Generating clips stopped being the hard part a while ago. Choosing between them did not. Every cut ClipSpeedAI produces arrives carrying a 0-100 grade, and this is the page that spells out in full what that number is made of and what it refuses to claim. A folder of fifteen candidates becomes a running order instead of an evening of second-guessing.
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 grade is not a separate step you run. It arrives attached to the work.
Paste a YouTube URL or upload a file — a podcast episode, a webinar, a stream replay, a talk. The demo covers anything under thirty minutes and paid plans go to two hours per upload. You do not choose moments or set anything up; the point of the exercise is to find out what the engine thinks is in there.
Each cut is scored 0-100. The number reflects the strength of the first few seconds, how the middle holds together, and whether the clip peaks somewhere useful rather than trailing off. Alongside it you get the clip itself, vertical and captioned, so you can immediately check the grade against your own eyes.
Publish the top of the list first, because your posting slots are the scarce resource and they should go to the strongest material. Then do the part almost nobody does: after a fortnight, compare what actually performed against what scored highest. Where your niche diverges from the general pattern, you have learned something no tool could have told you up front.
A single number is only useful if you know what went into it. Here is the honest breakdown.
Retention on short-form is decided almost entirely at the start, so the grade leans hard on how a clip begins. A cut that opens on a specific claim, a question, or a moment of tension scores above one that opens on somebody clearing their throat and easing into context, even when the second clip is more interesting by the twenty-second mark.
The middle is where a clip loses people quietly. Long unbroken sentences, tangents that go nowhere, and stretches with no change in energy all pull the number down. What is being read here is the transcript of the moment, so a stretch that looks flabby in words grades as flabby even when filler removal later tightens how it plays.
A clip whose best line lands in the final two seconds is structurally worse than the same line landing at the two-thirds mark, because the audience has to survive the run-up to reach it. The engine looks at where the strongest beat falls relative to the length of the clip, which frequently explains a score that looks surprising at first glance.
The most reliable thing about the number is the ordering it produces within one video. Clip A scoring 84 against clip B at 61 is a meaningful statement about which to post first. Comparing an 84 from a podcast against an 84 from a gaming VOD is a much weaker comparison, and you should not lean on it.
When clips are being cut from a running YouTube, Twitch or Kick stream, the grade comes with them in real time. During a long broadcast that is the difference between a queue of forty unlabelled fragments and a shortlist you can act on before the stream has ended.
The score is returned by the developer API, so you can filter programmatically — auto-publish everything above a threshold you set, route the middle band to a human for review, discard the rest. Through the MCP connector you can also just ask Claude for the clips above eighty. Details in the API documentation.
Each clip comes with a suggested title, which is a separate lever from the score. A strong clip with a limp caption underperforms a mediocre clip with a sharp one, and the grade knows nothing about the words you type into the upload box. See the AI title generator for how those are written.
The purpose of the number is to reduce a pile to a decision, quickly. If you post three times a week and the engine hands you fifteen candidates, twelve of them are never going out — the grade exists to make sure the right three are the survivors, not to make you feel better about the batch.
Volume is the model, so triage is the whole job. When you are posting several times a day, the cost of manually watching every candidate exceeds the cost of making them.
A grade is something you can put in front of a client. Saying we led with this one because it graded highest is a defensible answer in a way that saying it felt strongest is not.
One episode produces more candidates than the release week can absorb. Ranking them is what stops the same three obvious moments getting posted every time.
A five-hour broadcast generates far too many clips to review by hand. Scores let you skim the top of the list while the memory of the session is still fresh.
Webinars and conference talks are long and unevenly compelling. The grade finds the ninety seconds that justify the recording without you rewatching the whole thing.
If clips are flowing through an automated pipeline, a numeric quality signal is what makes the pipeline safe to leave unattended.
No system can tell you what a feed will do with a specific video. Distribution depends on the account posting it, the audience already attached to it, the hour it went out, what else was competing that day, and a ranking model nobody outside the platform has seen. Anyone selling a number that claims otherwise is selling confidence, not information.
What is genuinely knowable is comparative. Two clips from the same source, made for the same audience, differ in measurable ways: one opens on a hook and one opens on a preamble; one runs tight and one sags in the middle; one peaks early and one peaks after the audience has already left. Those differences are legible, and a score built on them ranks reliably even though it cannot forecast.
That is why the practical instruction on this page is always sort, never trust. The grade is a strong opinion about order. Your judgement about your own audience is the thing that turns an order into a schedule, and it should overrule the number whenever the two disagree for a reason you can name.
Three structural properties of the cut feed the grade. Every other page on this site that mentions a 0-100 number points back here rather than paraphrasing, because a definition that gets reworded on twenty pages stops being a definition. This is the full version.
The opening is the first few seconds, judged on how quickly the clip gives the viewer a reason to stay. What lifts it: starting on a claim, a number, a disagreement, or a question the viewer wants answered. What drops it: a greeting, a name-check, a windup, a sentence that only makes sense if you heard the previous one. Clips are cut along sentence boundaries rather than at arbitrary timestamps, so a weak mark on this axis is a judgement about the sentence you open on and not a complaint about a clip that starts mid-word.
Pacing is the middle, judged on whether the clip keeps moving. What lifts it: short sentences, a change of direction, a second idea arriving before the first has gone stale. What drops it: a single long explanation delivered at one energy, a tangent that resolves nothing, repetition of a point already made. This axis is why two clips of the same length covering the same argument can land far apart.
Peak placement is where the strongest beat falls relative to the length of the clip. A punchline at seventy per cent of the way through is worth more than the same punchline in the last two seconds, because everything after the peak is the viewer deciding whether to keep watching something that has already paid out — and everything before it is a debt they have to be willing to carry.
There is no published weighting between the three, and we are not going to invent one, because the number is not assembled from three sub-scores that get added up. The analysis pass reads the transcript of the moment and returns one figure for the clip. What we can tell you honestly is what pushes that figure in each direction, which is the list above.
It does not know your niche. A slow, quiet, technical explanation can perform superbly for an audience that came for exactly that, and it will not grade like a punchy reaction clip. If you work in a deliberate format, expect your good clips to sit lower than the general pattern and calibrate against your own history rather than against the raw number.
It does not know what you write in the caption, or which cover frame you choose. Those are decided after the clip exists and they materially change the outcome. A 90 posted with a lazy caption routinely loses to a 70 posted with a sharp one.
It does not know about timing or novelty. A clip about something that broke this morning has an advantage that no analysis of the footage can detect, which is precisely why live clipping and scoring work well together — the grade tells you which moment from the stream is strongest, and you supply the knowledge that the moment is currently being talked about.
And it does not know you. A creator with a distinctive voice will find their own reliable format scores in a narrow band, because the engine is judging structure and your structure is consistent. In that situation the useful signal is variance within your own clips, not the absolute value.
If you publish once a fortnight and already know which moment you want, ranking is solving a problem you do not have. Consider the case of a channel that posts one clip a week from one recording: the shortlist is short, you were in the room when it happened, and the number is confirming a decision you had already made. Grading earns its keep when there are more candidates than there are slots.
It is the wrong instrument for footage whose value is not in the talk. Say you cut a two-hour DJ set into drops, or a gym session into lifts — the analysis reads the transcript of the moment, and a moment with almost no speech in it gives that reading very little to hold. The grades cluster, the ordering they produce is close to arbitrary, and your eye is the better tool.
It is also the wrong instrument for comparing your account against somebody else's, or this quarter against last. The engine gets improved, so figures shift between versions, and nothing here is calibrated against a shared external scale. A 78 is a statement about the clips it was cut alongside, and it stops meaning anything the moment you lift it out of that batch.
And it is the wrong instrument if what you actually want is permission. The biggest mistake we see is somebody sitting on a clip they believe in because it graded in the fifties. Post it. A figure computed from the transcript of thirty seconds cannot know that your audience asked about exactly this last Thursday.
Watching every candidate yourself is the accurate method and nobody sensible disputes that. The obstacle is arithmetic. Imagine a three-hour broadcast that produces forty candidates: at thirty seconds of attention each, plus the cost of switching between them, most of an hour disappears before a single decision has been made — and the last ten decisions come out worse than the first ten, because by then you are tired and everything sounds the same.
Ranking inverts the order of that work. You watch the top eight properly rather than forty badly, and the attention you were going to spend anyway lands on the clips that have a genuine chance of being published. In practice that is the whole benefit: not that the machine judges better than you do, but that it judges everything cheaply, so your judgement gets spent where it changes an outcome.
Where a person still wins outright is context. You know your guest is about to appear somewhere large, that a particular phrase is currently a running joke in your comments, and that the clip which graded highest is a point you already made last month. None of that exists in the transcript. Here's why the division of labour holds anyway: those are override decisions, and an override is cheap when the default underneath it was already good.
For example, a habit that works on a weekly show is to accept the top three without argument, promote one clip out of the middle band because you know something the engine does not, and never open the bottom third at all. That is five minutes of work, and it produces a better week than an hour of second-guessing does.
Take the batch and split it in three. The top third goes in the queue this week. The middle third gets a second look, because that band is where a slow opening rather than a weak moment is usually what put the clip there. The bottom third is where you should be ruthless, because posting weak clips costs you more than not posting.
One thing to be clear about before you start trimming. The score is computed once when the clip is first cut and does not change when you re-render, so treat it as a verdict on the original selection rather than a live readout while you tighten. Trimming still helps the clip — it just helps it in front of an audience, not in front of the grader. Judge the re-cut version with your own eyes and post it on that basis.
Then build the habit that makes all of this compound. Log the score of everything you post and the result it got. After thirty or forty clips you will see where your audience diverges from the general model, and from that point on you are running the tool with a correction factor that is specific to you. That is worth considerably more than any single grade.
The first is treating the grade as a verdict on the moment rather than on the cut. A brilliant answer wrapped in a badly chosen thirty seconds scores like a mediocre one, because the analysis sees the clip you built and not the conversation behind it. When something you know is strong grades poorly, the disagreement is nearly always about the thirty seconds you framed rather than about the moment inside them — rebuild the cut and back your own read of it, rather than waiting for a stored number to change its mind.
The second is chasing the number. If you start selecting only for the structural qualities that score well, you drift towards a single shape of clip — cold open, quick escalation, punch, done. That shape works, which is exactly why everyone is already using it. A feed full of identical rhythm is its own problem, and no scoring engine will warn you about it.
The third is deleting the bottom of the batch without watching it. Low grades cluster around slow openings, and slow openings are the cheapest thing in the world to fix. Roughly speaking, the clips worth rescuing are the ones you remember from the recording and the engine did not.
The fourth is applying the score across formats. A talking-head clip and a montage of gameplay are not competing on the same axes, and reading one grade against the other will lead you to post the wrong thing. Compare like with like, inside a single source, and the number behaves.
The first myth is that the number responds to your editing. It does not. The grade is worked out once, at the moment the clip is created, and it is stored on the clip from then on. Trim the front, restyle the captions, re-render, export again — every one of those reads the stored figure and none of them replaces it. We had it written the other way round here for a while, which was wrong, and the version you are reading now is the correction rather than a hedge.
The second myth is that any two grades can be held up against each other. Clips cut from an uploaded video and clips cut from a running stream are graded by different machinery, for the reason set out in the observations below, and a stream clip at 88 is not making the same statement as a podcast clip at 88. Even within uploads, the safe comparison is between clips from one source video.
The third myth is that the number is a second opinion on the clip. It is not independent of the cut: the same analysis pass that decides where the clip should start and stop is the one that returns the figure, so the grader and the editor are the same judgement wearing two hats. That is worth knowing when a clip you like scores badly. You are not overruling an outside referee, you are disagreeing with the same reading that chose those boundaries in the first place.
Observations from operating the pipeline in production — not general advice.
An uploaded video is graded by the analysis pass that also picks the clip boundaries: one judgement, made by reading the transcript. A live stream cannot wait for that, so a clip cut from a broadcast is scored at the moment of the cut from signals that exist right then — how hard chat spiked, how loudly the room reacted in the audio, how many scene changes landed in the window, whether money came in. Both produce a 0-100 figure and they are genuinely different instruments, which is the honest reason we keep repeating that comparisons belong inside one batch.
The live scorer gives an explicit bonus when 2 independent signals fire on the same moment, and another when a 3rd joins them. So a window where chat spikes and nothing else happens will score under a window where chat spikes, the audio jumps and the shot cuts — even when the chat spike itself was smaller. That is deliberate rather than accidental: in practice one hot signal on its own is frequently a single viewer spamming, whereas three unrelated signals landing together is a room reacting to something real.
The first live scorer was purely additive, which behaved fine until we pointed it at high-energy streams: roughly 8 in 10 clips came back at the maximum and the ranking stopped carrying information, which is the worst thing a ranking can do. The fix was a soft knee — the lower and middle of the range passes through untouched, and everything above that point compresses with diminishing returns so the strong clips spread across the top instead of stacking on one value. Uniform scores are the symptom to watch for in any scoring system.
The grade is stored on the clip at creation and every later route reads it rather than writing it. We are spelling that out because the alternative is worse for everyone: a creator trims two seconds off the front, re-renders, sees an identical number and quite reasonably concludes the scoring is decorative. It is not decorative, it is simply a record of the cut the engine originally chose. If we ever make it recompute after an edit, that will be a shipped change announced on this page, not a quiet one.
Your instinct about your own material is genuinely valuable and also systematically biased — you are too close to it, you remember the context around each moment, and you tend to favour whatever you enjoyed recording. A second opinion that has never met you is useful precisely because it is uninformed about everything except the structure of the clip.
This is a real strategy and it is not stupid, but it has a cost. Weak posts on an account are not free, they consume the attention of people who already follow you, and they take the slot a better clip would have used. Scoring is a filter applied before that cost is incurred rather than after.
Analytics tell you the truth, which no predictive score can, and you should absolutely be reading them. The limitation is that they arrive after the decision has been made. The two work together: the score chooses what to post, the analytics tell you how the score is performing for your specific audience.
Several tools display a number. The questions worth asking about any of them are what it is made from, whether it is being sold as a prediction, and whether you can act on it programmatically. This one is built on hook, pacing and peak placement, is presented as a ranking instrument, and is exposed through the API. Run a video through and check the grades against clips you already know performed.
Run a video whose clips you have already published. Comparing the grades against what actually happened is the only test of a scoring engine that means anything.
⚡ Score My Clips — $1 trial
