Turn a long recording into short clips with Descript AI

Choose the question your clip should answer

A useful explanation can be buried inside a long lesson or interview. Descript’s Create clips tool uses AI to suggest short excerpts from an existing composition. Give it a specific question to look for, then review the suggested cuts against the recording. A strong opening alone is not enough: someone who missed the full video still needs to understand the answer.

Start with a recording you own or have permission to edit and publish. The source composition must be at least as long as the clip you request; Descript warns that recordings under one minute may not contain enough material to generate clips. This workflow is most useful when you already have a longer conversation or demonstration to work through.

Before importing a large recording, check your account’s media allowance and export options. Descript tracks media minutes separately from AI credits. Underlord requests can use different amounts of credits even for similar tasks, so start with a small batch and inspect Usage before repeating it.

Ask for a small set of candidates

In the right sidebar, open AI Tools, find Repurpose and select Create clips. Descript’s documented controls allow 1–20 clips, with a requested length from 10 seconds to 5 minutes. You can choose a layout, including portrait 9:16, and enter a topic, goal or selection criteria before clicking Submit.

For a first pass, try two clips of 30–60 seconds. These are suggested starting settings, not a platform requirement. Each result appears as an editable composition in the Clips from [composition name] folder. You can also ask Underlord to create clips, but the same review work applies.

Keep the condition attached to the advice

Open each candidate and listen to the source immediately before and after it. Check what words such as “this,” “they” or “in that case” refer to. Restore the setup if the excerpt depends on an earlier question. If a comparison loses one of its alternatives, the remaining line may sound more absolute than the speaker intended.

Consider this fictional teaching line: “When the sky is overcast, place the subject close to the window. If direct sun falls across their face, move them back.” A cut that starts at “place the subject close to the window” removes the condition. Keep the qualifying phrase, or choose a different passage that makes sense within the available time.

Watch the cut as well as reading it. An answer may depend on a hand gesture or a demonstration that appears a few seconds later. Keep that evidence in the clip. If you cannot preserve the meaning within your chosen duration, allow more time or reject the candidate.

Correct the transcript before styling captions

Descript builds captions from the script. Listen for names, numbers and missing words before choosing a caption style. To fix a transcription error, select the word and use Correct. Descript documents that this changes the displayed text without altering the underlying audio or video. It does not repair a factual mistake the speaker actually made.

Open the Captions panel and choose whether to apply a style to the current scene or all scenes. When changing an existing style, select the captions layer first: choosing a preset without that selection adds another layer. Check the result for overlapping text and read the captions at normal playback speed.

Check what the vertical crop removes

A portrait layout changes which parts of a landscape recording are visible. Play the entire clip and check that the person, their gestures and any object being discussed stay in view. For a screen demonstration, a wider view or a different composition may explain the action better than a tight crop of the speaker.

Preview the upload in its destination app before posting. Check that interface controls do not cover a key caption or the detail being demonstrated. Judge the actual preview instead of relying on one fixed margin for every platform.

Review one exported clip before making the rest

Export your strongest candidate and watch the file on a phone, first with sound and then muted. Listen for a clipped first word, an abrupt ending or a cut that makes the sentence misleading. Without sound, check whether the captions and visible demonstration still carry the explanation.

Keep the source timestamp and a short note about any restored context. If the first export needs changes, fix those before generating a larger batch. Finish with one complete, understandable answer that you are ready to publish.