How to edit a podcast or interview video with ChatGPT or Claude (edit from the transcript)

Transcribe each raw file, put every camera and mic on one timeline, have ChatGPT or Claude plan the cut and write it as a clip list, then render it with free ffmpeg. You end up with a tight edit cut on whole sentences, a written edit plan, and a quick way to make new versions from notes with times.

Christian MartinFrom the playbook Edits raw podcast and camera footage into a tight cut, then fixes your notes by Christian Martin ✓ · Updated Nov 19, 2026

What you need

  • The raw files for one recording: one file per camera, mic or screen recording, copied to your computer (not streaming from a cloud drive)
  • One line per file saying who or what is on it, such as "Maya, camera and mic", "Wide camera", "Screen"
  • Whether it was recorded in one room or remote (Zoom, Riverside, each person on their own computer)
  • The brief: who watches, the one point of the video, the call to action, the size (16:9, 9:16 or square) and the length you want
  • Anything that must stay in or must come out
  • ffmpeg installed on your computer (free), or a ChatGPT or Claude plan with a code tool that can run it
  • A way to get a transcript with times for each file that has sound: Riverside, Descript or Zoom transcript exports, free Whisper on your computer, or a paid speech-to-text service

The steps

  1. Write the brief and label the files

    Paste the brief into ChatGPT or Claude: audience, the one point, the call to action, size, target length, what must stay in or come out. Then list every file with one line each: file name, who or what is on it, and whether it has sound. Ask the AI to repeat the source list back as a table with a short label for each file (maya, dev, wide, screen). Check every label against the actual footage before you go on. A wrong label puts the wrong name on every word of that transcript, and that's one of the two main ways an edit goes bad.

  2. Measure the files and transcribe each one once

    Ask the AI for the ffprobe command that prints each file's length and whether it has an audio track, then run it and paste the output back. Add up the minutes of audio. If you pay for transcription, check today's price and work out the cost before you start. Transcribe each file with sound one time, with times on every word or at least every sentence. Files with no sound need no transcript. Paste each transcript in under its label so every line carries the speaker's name. Fix any wrong speaker names now, before any planning.

  3. Put every source on one clock

    Choose one file as the master clock (usually the main host's mic). For each other file, work out its offset: source time = master time + offset. Same room: ask the AI for a short Python script that cross-correlates the audio of two files. Run it and paste the result. A clear single peak means a sure match. A flat or weak match means it isn't one. Remote: the files share no sound, so an audio match gives false numbers. If all recordings stopped at the same moment, offset = that file's length minus the master file's length. Then check one turn: a question that ends on one mic and the answer that starts on the other must land in order, with a gap under one second. Screen recording: find a spoken cue like "let me share my screen" and match it to the moment it shows up. Write down each offset and how you checked it.

  4. Plan the cut on paper

    Ask the AI to merge all the transcripts onto the master clock (speaker, time, words) and then write a short edit plan: the story in a few lines, which takes to keep, and what to cut and why. Read the plan against your brief. Does the one point come through, and does the call to action survive? Push back on anything it cut that you want kept. Save the plan, because you'll come back to it on every version.

  5. Write the edit as a clip list

    Ask for the edit as a JSON clip list: format (width, height, fps), each source with its offset in ms and a gain in dB, which mics go in the audio mix, and clips with in_ms, out_ms and which source is on screen. Give it these rules: cut on whole sentences. Shorten pauses longer than about 0.7 s to about 0.35 s. When a line is said twice, keep the last complete take and drop false starts. The camera follows the speaker, with no switch for reactions under about 1.5 s. Show the screen recording, with the whole picture visible, while the speaker talks about it. On a remote podcast, mix every mic into every clip and raise a quiet person's gain. Spot-check a few clips against the transcript to make sure none starts or ends mid-sentence.

  6. Render a short test cut first

    Ask the AI for an ffmpeg script (bash or Python) that turns the clip list into a video in chunks of about 30 seconds and then joins them. Before the full render, have it build a 2 to 3 minute test cut from a few clips spread across the recording. Run it and watch it. Check that lips match the sound, that each camera switch lands on the right person, and that the framing is right. A wrong offset is cheap to fix here and costly once it's in the whole video.

  7. Render the full cut and check every chunk

    Run the full render, with chunks in parallel if your computer can handle it. Ask the script to print the expected and actual length of every chunk. Render any short chunk again and never join one that came out short. Read sources from local disk only, because streaming from Dropbox or Google Drive can stop early without an error. Keep every chunk file and name it by its clips so the next version can reuse it. Save this as v1 and note its length.

  8. Work the notes with times

    Watch v1 and write notes with times, such as "v1 12:04 cut this story" or "v1 3:10 stay on Dev". You can also paste a block of transcript words to cut. Give the AI the notes with the version number, the v1 clip list and the plan. Ask it to tie each note to the clips at that time on that version, change the list, and tell you which chunks changed. Re-render only those chunks and reuse the rest. Ask for a short change list (what changed, and where it now sits in v2). Group several notes into one render, and repeat until the cut is done.

  9. Finish and export

    Once the cut is approved, ask for the finishing commands: ffmpeg loudnorm to about -14 LUFS for social or -16 LUFS for a podcast, and short audio fades at hard cuts if you want them. Export each size you need (16:9, 9:16, square). Ask the AI to build an SRT caption file from the transcript, retimed to the final cut. Make titles or lower thirds in a motion-graphics tool such as Remotion and drop the finished file in as a whole clip. Keep one folder per video with the brief, source list and offsets, transcripts, plan, clip list per version, notes, renders and finals. Don't post until you've watched the final. If it runs as an ad, put full UTM tags on its link (utm_source, utm_medium, utm_campaign, utm_content).

A prompt to start with

You are my video editor. You can't watch video, so we edit from the words: transcripts with times, a clip list, and ffmpeg commands I run on my computer.

Brief:
- Who watches: [audience]
- The one point: [main point]
- Call to action: [CTA]
- Size: [16:9 / 9:16 / square]. Length: [full recording tightened / target in minutes]
- Recorded: [same room / remote on Zoom or Riverside]
- Must stay in: [list]. Must come out: [list]

Source files (one line each: file name | who or what is on it | has sound yes/no):
[file1.mp4 | Maya, camera and mic | yes]
[file2.mp4 | Dev, camera and mic | yes]
[screen.mp4 | Screen | no]

Lengths from ffprobe: [paste output]

Transcripts with times, one per file with sound:
[paste each transcript under its label]

Do this in order and stop after each part so I can check it:
1. Repeat the source list as a table with short labels.
2. Work out each source's offset on one master clock (source time = master time + offset). For remote recordings, don't use audio matching. Use the length difference if all files stopped together, then check one question-and-answer turn lands in order with a gap under 1 second. Tell me how you checked.
3. Write a short edit plan: the story in a few lines, the takes you keep, and what you cut and why.
4. Write the clip list as JSON (format, sources with offset_ms and gain_db, audio mix, clips with in_ms, out_ms, video). Rules: cut on whole sentences. Shorten pauses over 0.7 s to 0.35 s. Keep the last complete take of a repeated line. Camera follows the speaker, with no switch for reactions under 1.5 s. Show the screen while it's being discussed, with the whole picture visible. Mix every mic into every clip.
5. Give me an ffmpeg script that renders a 2 to 3 minute test cut first, then the full cut in about 30-second chunks, prints expected vs actual length for each chunk, and keeps the chunks so later versions only re-render what changed.
6. Tell me the times to watch first and what isn't done yet.

Paste it into ChatGPT or Claude and put your own facts in the [brackets].

An example of the finished work

Harbor Lane Bookkeeping, podcast episode 14, "Clean Books Before Tax Season". Remote recording on Riverside, 58 minutes raw, cut to 41:12.

SOURCE LIST
- maya: Maya Ortiz (host), camera and mic. Master clock, offset 0 ms
- dev: Dev Patel (guest), camera and mic. Offset -143,200 ms, gain +5 dB (quiet mic)
- screen: Maya's screen share of the QuickBooks dashboard. Offset +1,812,400 ms

HOW THE OFFSETS WERE CHECKED
Both recordings stopped at the same moment. Maya's file is 58:41.2 and Dev's is 56:18.0, so Dev's offset = 3,378,000 - 3,521,200 = -143,200 ms. Turn check: Maya's question "so what breaks first in March?" ends at master 14:02.6, and Dev's answer starts at 14:03.1, a 0.5 s gap in the right order. Screen: Maya says "here's the reconciliation tab" at master 31:40.0, and the tab opens on screen at that moment.

EDIT PLAN (v1)
Story: small firms leave reconciliation until March, three things break, and Dev shows the 20-minute weekly fix. The call to action is to book a free books review at harborlane.com/review.
Kept: the cold open on "the March panic" (12:10–12:48), the three breakdowns, the screen walkthrough, and the closing call to action (second take).
Cut: 9 minutes of pre-show chat and mic checks, Dev's story about his old firm (about 2 minutes, off point), four false starts, and the first take of the call to action.

CLIP LIST (excerpt, v1)
{ "format": {"width":1920,"height":1080,"fps":30},
"sources": {"maya":{"offset_ms":0,"gain_db":0},"dev":{"offset_ms":-143200,"gain_db":5},"screen":{"offset_ms":1812400,"gain_db":0}},
"audio": ["maya","dev"],
"clips": [
{"in_ms":730000,"out_ms":768000,"video":"maya"},
{"in_ms":842600,"out_ms":871300,"video":"dev"},
{"in_ms":1900000,"out_ms":1958500,"video":"screen","fit":"contain"} ] }

RENDER LOG
Test cut: 2:40 from 6 clips. Sync and framing OK.
v1: 41:58, 84 chunks, all lengths match. One chunk came out 4 s short reading from Google Drive and was re-rendered from local disk.

NOTES ON v1 → v2
- v1 12:04 cut the payroll story → removed 38 s, 2 chunks re-rendered
- v1 22:15 stay on Dev, Maya's "mm-hm" switch is distracting → merged 2 clips, 1 chunk re-rendered
- v1 33:50 keep the full dashboard on screen → set fit to contain, 1 chunk re-rendered
v2: 41:12. 4 of 83 chunks re-rendered, 79 reused.

FINALS
- episode14_16x9.mp4 (loudnorm -16 LUFS, for YouTube)
- episode14_9x16_clip1.mp4 (the 0:58 "three things break" short, -14 LUFS)
- episode14.srt (captions, timed to v2)
Not posted yet. Waiting on Maya's final watch.

The business in this example is made up.

Or hand the whole job to the playbook

Christian Martin wrote these steps as a playbook on Iggy. It runs in your own ChatGPT or Claude, and Iggy gives it the facts of your business, so you do not explain them each time. In the free webinar, we set Iggy up together, in real time.

Mistakes to avoid

  • Guessing at file labels or leaving them as "cam1" and "audio2"Write one line per file saying who or what is on it, and check it against the footage before you transcribe. The label becomes the speaker name on every word.
  • Lining up remote recordings by matching their audioRemote files share no sound, so an audio match gives false numbers. Use the length difference when all files stopped together, then check one question-and-answer turn lands in order with a gap under one second.
  • Rendering the whole hour before checking syncRender a 2 to 3 minute test cut from clips spread across the recording first. A wrong offset shows up there in minutes.
  • Reading source files straight from Dropbox or Google Drive during the renderCopy every source to local disk first. Streaming can stop early without an error and leave you with a short chunk.
  • Joining chunks without checking their lengthHave the script print expected and actual length for every chunk. Render any short one again before joining.
  • Sending notes without a version number, so times driftWrite "v1 12:04 cut this story" and give the AI that version's clip list, so it ties each note to the right clips.
  • Re-rendering the whole video for every round of notesKeep the chunk files named by their clips and re-render only the chunks whose clips changed.
  • Cutting mid-sentence or keeping the first take of a repeated lineTell the AI to cut on whole sentences and keep the last complete take, then spot-check a few cut points in the transcript.

Questions and answers

Can ChatGPT or Claude edit video directly?

Not by watching it. They edit from the words. You give them transcripts with times, they plan the cut and write a clip list, and ffmpeg (free) turns that list into the video. Some plans let the AI run ffmpeg in its own code tool. Otherwise it writes the commands and you run them on your computer.

Is this a replacement for a video editing service?

For podcasts, interviews and talking-head videos, it covers the slow part: syncing cameras, cutting false starts, trimming pauses and following the speaker. Wedding films, music videos and heavy motion graphics still need a human editor's eye.

How long does it take?

An hour of two-camera footage takes an editor about a day by hand. Done this way, most of your time goes into checking the plan and watching the test cut and v1. Each new version then takes minutes, because only the changed chunks re-render.

What does it cost?

ffmpeg is free. The main cost is transcription. Free Whisper on your computer costs nothing, and paid speech-to-text services charge by the minute of audio. Count the minutes of files with sound, check today's price, and transcribe each file only once.

Do I need word-level timestamps?

Word-level times give the tightest cuts and pause trimming. Sentence-level times from a Zoom, Riverside or Descript export still work for cutting on whole sentences, but the cuts will be a little looser.

Can I make vertical shorts from the same recording?

Yes. Ask for a second clip list at 1080x1920 with a crop on each speaker (x, y and zoom), using a short section from the edit plan. The offsets and transcripts stay the same.

How do I get captions?

Once the cut is final, ask the AI to build an SRT file from the transcript, retimed to the final clip list. Upload it with the video or burn it in with ffmpeg.

What if the cameras drift out of sync later in a long recording?

Check one turn near the start and one near the end. If the gap grows, the clocks drift. Ask the AI to set a second offset for the later part of that source, then re-check in the test cut.

More how-to guides