Feed it a long recording — a match, a tournament stream, a conference day, a four-hour broadcast — and the AI marks the moments that stand on their own, cuts each one out, captions it, and grades it 0-100 so the review is a ranked list instead of a scrub through the whole thing.
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 slow part of highlight work is the watching. That is the part being removed.
Upload the file or paste a URL if the footage is already published. Paid plans handle up to two hours in a single go; the free demo takes anything under thirty minutes, which is enough to sanity-check the output on one period, one session or one talk before you commit a whole event to it.
The model goes through the full runtime rather than sampling it, which is the only way to find the thing that happened at minute 97 while everyone had stopped paying attention. Segments that hold up out of context get marked, and each is cut around a complete thought rather than a fixed number of seconds either side of a timestamp.
Highlights come back captioned and framed, each with a 0-100 grade. Sorting by score turns thirty candidates into a five-minute review. Take the strongest few, adjust the in and out points on anything that needs tightening, and download or schedule them from the same screen.
Long footage has specific problems. These are the ones being solved.
Highlight tools that skim keyframes miss anything that is interesting for what was said rather than for how it looked. Reading the whole transcript and the audio is slower to run and much better at finding the moment nobody thought to note down at the time.
Getting forty candidates out of a six-hour event is only half a result, because forty candidates is still forty things to watch. The 0-100 grade decides the order you open them in, so the review becomes a top-down skim you can stop the moment you have enough for the week. What feeds that number is defined in full on the viral score checker; on this page it is simply the thing that stops a long event producing a second pile of work.
The tell of an automatic highlight is a clip that opens two words into a sentence. Cutting on sentence boundaries means each highlight begins on a complete thought and lands on the end of one, which is what makes it feel chosen rather than sliced at a timestamp.
Event and match audio is rarely clean — crowd noise, a distant mic, someone shouting off camera. Word-level captions synced to the speech keep the highlight legible when the sound is not, and they carry the clip for the large share of viewers watching muted.
Wide event footage is mostly empty space, so a fixed vertical crop tends to land on a table edge or the gap between two chairs. Face tracking moves the crop to whoever is speaking, which is the difference between a usable vertical highlight and one you throw away.
A panel wants a split or a grid, a screen-share session wants the slides readable, gameplay wants the camera in picture-in-picture. Choosing the layout from what is actually on screen prevents the standard failure of jamming every kind of footage into the same vertical template.
Long recordings are full of pauses, restarts and throat-clearing. Removing them tightens every highlight by several seconds, which matters most in this format — a highlight that takes eight seconds to get going is not a highlight.
If the source is a broadcast rather than a file, clips can be produced while it is still running. For a tournament day or a conference livestream that means highlights go out during the thing, not in a recap post nobody clicks. See the livestream clip generator.
The MCP connector lets you ask Claude to pull highlights from a recording and hand them back in the chat. For batch work — a season of matches, a year of webinars — the API is the sane route.
A two-day event produces dozens of hours nobody will rewatch. Highlights are what the marketing team actually needed, and they are needed the same week.
Bracket days generate more usable material than any human can cut in time. Casting audio gives the model plenty to work with, which makes commentated esports one of the strongest cases here.
A long teaching session contains several self-contained explanations. Pulled out individually they become the search-friendly answers people find you with.
Recorded demos and QBRs hold the objection-handling moments worth circulating internally. Highlights make a two-hour call reviewable in six minutes.
One long broadcast reliably contains a handful of passages that travel on their own. Captions do the heavy lifting, since most of that viewing happens with the sound off.
Panel discussion is dense in exactly the way detection likes. The split and grid layouts keep more than one speaker on screen so an exchange still reads vertically.
Highlight work is not hard, it is long. A six-hour recording might hold twelve minutes worth keeping, which means the labour is almost entirely in watching the other five hours and forty-eight minutes to find them. Nobody can do that at a reasonable cost, so in practice most long footage is never highlighted at all — it sits in a drive as an asset that was expensive to capture and produced nothing.
Automating the cut is only marginally useful. Automating the watching is the thing. Once a machine has read the entire runtime and produced a ranked shortlist, a person spends their attention on judgement — is this the moment we want to be known for — rather than on transport controls.
That is also why the score matters more here than on shorter sources. With a twenty-minute video you can review every candidate. With a full event you cannot, so an ordering that is roughly right is worth more than a perfect extraction with no ordering at all.
Detection reads what is said. That makes commentated footage — an esports cast, a broadcast match with a commentary track, a livestream where someone is reacting to the play — a strong case, because the commentator reliably gets loud and specific exactly when something worth clipping happens. The transcript effectively labels the highlights for us.
Footage with no commentary is the weak case, and it would be dishonest to pretend otherwise. A youth game shot from a tripod with only crowd ambience gives the model very little to distinguish a goal from a throw-in. We do not run object tracking on a ball or detect a scoreboard change, so if that is your footage, this is not the right tool and you should not spend a trial finding out.
The line is simple enough to apply before you upload: if a person is describing what is happening, the highlights are findable. If the only signal is on the pitch, they are not.
This produces separate highlight clips, each standing alone, ready to post. It does not assemble them into one montage with a music bed and transitions. That is a deliberate scope choice rather than a missing feature — short-form platforms reward individual moments, and a five-minute compilation performs poorly in a feed built for one idea at a time.
If you do want a traditional highlight reel for a website or an opening slot at an event, the practical route is to take the top-scoring clips out of here and assemble them in whatever editor you already use. The tedious part, finding and cutting them, is the part that has been removed.
The same applies to music. Clips come out with the original audio and burned-in captions, not with a licensed track laid under them. Adding a bed is thirty seconds of work in any editor and keeps the licensing decision where it belongs, with you.
Somebody who sat in the room will beat this on any individual highlight, and it is worth being specific about why. They know which speaker the client is paying attention to, they heard the line that got the laugh, and they remember that the best answer of the day came after a question the microphone never picked up. A model reading a transcript afterwards has none of that context and cannot acquire it.
What the manual route cannot do is keep up with runtime. Reviewing a nine-hour bracket day honestly costs real time plus a second pass, so a hand-cut event is two working days before a single file is exported. Most teams never spend those two days. They cut the keynote, promise the breakouts, and the breakouts quietly never happen — which is why the realistic comparison is not "machine clips versus better human clips" but "machine clips versus no clips".
Most people expect the division of labour to fall the other way round: the machine does the mechanical cutting and the human finds the moments. It inverts. Finding is the part that is mechanical at this scale, and choosing which three of fifteen strong candidates represent the event is the part that needs someone who was standing at the back of the room.
So the arrangement that holds up is sequential rather than either-or. Let the full-runtime pass produce the ranked shortlist, then put your editor on the top fifteen candidates instead of on six hours of transport controls. Trimming a second off the front of a good clip is a thirty-second job; locating that clip was the two days.
If the footage is one talk under twenty minutes and you already know the moment you want, use the trimmer you already own. Running a whole selection pipeline to extract a clip you could scrub to in ninety seconds is not a saving, and we would rather say that than take the upload.
If the event is being streamed and you mainly want reaction moments, the platform clip button is free and immediate. Twitch and Kick both ship one, and a moderator with a hotkey produces perfectly serviceable markers while the show is running. What that route cannot give you is captions, a crop that follows the speaker, or any ordering across a nine-hour day — so in practice the two get used together, with the hotkey covering what a human noticed live and the full-runtime pass covering the eight hours nobody was watching.
For a flagship event with a real budget, a human clipper or an agency is a legitimate answer and often the right one. The economics only turn when the volume does: one keynote a year is a freelancer, twelve sessions a quarter is a pipeline, and the awkward middle is where teams overspend on both.
And you can build a version of this yourself. An open transcription model plus a script that cuts on timestamps gets you a rough shortlist in a weekend. The parts that take much longer than a weekend are the ones nobody budgets for: word-level caption timing, speaker-following crops, and a scoring pass that is consistent enough to sort by. Worth knowing before you start, because the transcript step is the easy 20% that makes the project look finished.
The most expensive mistake is feeding it bad audio. Highlight selection is downstream of the transcript, so a camera mic picking up a hall from thirty feet away produces a mangled transcript and, inevitably, mangled choices. If there is a desk feed, a board mix or a lapel recording available, use that file instead — it changes the output more than any setting.
The second is uploading raw footage with the pre-roll still attached. Twenty minutes of an empty stage while people find their seats is twenty minutes of nothing to score, and on a demo capped at thirty it is most of your test. Trim the top before you submit.
The third is assuming the top-scoring clip is the one to lead with. Ranking is reliable at separating the strong from the weak and much less reliable at separating first place from third. Watch the top few and choose with your own judgement about which one represents the event.
The fourth is running silent footage through it and concluding the tool is broken. It is not broken, it is doing what a speech-driven system does with no speech, and no amount of retrying changes that. Check whether someone is narrating before you upload.
Split the day the way the schedule already split it. One file per session, per match or per talk gives cleaner results than one enormous recording, because each piece gets judged against itself instead of against the whole day, and a strong moment in a quiet session is not buried under a louder one from the keynote.
Run the first session on its own and review it properly before submitting the rest. Ten minutes spent checking whether the framing holds up on your setup — your stage, your camera position, your two-shot — will tell you whether to change layouts before you have processed twelve files that all need the same correction.
Then queue the remaining sessions, come back once, and sort everything by score across the whole event. Pull the top clip from each session rather than the top ten overall, which keeps your coverage broad instead of six clips from the one speaker who happened to be the most quotable.
Schedule the output across the days after the event rather than the hour it ends. Interest in an event lasts longer than the event does, and a clip released on the Thursday reaches the people who were too busy on the Tuesday.
Choose the aspect ratio from how far away the camera was, not from where you intend to post. A lectern shot from the third row crops to vertical perfectly well. The same talk captured from the back of a ballroom does not, and the honest export there is 1:1, which keeps enough of the stage for the speaker to still read as a person, or 16:9 for a recap page or a screen in the lobby. Any clip can be re-rendered into another ratio later without regenerating it, so the ratio is a reversible decision in a way the camera position is not.
Pick the layout from what is on screen rather than applying one choice to the whole event. A two-person exchange wants split-screen, a four-person panel wants the grid, a workshop that lived in its slides wants the screen-share layout so the deck stays legible in a tall frame, and a casted tournament wants picture-in-picture so the play and the reaction are both visible. One event will legitimately use three of these, and the batch that gets a single layout applied across everything is the batch with an unusable third.
Set the caption style before you render the event, not after. Room audio is the norm here and the burned-in text is what carries the clip on a muted phone, so the style that survives contact with a busy hall is the high-contrast one rather than the one that matched your slide deck. If you want a mark on the output, put the event or team logo in the Brand Kit — paid exports come out clean by default, and your own badge is the only one worth carrying on a conference clip.
Render vertical first and treat every other ratio as a second pass. Most event clips only ever need 9:16, and the two or three destined for the recap page or a sponsor reel can be re-rendered once you know which ones they are. Doing it the other way round, every ratio for every clip on the way out, is how a twelve-session event becomes ninety files in which nobody can find anything.
Observations from operating the pipeline in production — not general advice.
Selection does not read your runtime, it reads a transcript that is truncated before it ever reaches the model, so everything in the file competes for a fixed number of characters. Event recordings waste them more than any other source we handle: twenty minutes of an empty stage, a mic check, a compere reading housekeeping off a card. All of that transcribes, all of it occupies the window, and none of it can score. In practice, trimming the top of the file before you submit changes the shortlist more than any option on this page, because the characters you free up are spent reading further into the session instead of into the queue for coffee. The character limit itself is set out on the best moments finder.
Two hours per submission is a limit on an uploaded file, and a tournament day is not a file. A live session works differently in kind: it follows a public broadcast on a window that keeps being extended while the stream runs, and it commits to each clip inside the part of the show that is currently airing rather than from one read of a finished transcript. That is the reason a nine-hour bracket is easier to clip as it happens than it is to clip the next morning, and why we tell event teams to connect the channel before the doors open. The cost is that it is not retrospective: a session you start at 4pm cannot go back for the thing that happened at 11am.
Face tracking moves the crop to whoever is talking, which assumes there is something in the frame worth moving to. A hall camera thirty metres back breaks that assumption: a 9:16 window is roughly a third of the width of a 1080p frame, so when the speaker occupies a tenth of the original shot, every candidate crop contains the same small figure and the choice between them is cosmetic. With nothing it can hold confidently, the framing stays put, because a still crop that is slightly off is far easier to watch than one drifting across a stage hunting for a face. The reason we push event teams to run one session before the other eleven is that this is settled by your camera position rather than by a setting.
Every event team has a version of this: a marker key on the switcher, a runner with a notepad, a channel where people drop timestamps as things happen. It is real signal and worth keeping. Its structural hole is that it only records the minutes when somebody was in that room, paying attention and not currently rescuing a microphone — so it is dense on the keynote and empty on the third breakout, which is where the unexpected thing usually happened. A full-runtime pass covers the sessions nobody was assigned to. The productive arrangement is to use the log as a checklist against the ranked shortlist afterwards: anything your team flagged that the shortlist missed is worth a look, and anything the shortlist found that nobody flagged is the reason you ran it.
Volume-based highlight tools find the moments where something got loud, which correlates with interest and is not the same thing. A quiet, devastating answer registers as nothing. Reading the language finds the substance as well as the noise, and drops the loud moments that turn out to be an alarm going off.
Uploading every session in full is worth doing and most event teams already do it. What the archive cannot do is reach anybody who was not already looking for it: chaptered two-hour recordings serve the delegate who attended and wants the slide again, and nothing about them travels. Highlights are the part aimed at the people who were not in the building, who will never scrub a timeline for the good ten minutes, and who are the audience the event was supposed to grow. The two outputs answer different questions, and only one of them has any reach.
Most are tuned around twenty-to-sixty-minute talking-head video. The differences that show up on long footage are full-runtime review rather than sampling, ranking that survives a large candidate set, and the ability to work on a live event instead of waiting for the recording. Run one event through it and judge on the shortlist.
The best test is the file you have been avoiding. Run it once and see what the shortlist looks like against what you would have found yourself.
⚡ Find My Highlights — $1 trial
