Edit Videos Faster With AI (For Non-Editors)

Record in the Descript desktop app or import your source video. Descript edits through the transcript, so accuracy matters.

Open Settings with Command/Control + ,, choose AI models, and select the most accurate transcription model available. If you know the language, select it instead of relying on automatic detection.

Add unusual product names or technical terms to the transcription glossary before processing. Model names vary by plan, so check Descript’s current automatic transcription (https://help.descript.com/hc/en-us/articles/10249424286477-Automatic-transcription) and app settings (https://help.descript.com/hc/en-us/articles/10606771145869-App-settings) documentation.

Pro tip: When you make a mistake while recording, say “cut that” or a unique marker word such as “strawberry.” You—and the AI—can search for that signal later instead of hunting through the entire video.

Open your project, confirm the source composition name and duration, then duplicate the composition. Rename the copy [GUIDE] - Clean Edit and make every cleanup change there.

Use the duplicate as your safety copy. For early cuts, prefer Ignore because struck-through text can be restored.

See Descript’s guides to composition duplication (https://help.descript.com/hc/en-us/articles/10612359826061-Duplicate-a-project-composition-selection-or-layer) and Ignore versus Delete (https://help.descript.com/hc/en-us/articles/10164872017933-Deleting-vs-ignoring-script-text).

If you want Underlord—Descript’s AI co-editor—to perform the first pass, paste this prompt:

Create a duplicate of the current source composition named "[GUIDE] - Clean Edit" and edit only that duplicate. Leave the source unchanged.

On the duplicate, perform a conservative first pass:
- trim dead air at the beginning and end;
- remove clear repeated takes and sentence restarts;
- remove spoken editor notes such as "cut that";
- remove only nonessential filler words.

Preserve meaning, screen actions, loading time, and useful comprehension pauses. Prefer reversible ignored or struck-through transcript edits. Do not shorten all word gaps yet. Do not add captions, publish, export, or modify any other composition. Report what changed and flag uncertain cuts for human review.

Before continuing, verify that the source still exists at its original duration and that only the named duplicate changed.

On the clean duplicate, open AI Tools → Remove filler words. Preview a few results before using bulk removal, and enable Avoid harsh cuts to prevent clipped speech.

Not every filler should disappear; some preserve emphasis or cover a visual action. Use the Remove filler words (https://help.descript.com/hc/en-us/articles/10164806394509-Remove-filler-words) panel to preview each candidate.

After the pass, watch the edited composition at up to 2× speed and listen across every cut boundary. If a word sounds clipped or the delivery becomes rushed, restore that edit before moving on.

### Shorten Long Pauses

Open AI Tools → Shorten word gaps and start by finding gaps longer than 1.0 second. Review individual candidates before using Shorten all.

Our recommended starting points are 0.25–0.5 seconds inside a sentence and 0.5–0.8 seconds between ideas. These are not Descript defaults—just what we recommend.

Keep longer pauses when the viewer needs time to read the screen, watch a click, wait for loading, or absorb a point.

For a safer delegated pass, use:

On "[GUIDE] - Clean Edit" only, review word gaps longer than 1.0 second. Shorten unintentional hesitation gaps conservatively, but preserve pauses needed for emphasis, screen actions, loading, or a clear break between ideas. As a starting point, use 0.25-0.5 seconds within a sentence and 0.5-0.8 seconds between ideas. Keep changes reversible. Do not alter the source, publish, or export. Report the gaps changed and any pauses left for review.

Descript’s Shorten word gaps (https://help.descript.com/hc/en-us/articles/10164807277453-Shorten-word-gaps) tool supports individual and bulk actions. Review individual gaps before automating them.

Run Edit for Clarity at Low. Preview each suggested removal, then accept or reject it.

Removed phrases appear as ignored text so they can be reviewed or restored. See Descript’s Edit for Clarity (https://help.descript.com/hc/en-us/articles/36841959272717-Edit-for-Clarity) guide.

Then watch the entire clean composition and check:

• Meaning: Did any cut remove context or change the claim?
• Retakes: Is the intended take present and the false start gone?
• Word boundaries: Did a cut clip a consonant, breath, or word ending?
• Pacing: Does the speaker still sound human rather than rushed?
• Visual timing: Is there enough time to read, click, or watch the screen action?
• Source safety: Does the untouched composition still have its original duration?

Studio Sound is optional and affects the source file, not just the composition. Adjust it carefully; Descript’s Studio Sound (https://help.descript.com/hc/en-us/articles/10327603613837-Studio-Sound) guide explains the scope.

For a final no-change verification pass, use:

On "[GUIDE] - Clean Edit" only, identify any remaining hesitation gaps of at least 1.0 second. Do not shorten them.

Report each safe candidate with its timestamp, current duration, suggested duration, and reason. Separately list gaps that must remain for screen action, loading, emphasis, ignored content, or a clear break between ideas.

Then verify that the source composition still exists at its original duration, only the named duplicate changed, and no caption, publish, or export action occurred. Do not make further edits after reporting.

Run this process manually once or twice before automating it so you know which edits the skill should make.

Connect Descript from the ChatGPT app marketplace and authorize the correct Drive. ChatGPT requires Plus or higher, and workspace admins may need to enable third-party connectors.

Follow Descript’s current connection instructions (https://help.descript.com/hc/en-us/articles/46215649218445-Connect-Descript-to-ChatGPT).

Then invoke Skill Creator in ChatGPT or Codex and paste:

Use Skill Creator to build a reusable video-editing skill that works through the connected Descript app/MCP.

The skill must duplicate the named source composition and edit only the duplicate. It should remove dead air, clear retakes, spoken edit notes such as "cut that," and nonessential filler words; enable conservative filler-word handling; shorten only unhelpful word gaps; preserve meaning, screen actions, loading time, emphasis, and useful pauses; and keep early transcript edits reversible.

Require verification after every pass. The source must remain unchanged, uncertain cuts must be reported for human review, and the skill must stop before captions, publishing, or export. If the available tool can only shorten all remaining gaps globally, report the candidates and stop instead of applying that action.

Name the skill "descript-clean-edit" and include a clear example showing how to run it with either a Descript project link or an exact project and composition name.

Keep your script, outline, and editing notes in the same ChatGPT or Codex project folder so the assistant can reference them. To run the finished skill, paste:

Use $descript-clean-edit on this Descript project: [PROJECT LINK OR EXACT PROJECT NAME].

Source composition: "[SOURCE NAME]"
Create: "[GUIDE] - Clean Edit"

Leave the source untouched. Run the staged conservative cleanup: obvious retakes, editor notes, and filler words first; long gaps second. For the final hesitation-gap pass, identify and report individual candidates only—do not apply a global shorten-word-gaps action. Verify the project after every pass, preserve screen actions and useful pauses, and stop before captions, publishing, or export. Report the source duration, clean-edit duration, and everything that still needs human review.

The job may take 10–15 minutes, and progress can be limited. Older projects can fail, so inspect the completed duplicate and verify the source.