How Sharpy decides where to cut

What does "cutting on noise" mean?

Most automatic cleanup works on the sound wave. The tool looks for stretches quieter than a set level and longer than a set length, then removes them. It is fast and predictable, and it has no idea what anyone is saying.

Here is the ten-second raw take from the demo on Sharpy's home page:

So the, um, the thing nobody tells you about editing is… the thing nobody tells you about editing is that the cut is a decision. (a 1.80-second pause) And every decision needs a reason.

A silence threshold of 0.3 seconds does this to it:

#	Source in	Event	Why
001	00:00:00:13	CUT	silence 0.38 s
002	00:00:00:22	KEEP	the “um” and the false start: they’re loud
003	00:00:01:08	CUT	silence 0.34 s
004	00:00:06:11	CUT	silence 0.32 s, in the middle of a thought
005	00:00:07:11	CUT	silence 1.80 s

A silence threshold of 0.3 seconds, run over the raw take.

The take is shorter, and worse. The stumble and the false start are still in, because they are loud. A breath has gone from the middle of a sentence. And the pause before the last line is gone too, so the line that should land arrives with no beat before it.

What does "cutting on meaning" mean?

Cutting on meaning starts from what is being said. Sharpy reads the same take the way an editor would: a stumble at the start, a sentence begun and then begun again, a breath inside a thought, an idea that ends, and a long pause before the closing line. Then it decides:

#	Source in	Event	Why
001	00:00:00:07	CUT	stumble: “the, um”
002	00:00:01:16	CUT	false start, the cleaner take is kept
003	00:00:06:11	KEEP	the breath inside the thought
004	00:00:06:18	MARK	the idea ends here, so the next cut can land
005	00:00:07:11	TRIM	dead air 1.80 s → 0.42 s, the low end of this speaker’s pauses

The same take, cut by Sharpy. Every event names its reason.

Side by side, the difference is plain:

	Raw take	Threshold cut	Sharpy
Runtime	10.68 s	7.84 s	5.69 s
The stumble, "the, um"	In	In	Cut
The false start	In	In	Cut, the cleaner take kept
The breath inside the thought	In	Cut	Kept
The pause before the last line	1.80 s	Gone	0.42 s
Reasons written down	None	"Silence"	One for every decision

Sharpy's cut is two seconds shorter than the threshold cut, and it is the one that sounds like a person talking.

What does Sharpy look at before it cuts?

Everything a good editor would, at once:

The words and their timing. Two transcription engines run on your Mac and vote on every word, and every word keeps its timing, so a cut falls between two words rather than inside one.
The delivery. False starts, restarts, repeated words, stumbles, filler words, laughter and where the stress falls in a sentence.
The pauses, and what each one is for. Sharpy labels every gap: a sentence running on, a hesitation, a breath, emphasis, a reveal or a change of topic. It learns the rhythm of the recording at each kind of place in a sentence, so a trimmed pause still sounds like the person, and a pause for emphasis is left alone.
The picture. Who is on screen, how they are framed, whether their eyes are open, and any text in the shot. It reads on-screen text with 2.82% of it wrong, where Apple's own text reader gets 11.00% wrong.
The shape of the piece. Where an idea ends, and what the format needs, from a 60-second vertical clip to a long interview.

It watches 1.7 times faster than real time on the base Mac mini M4, and it answered 18 of 18 held-out questions about the footage correctly. Every one of those numbers was measured on that Mac.

A man in a linen shirt looks away from the camera, gathering a thought between two sentences. — The pause before the next sentence is part of how someone speaks. Sharpy trims it to their rhythm instead of cutting it out.

How does it choose between two takes?

When a line is read more than once, Sharpy lines the takes up and scores each one: how closely it follows the script, how fluent and clear it is, how it sounds, how it is framed and whether the speaker's eyes are open. The best read goes into the edit. The others stay in the edit decision list, with the reason they lost.

A restart is simpler. When a sentence is begun and then begun again within a couple of seconds, the second attempt is almost always the one the speaker meant, so Sharpy keeps it and cuts the first.

Why does every cut need a reason?

Because an edit you did not make by hand is only useful if you can check it. Editors have used edit decision lists for decades to record every edit and move it between systems. Sharpy's adds the column they never had: why.

The reasons do three jobs. They let you review an edit in a minute instead of scrubbing through it. They let you disagree with one decision without redoing the rest. And they keep Sharpy honest: before it renders, it checks that every event in the list has a reason. If one does not, there is no render.

Can you change a decision?

Yes. The edit decision list is yours to read and to argue with. Tell Sharpy to keep the second take of the intro, or to leave a pause alone, and it plans the edit again and renders it. You can do the same through your own AI agent: Sharpy works with Claude, Codex or any MCP client.

Before it renders, Sharpy also asks you the few things only you can decide: the finished size and quality, whether to mirror the picture, and how firmly to deal with dead air and stumbles. The decisions are Sharpy's to make and yours to overrule.

To see where this fits in the bigger picture, read what an agentic video editor is.

Frequently asked questions

Does Sharpy remove every "um"?

No. It cuts the filler sounds that break a sentence, such as um, uh and erm. It keeps "mm-hmm" and "uh-huh", which usually mean yes, and it lists phrases like "you know" and "I mean" for you to check rather than cutting them, because they often carry meaning. Removing filler words and dead air on a Mac has the full list.

How short does it make the pauses?

It shortens dead air of 0.6 seconds or more, and never closes a gap completely: each pause comes down to the tightest quarter of the speaker's own gaps between sentences, so the edit is tighter and still sounds like the person. In the demo, a 1.80-second gap becomes 0.42 seconds.

Does it work with more than one speaker?

Yes. It tells speakers apart by voice, and it reads every pause by where it falls in a sentence and what it is for.

Can I make it cut tighter or looser?

Yes. The format sets the pace, so a 60-second vertical clip is cut tighter than a long interview, and you can ask for tighter or looser in plain language.

What is an edit decision list?

A list of every edit in a video: where each piece comes from, what happens to it and where it goes. Sharpy's list also says why each edit was made.