Yes, AI can edit a podcast video, and at Truth Work Media it does real work on real client episodes. Claude, the AI model from Anthropic, drives DaVinci Resolve through its scripting API to sync cameras, cut multicam episodes, build captions, render motion graphics and measure audio. A human editor or producer still sets the taste, picks between options, finishes the edit, and nothing is published until a person approves it.
Last updated: October 2026
The short version: AI podcast video editing works best when the AI does the build and the measurement and a human makes the calls. On one interview show, Claude synced two cameras to a 14-channel recorder with under 10 ms of drift over 44 minutes and cut the episode to match one the client already liked. On another, it leveled two voices to within 0.2 dB. A human still signed off on both.
What AI podcast video editing looks like in our studio
Claude works inside the same professional tool our editors use, DaVinci Resolve Studio, by writing and running Python against Resolve’s scripting API. It builds projects and timelines, syncs cameras to separate recorder audio, cuts multicam and renders.
For motion graphics, it writes code. For audio, it measures every mic before touching an EQ. For anything involving faces, it runs face detection on rendered frames to check its own work.
Our hub article, How We Use AI to Produce Podcasts, covers the rest of the system. Here we stay on video and audio.

How does Claude build a full episode edit?
Take a full episode of a two-host interview show. The raw material was two cameras plus a 14-channel audio recorder. Camera audio is a sync reference for us, almost never the audio we use, so the first job is lining everything up to the recorder.
Claude found that the guest camera started 2.712 seconds after the host camera, synced both to the recorder, and measured drift at under 10 milliseconds across the 44 minute recording. That is precise, tedious work, and a machine is good at it.
Then the edit. The client had a reference episode they already liked, so the job was to match it, not reinvent it: a sponsor spot, a logo intro, the host’s intro to camera, then the conversation cutting between host, guest and a full-bleed split screen with a thin white divider. Median shot length about 5 seconds, with the show’s logo bug in the corner. Matching an approved reference removes most of the guessing, because Claude can measure it and build to it.
Rough cuts from a transcript
A creator sent us a four-part competition docuseries to rough cut. Whisper produced word-level transcripts, Claude built an edit decision list in code from them, then assembled and rendered the cut. Graphics cues were rendered separately as transparent overlays. The style reference was a well-known travel competition series, and that edit produced several of the rules below. For how we plan an edit before anyone cuts, read How to Use AI to Plan a Podcast Edit.
Can AI make captions and Shorts copy?
Yes. One editor we work with builds in a Blackmagic Cloud project with her own subtitle style already set. Claude duplicates her styled timeline, so the new one inherits her subtitle look, then imports an SRT caption track into it. She still finishes the font, intro and outro, music and transitions, and does the export.
For the words around the video, our AI content system, TruthFlow, writes Shorts copy with the rest of the episode content. Shorts captions are one paragraph, pull context from the full episode transcript and not just the clip, and keep the editor’s own titles word for word. More in How to Write Podcast Show Notes With AI.
Can AI create logo animations and podcast intros?
It can, and the method matters. Claude writes motion graphics as code: an HTML page rendered frame by frame in headless Chrome, then mixed with ffmpeg. Because it is code, every frame uses the client’s exact brand colors and fonts from their brand kit.
We have built promos this way for our own brand and for client episodes, plus a two-minute 3D animated conference booth loop for a tax advisory firm, built in three.js. We also build two-host intros from licensed motion templates in Resolve, back-timed so the song’s ending lands on the intro’s end, plus YouTube end screens and per-show Shorts cover brand kits.
How Claude handles music it cannot hear
Claude cannot listen to music. So when an interior designer launching a podcast needed YouTube intro options, Claude analyzed the waveform instead. It found music edit points using chroma and mel similarity on the beat grid, then timed the animation to a measured beat map.
We built three options, at 8.4, 9.3 and 9.6 seconds, each cut to a different music track she supplied, each matching her cover art and using B-roll of her talking. Her face was tracked with Apple Vision so the graphics framed her, and a script re-detected faces on the rendered frames to verify nothing covered her. She picked option 2.
Measuring a beat is not the same as feeling one. The numbers get the cuts close. A person decides whether it works.
Why we always show 2 or 3 options
We show 2 or 3 options (motion tests, intro options, storyboards), the client or producer picks, then we refine. On the booth loop, our producer picked a mix of two of the three motion tests and approved an 8-beat storyboard before the full build started.

How do brand guidelines become hard rules?
Most of our video rules started as a human reaction to something the AI made. We write each one down, and where we can, make it checkable.
| Rule | What it means in practice |
|---|---|
| Graphics never cover faces | Face detection picks the overlay’s side; faces are re-detected on rendered frames |
| Explainer graphics never sit in silence | They run over the person explaining the same thing |
| A card read on camera appears on screen | The viewer sees what the person is reading |
| Split screens are full-bleed | Edge to edge, not boxed |
| Runtime is “long enough for the story” | No padding |
| No flares or sparkle | Kept out of all motion graphics |
| Text on TV and booth videos is sized for distance | At 1080p, headlines about 96 to 118 px, list items about 44 px |
| No sustained low hum sound effects | Short transients only |
| Music at one steady low level, no ducking | Our producer rejected ducking twice as “pumping” |
| Personal to-camera videos skip the logo intro | Fade from black straight into the speaker, short logo outro, light music about 15 dB under the voice |
Once written down, Claude follows each rule every time. We learned the same thing on the copy side, covered in What AI Should Never Do in Your Podcast: a rule works when it is specific.
How does Claude polish podcast audio?
Each speaker’s mic is measured and EQ’d on its own, never through one shared chain.
On one episode, the guest’s mic was 6 to 9 dB brighter at 5 to 6 kHz than the host’s and had about 15 dB more low rumble. One shared EQ with a presence boost made him harsh.
That episode also exposed a Resolve gotcha. A mono clip on a mono track plays about 3 dB quieter than on a stereo track, so two files we had matched came out about 2.6 dB apart in the timeline. The fix had three parts:
- Per-mic EQ based on each mic’s own measurement.
- Loudness matched in LUFS (K-weighted, per ITU-R BS.1770), not plain RMS.
- The same track type for every voice.
The result: the two voices landed 0.2 dB apart.
Every delivery now runs a 16-check pass/fail verification, covering voice balance, voice isolation being on, no peaking, true-peak limiting and the loudness target. The final file is re-measured fresh for each video. If you are choosing mics, our podcast equipment guide covers what we recommend.

What can’t AI do in podcast video editing?
Taste. Our first AI-built promo came back from our producer as “feels ok, not award winning.” Version 2 shipped the same day, but the call that version 1 was not good enough was human.
Listening. Claude cannot hear music or voices. It reads waveforms, beat maps and loudness meters. That gets it close, and a person still listens.
Transcription errors that become edit errors. On one edit, Whisper merged a cut-off false start and the full retake into one sentence, so a repeated phrase survived into the cut. A human caught it at about 30 seconds in. Now we re-transcribe the finished voice track to catch repeats before delivery.
Resolve’s API limits. The scripting API cannot place text titles on a chosen track, so name tags and titles go on in an ffmpeg finishing step. Color management set through the API had no effect on iPhone HLG footage, so we generate an HLG to Rec.709 LUT instead.
Cloud projects. Blackmagic Cloud syncs edits but not media, so Claude builds only from clips the editor already has.
Sign-off. Nothing is published until a human has approved it. When a client’s team catches something, we turn that catch into a new rule.
Who does what on an AI-assisted video edit?
| Video task | What Claude does | What a human does |
|---|---|---|
| Camera and audio sync | Syncs cameras to recorder audio, measures offset and drift | Reviews the synced timeline |
| Multicam episode edit | Cuts host, guest and split screen to match a reference | Supplies the reference, watches the cut, signs off |
| Rough cut from a transcript | Builds an edit decision list from word-level transcripts, assembles, renders | Judges pacing and story, catches transcript errors |
| Captions | Duplicates the editor’s styled timeline, imports the SRT | Finishes and exports |
| Logo animations and promos | Renders motion graphics as code in exact brand colors and fonts | Picks an option, decides if it is good enough |
| Intros cut to music | Finds edit points from waveform analysis and a beat map | Listens and picks |
| Face-safe graphics | Tracks faces, re-detects them on rendered frames | Final visual review |
| Audio polish | Per-mic EQ, LUFS matching, 16-check verification | Listens, sets the taste, signs off |
Our rule of thumb: Claude does the build and the measurement. A human sets the taste, picks the options and signs off.
Frequently asked questions
Can AI edit a full podcast episode by itself?
No. Claude syncs, cuts and renders, but a human finishes the edit and a person approves it before it is published.
What AI do you use for podcast video editing?
Claude, from Anthropic, driving DaVinci Resolve Studio through its Python scripting API, plus Whisper for word-level transcripts and ffmpeg for finishing.
Can AI pick music edit points if it cannot hear?
It can get close. Claude analyzes the waveform on the beat grid and times cuts to a measured beat map. A person still listens and picks the final option.
Does AI video editing keep graphics off people’s faces?
In our workflow, yes. Face detection decides which side a graphic goes on, and a script re-detects faces on the rendered frames to confirm nothing is covered.
How do you get two podcast voices to the same volume?
EQ each mic on its own, match loudness in LUFS rather than plain RMS, and put every voice on the same track type in Resolve. On one episode that brought the voices within 0.2 dB.
Will AI editing make my show look generic?
Not if it is built to your references: an episode you already like, your exact brand kit, and 2 or 3 options for you to choose from.
Want this done for your show?
We produce video and audio podcasts with AI doing the build and the measurement and our team making every call that needs taste. See our pricing, or read the full picture in How We Use AI to Produce Podcasts.
About Truth Work Media
Truth Work Media is a podcast production company in South Bend, Indiana, founded in 2014, producing shows for businesses, professional services firms and nonprofits nationwide.
Podcast production by Truth Work Media, https://truthworkmedia.com


