The recording is fine. The problem is everything around the good part: the webinar that ran long, the talking-head take with three false starts before the real sentence, the customer testimonial whose best forty seconds sit in the middle of six minutes. Nobody on the team edits video, and a timeline editor is its own skill.
Cut: So, um, where do I even start.
It took us a few months, but it works.
Cut: Anyway, yeah, so that's that.
The claim keeps its condition
Cut: So, um, where do I even start.
It works.
Cut: Anyway, yeah, so that's that.
The condition is gone, and nobody notices in the script
Descript turns the recording into a transcript you edit like a document. Delete a sentence and the matching video goes with it. That leaves the decision that takes the time, which is what to cut. Claude can read the transcript and propose a cut list against a target length. You make the cuts, and you watch the result before anyone else does.
Watch Descript cut video by deleting textDescript's own video, on its Edit like a doc help page, checked 2026-10-23Import the recording and correct the transcript
Start a Descript project and add the recording. Descript transcribes it, and from then on the transcript is the edit: delete a sentence and the matching video goes with it.
Read the whole transcript once with the video playing before you cut anything. Fix misheard names and product terms in Correct mode (
Opt+Con a Mac,Alt+Con Windows). Correcting changes the text without touching the audio or video underneath.Do it now, not later. The captions are built from this text, and Claude will read whatever spelling is there.
Checkpoint
A transcript with every name spelled right.
Strip the filler words and fix the sound
Remove the filler words with Descript's AI tools. It lists every "um" and "uh" it found, each with a timestamp and a preview. Turn on Avoid harsh cuts: it skips any filler word that can't come out without clipping the words next to it.
Review the list rather than taking all of it. In a testimonial, the pause before an answer is sometimes the most believable second in it. Filler detection works on English transcripts only.
Then switch on Studio Sound for the script layer. It reduces background noise and echo in recorded speech. On current plans, both filler removal and Studio Sound use AI credits.
If the waveform goes flat once it's on, the noise was too loud to separate from the voice: lower the Intensity, or turn it off for that file.
Checkpoint
A transcript without the ums, and audio you've listened to with Studio Sound on.
Ask Claude for a cut list
the step that pays off
Export the transcript as a
.txtfile, with timecodes at paragraph breaks and speaker labels included (both are export settings). Keep this file. It's the original you'll check the edit against at the end.Before a testimonial transcript goes anywhere, check the customer agreed to it being edited and published, cut anything they said off the record, and read Claude's data settings before your first upload.
Start a new Claude chat, attach the file and run the prompt below with your target length and where the video is going. Compare the first and last timecodes it reports with the file, so you know it read all of it.
The rule doing the most work is the one about conditions. The usual failure in a tight edit is a claim that loses its condition: "it took us a few months, but it works" becomes "it works", and nobody notices in the script.
Step 1: Get a cut list that hits your target length
Checkpoint
A quoted cut list with an estimated length, and up to three clip candidates.
Make the cuts, then caption and export
Work down Claude's list. Search the script for each quoted passage, select it and delete it.
Ignoring text also cuts it from the video, but the struck-through text stays in the script, so it's easy to find and restore. To keep a passage in, leave it alone.
Editing in Descript is non-destructive, so a cut you regret comes back: restore the removed media.
Then add captions and pick a style. Captions come from the script, so they're already timed to the cut you just made.
Export it locally as an MP4. On the Free plan, exports carry a Descript watermark and are limited to 720p, which matters for a video that goes on your own site.
Pro tip
Descript also embeds a subtitle track in the MP4, which viewers can switch on where the platform supports it. That track isn't burned in; the captions layer is.
Checkpoint
An MP4 at your target length with captions on screen.
Watch it once at full speed before it goes out
Watch the export start to finish at normal speed, sound on, the way a viewer will. A cut that reads cleanly in the script can land mid-gesture or mid-breath, and only the video shows it. Fix each jarring jump in the script and export again.
For a testimonial, do one more pass. Don't edit a customer into saying something they didn't. Export the edited transcript (leave ignored text out) and run the check below against the original from the Claude step. Then send the customer the final cut and get their yes before it goes on your site.
Step 2: Check the edit did not change what anyone said
Checkpoint
One full watch, and for a testimonial, a meaning check and the customer's yes.
The takeaway
You end up with a captioned MP4 at the length you asked for, a cut list you can read back to see why each passage went, and, for a testimonial, a line-by-line check that the edit kept what the customer meant.
The limit: this fixes what was said badly, not what was filmed badly. Poor light, a cropped head or a speaker who never got to the point won't be rescued by the transcript. And a text edit can't see the picture, which is why the last step is a full watch and not a skim of the script.
Variations
Shorts from a long webinar. Descript's help article says Underlord, its AI co-editor, can find social clips and reformat them vertically. OpusClip does the same job from outside Descript: upload the file or drop in a video link, and it picks moments and adds captions. Its own FAQ says it also rearranges those moments and adds a call to action, so check each OpusClip clip against the original before it goes out. Run the meaning check on every clip either way. A clip is a cut with everything around it removed.
A written quote for the website. The fourth section of the cut list is a set of self-contained moments, quoted. One of them, approved by the customer, is a text testimonial with a video behind it.
Don't regenerate a customer's words. Descript's Regenerate can heal jump cuts and, with an authorized custom AI speaker, change what someone said. Descript asks for consent the first time you regenerate video, because it modifies a speaker's visual performance. On a testimonial, leave the jump cut showing. A visible cut is honest; a smoothed one is a performance the customer didn't give.
Don't hand the whole edit to an agent. Underlord's own help article calls it beta and says to check its work as you go. Use it for captions and clips, and keep the cut decisions in a list you can read.