the next sentence. That is the boundary path failing open rather than cutting at random. When no run of complete thoughts fits inside the length the clip is aiming for, the originally chosen window is kept rather than distorted, and occasionally that window ends untidily. Pull the out-point back a second in the editor and re-render — the transcript already exists, so a re-render is a cheap operation rather than a second pass over the call.
The comparison people make is editing against automation, and that is the wrong axis. Once you know where a moment sits, cutting ninety seconds out of a recording and captioning it is a modest job that a competent person finishes quickly. Finding the moment inside a ninety-minute call is the expensive part, and it is expensive in a particularly unhelpful way: it is dull, it cannot be done while distracted, and it produces nothing until it is finished.
That asymmetry is why automating the search changes team behaviour and automating the edit does not. In practice the value shows up as the number of recordings that get touched at all, rather than as a better version of the clips somebody was already making. A marketer who reviews eight ranked candidates will do it every week. The same marketer will not scrub eight recordings a week, and after a fortnight of pretending otherwise the whole programme stops.
What a person still does better is everything downstream of selection. Judging whether a customer would be happy to see themselves on a public feed. Noticing that a clip is technically excellent and full of internal codenames. Deciding that the best moment of the call is the one you should not publish. Most people assume automation is weakest at taste and strongest at accuracy; on meeting footage it is the other way round, and the review step exists precisely because the machine has no opinion about what is appropriate to send.
The honest hybrid, then, is machine search plus human veto. Let the pass read the whole recording and rank what it found, then spend your time on the four decisions at the top of the list rather than on the ninety minutes underneath them.
Clipping is one answer to "we recorded something useful and did nothing with it", and it is not always the best one. A meeting-notes tool that produces a summary, decisions and action items beats a clip comprehensively when the output needs to be searchable or acted on — nobody has ever found a decision faster by scrubbing video. If the recording exists to inform people rather than to persuade them, write it up and skip this entirely.
The highlight or snippet export built into most call platforms is the closest neighbour. It is a trimming tool: you supply the timestamps and it hands back a landscape file with no captions. That is genuinely enough for an internal Slack post, and it is free. It stops being enough the moment the clip is going somewhere it has to survive silent autoplay next to professional video.
Re-recording is the underrated alternative. If the thing you want is one clean sixty-second explanation, asking the presenter to record exactly that after the call often beats extracting an approximation of it from ninety minutes of live speech. The trap is treating the archive as the only source of truth when a purpose-built take would be shorter to produce and better to watch.
And there are adjacent versions of this job with their own pages, because a public webinar and an internal all-hands are not the same problem: the webinar clip generator is written for the marketing case, while a recorded panel or fireside is closer to the interview clip generator. If you only need to know what was said, the video summarizer is the honest recommendation over any of them.
A meeting recording is not the same kind of material as a podcast episode. The people in it agreed to be on a call, which is not the same as agreeing to appear on TikTok. For anything with a customer, a candidate or a colleague in it, ask before publishing outward — it takes one message and it avoids a genuinely bad afternoon.
Some jurisdictions require all parties to consent to a recording in the first place, and plenty of companies have internal policy about what leaves the building. None of that is something a clipping tool can decide for you, and any page that skips over it is not being straight with you.
Internal use is the easy case and it is where most of the value is anyway. Clipping an all-hands for the people who missed it, or a training session for a new starter, raises none of these questions. How your uploaded files are handled is set out in the privacy policy.
Observations from operating the pipeline in production — not general advice.
The transcriber splits speech into utterances on pauses rather than on sentences. On one source we measured, only 37 of 62 utterances actually ended in terminal punctuation, and meeting audio is the worst case for this of anything we handle — people stop mid-clause to think, to let somebody else in, or to find the right tab. So utterances are merged forward until a thought genuinely closes, with a sixty-second bail-out for a stretch where no terminator ever arrives. One consequence is easy to miss: a trailing run of words that never closes is discarded rather than used, which is why the last thing said before somebody hits stop almost never becomes the end of a clip.
The file you watch back in your library plays the raw timing of the call. Silence removal is applied at export, at a 700-millisecond threshold, and a source of sixty seconds or under skips it altogether on the grounds that anything that short was already cut tight by a human. The split exists because running it earlier produced two different timelines for one clip — the preview and the editor disagreed about where words sat, and burned captions drifted against the audio a little more at every cut point. In practice, on a recording where 700-millisecond gaps are constant, this means the exported file is noticeably tighter than the version you approved, and that direction is deliberate.
The recommender counts distinct faces and, past two, stops discriminating: three people, five people or nine all resolve to the 3-up arrangement at 0.93 confidence. Four-up exists and is never chosen automatically, because a fourth pane takes width away from the other three inside a 1080-pixel frame and nothing in the signal tells the recommender whether the fourth participant said anything worth showing. The reason this matters on a call rather than on a podcast is that meeting attendance is arbitrary — the number of people in the recording has almost no relationship to the number who spoke. A four-way conversation is one override away from the layout you wanted, and a twelve-tile gallery is the point at which no override helps.
A replay link asks for an hour of somebody's time and gets a click-through rate to match. A ninety-second clip asks for ninety seconds and answers one question. For follow-up after a webinar, the clip converts and the replay does not, which is not a subtle difference.
Several call platforms now let you mark a highlight or export a snippet. Those are trimming tools: you still have to know where the good part was, and the output is landscape with no captions. The finding, the vertical reframe and the captioning are the labour, and that is what is missing.
An editor will do a better job on any single clip. For a weekly series the arithmetic stops working — you are paying per clip on content that mostly serves an internal audience, and the turnaround puts the clip out after the moment it referred to has passed.
A written summary is genuinely better for decisions and action items, and plenty of tools do it well. It is worse for anything where the delivery is the point: a customer saying it in their own words, a founder being unexpectedly candid, a teacher explaining something clearly. Use notes for the record and clips for the moments.
Pick a webinar nobody rewatched. See how much of it was actually worth sending to someone.
⚡ Get A.I Clips — $1 trial
