A five-hour session holds maybe eight seconds of something anyone would rewatch. Feed the whole thing in and get back vertical clips with the webcam sitting over the gameplay, captions burned on, silence trimmed, and a grade on each so the good ones surface without you scrubbing the recording.
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
Bring whatever you already have — an upload, a VOD link, or a live channel.
Upload a local OBS or capture-card recording, or paste the URL of a stream VOD or a YouTube upload. Nothing has to be pre-trimmed and nothing has to be transcoded first. Paid plans handle two hours per submission, so a long session goes in as a couple of pieces split wherever a match or a map ends.
Instead of you dragging a playhead across four hours hunting for the round you half remember, it goes through the recording and marks the segments where something happened — the shout, the laugh, the sudden shift in how you were talking. Every candidate is cut so it begins on a full sentence rather than fading up mid-word.
Gameplay with a webcam corner becomes picture-in-picture, with your face placed over the action rather than cropped off the side. A menu or capture-card-only recording with no camera gets a straight fill. Two players on a co-stream can go split-screen. You can override any of it and re-render.
Each clip carries a 0-100 grade, so twenty candidates from one session collapse into a ranked list rather than a folder you have to review twice. Download the top few or push them to TikTok, Reels and Shorts on a schedule that keeps posting on the days you are not playing.
Gaming video breaks in specific ways when it goes vertical. These are the fixes.
The webcam is usually the most valuable rectangle on a gaming stream and it is also the first thing a naive vertical crop deletes, because it sits in a corner. The picture-in-picture layout lifts it out and places it over the gameplay, so a viewer sees the play and your response to it in the same frame.
Game interfaces put critical information at the edges — killfeed top right, minimap top left, ammo and health along the bottom. Zoom-cropping a widescreen game to 9:16 removes about half the width and takes that information with it. Layout selection is about deciding what survives, not just where to point the crop.
The set runs from a plain fill or fit, through split screen and the three-up and four-up grids, to a screenshare treatment and gameplay picture-in-picture. A racing sim, a four-player party game and a solo horror playthrough are three different framing problems, and forcing all of them through the same vertical crop is why so much automated gaming content looks identically bad.
What marks a moment is mostly you: the volume jump, the swearing, the "no way" that follows something going right. The model works from speech and audio, which is why commentated footage clips well and why a silent recording gives it far less to grade.
Gunfire, music and alerts all compete with your mic, and a lot of viewers have the sound off anyway. Word-by-word captions burned into the file keep the clip legible in both cases, in eleven styles so it can match a channel look rather than shouting default yellow at everyone.
Gaming sessions are full of lobby waits, respawns and menu navigation. Silence and filler are stripped from each clip automatically, which on this kind of source recovers a surprising amount — the difference between a clip that lands in four seconds and one a viewer abandons during a loading bar.
Twenty candidate clips from one night is a good problem and still a problem. Grading each on its opening, pacing and peak gives you a running order, so the best clip goes out first instead of third behind two that were merely fine.
Connect a YouTube, Twitch or Kick channel and clips are produced during the stream itself. Details for each platform are on the Twitch clip generator and Kick clip generator pages.
The Brand Kit holds your logo, colours and fonts and stamps them consistently. For a gaming channel this is most of what makes a stranger recognise their third clip from you as being from the same person as the first.
The same engine sits behind a developer API, and an MCP connector lets Claude cut clips for you conversationally. If your recordings already land in a folder or a bucket on a schedule, clipping can hang off that — see the developer docs.
Short bursts of intensity separated by long quiet stretches. Commentary is what marks the difference, and commentary is what the model reads.
The clip is your face, not the game. Picture-in-picture with a large webcam is usually the right call here.
The moment is a single overtake or a last-second goal. Sentence-boundary cutting matters because the callout usually starts before the play does.
Slow footage with occasional disasters. Scoring earns its keep on long sessions where the highlights are genuinely sparse.
Split and grid layouts keep more than one player on screen, which is the whole point when the funny thing is somebody else's reaction.
Volume work where consistency beats craft on any individual clip. The API makes it schedulable rather than nightly manual labour.
Almost every complaint about automated gaming clips traces back to one decision: what to throw away when a 16:9 frame becomes 9:16. That conversion discards roughly half the width, and games put meaningful things in exactly the places that get discarded. A centre crop of a shooter keeps the crosshair, which is convenient, and loses the killfeed that explains what just happened, which is not.
The webcam is the worse loss. It lives in a corner by convention, so a naive crop removes it entirely, and what remains is anonymous gameplay footage that could belong to anyone. Your channel is the person, not the game — a clip of gameplay with no face in it competes against every other clip of that game, on nothing but the play itself.
Choosing between seven layouts rather than applying one crop is what fixes this. A game with a facecam becomes picture-in-picture. A capture-card recording with no camera at all gets a fill or a fit, depending on whether the edges of the HUD matter. Two people become a split. It is less glamorous than the moment detection, and it has a bigger effect on whether the clip is watchable.
Be clear about the mechanism, because it determines whether this tool will work for you. Moments are found through speech and audio: what you said, how loudly, how the energy of the room changed. Commentated gameplay is therefore the ideal input, and the more expressive the commentary, the better the shortlist.
It does not read game state. There is no killfeed parsing, no score tracking, no detection of an ace or a triple kill from the pixels of a scoreboard. If you play in total silence, a great round and a dull one look much the same to it, and you will get a thin set of clips back. That is a real limitation and it is better stated up front than discovered after paying.
The upside of an audio-first approach is that it generalises. A model tuned to one shooter's killfeed is useless on a farming sim. Reactions work in every genre, which is why the same detection handles a horror playthrough and a rocket-league match without anyone teaching it either game.
Starting too early is the biggest one. Eight seconds of walking down a corridor before the fight is eight seconds a scrolling viewer will not give you. The cut should land just before the callout that signals something is about to happen, which is why the front edge is the edit worth making by hand when the automatic one is off.
Second: no face anywhere in the frame. Gameplay-only clips have to beat every other clip of that game on execution alone, and unless the play is genuinely exceptional, they do not. Keeping the webcam in shot is the cheapest possible differentiator.
Third: needing the score to understand the moment. If the clip only makes sense once you know it was match point, either the commentary has to establish that or a title card does. A viewer with no context will not infer stakes from a HUD they cannot read at phone size.
Fourth: unreadable captions over a busy scene. Gameplay is visually noisy, so caption styling that survives a clean talking-head shot can disappear entirely over an explosion. Pick a style with a solid background or a heavy outline for anything with a lot of motion behind it.
Submit the recording the same night, while you can still remember roughly what happened. Reviewing a shortlist is much faster when the session is fresh — you will recognise the good moment from its thumbnail alone, and you can discard three-quarters of the list in under a minute.
Take four or five, not twenty. Post one immediately and schedule the others across the days between sessions, so the account keeps moving even when you do not play. The clips you hold back are not wasted; they are the buffer for the week something goes wrong with your PC.
Vary what you post. Two clutch plays, a funny failure and a moment of straight commentary will outperform four clips of the same thing, because a feed algorithm and a human being both get bored of a channel that only does one trick.
Every month or so, run a session from the archive through. Old gameplay is only old to you — a stranger seeing it for the first time has no idea it happened in April, and most channels have far more unclipped footage than they will record next month.
The comparison people expect is about cut quality, and on any single clip a person wins. An editor who knows the game starts two seconds earlier because they know the callout matters, holds the reaction a beat longer, and picks the take where your voice cracked. Nothing here beats that on one clip.
The comparison that actually decides your week is about search rather than craft. Cutting by hand means watching your own session back, and that is where the hours disappear — not in the trim, in the scrubbing. Say you record a four-hour ranked session and remember four moments from it: locating them costs most of an evening, and realistically you find three.
Automated selection changes the unit of work rather than the ceiling on quality. You read a ranked shortlist instead of a timeline, which takes minutes, and then you hand-edit the two that deserve it. Most streamers who stay with automation land on exactly that hybrid instead of picking a side — the machine does the watching, a person does the last ten per cent on the clips that represent the channel.
One job still belongs entirely to manual editing. A montage — music underneath, cuts landing on the beat, a build across ninety seconds — is a composition, not a selection, and no shortlist produces one. Use the shortlist to gather the raw moments and assemble the montage somewhere else.
Silent play is the honest disqualifier. Detection reads speech and audio, so a no-mic session gives it nothing that separates a clutch from a reload, and it returns a thin list rather than inventing one. It turns out that is the better failure of the two: a tool willing to fill twenty slots from silent footage costs you an hour of review and hands back nothing.
Highly visual, low-talk genres sit in the same bucket. For example, a precision platformer run where the whole achievement is mechanical and the room is quiet reads perfectly to a viewer and is close to invisible to a detector working from a transcript.
The economics change too if your channel ships one heavily produced short a week rather than daily volume. Coverage is worth most to somebody with more footage than attention. At one clip a week you have the attention, and the same effort spent inside an editor buys more.
And if you cannot record a camera at all — a console capture with no webcam, a shared machine, a handheld — you lose the differentiator that makes picture-in-picture worth having in the first place. The tool still runs, but what comes back competes on the play alone, against every other clip of that game, which is a much harder market than competing as yourself.
The reason most gaming channels post so little short-form is not that the footage is bad. It is that reviewing your own five-hour session is miserable work. You remember that something good happened around the middle, you scrub, you overshoot, you watch four minutes of a lobby, and an hour later you have found two of the four things you were looking for.
Handing the review to a model changes the unit of work from hours of scrubbing to a few minutes of reading a shortlist. The clips are already cut, already vertical, already captioned, so the decision in front of you is only which ones to post.
That has a compounding effect on a channel. Posting five clips a week from sessions you were streaming anyway is a completely different growth curve from posting one clip whenever you can face the edit, and the difference has nothing to do with how good you are at the game.
Observations from operating the pipeline in production — not general advice.
The per-chunk corner-cam detector is not reliable enough to be trusted alone — on busy scenes it misses the cam often enough that re-deciding the layout for every clip produced visibly inconsistent nights. One slots stream cut five clips as a correct facecam-over-game split, then a fresh detection missed the cam and the last two rendered as a plain vertical crop, from identical footage. So the decision is now pooled rather than per clip: an identity needs 3 chunk votes with a clear majority before it locks, and 5 consecutive contrary chunks before it is allowed to flip. A single miss cannot change your framing halfway through a session.
The first version of facecam detection scored regions on brightness, motion and texture, which sounds reasonable until you remember what a game screen contains. It flagged sponsor banners, HUD panels and in-game overlays as the webcam, and clips shipped with all three sitting in the picture-in-picture slot. The replacement runs an actual face model over 8 frames sampled across the chunk and clusters the detections: a real cam sits in the same spot in every frame and forms a tight cluster, while characters and NPCs scatter or appear once. If no cluster persists across at least half the sampled frames, the answer is no facecam and it renders a centre crop rather than guessing.
A picture-in-picture render only happens when there is a corner cam, a content screen to put underneath it, a tile of sane dimensions, and a face actually verified inside that tile. Each of those guards is there because a specific clip failed without it. An IRL streamer in a busy room read as a content screen and got split with nothing meaningful in the lower half. Another stream locked onto a top-left "cam" that was two people on a couch behind the streamer. A tile covering most of the frame is not a tile at all. Fail any one test and it renders a clean single crop, because a bad split looks worse than no split.
Instant-replay capture is excellent at what it does: preserving the last thirty seconds when something unexpected happens. It relies on you noticing in time, it saves widescreen files with no captions, and it leaves you a folder of raw clips that still need editing. Capture and publishing are two different jobs.
Those tools can produce a better single clip than any automation, and if you are building a montage with music and beat-matched cuts you should use them. What they do not do is watch your session for you. The step being replaced is the search, and the search is the part that takes the evening.
Someone who knows your game and your sense of humour will beat a model on selection, reliably. The trade is cost per clip and turnaround measured in days, which is awkward for content tied to a patch or a meta that moves weekly. Automation for the daily posts and a person for anything that represents the channel is a workable split.
Most handle a talking-head video well and treat gameplay as an edge case, which shows up immediately in a vertical crop that has thrown your facecam away. Ask specifically about layout handling and about live clipping, then run a session through and look at where the webcam ended up.
Pick a recording where you know something good happened, and check whether the shortlist found it and kept your face in the shot.
⚡ Clip My Gameplay — $1 trial
