Key Takeaways

  • Short-form humor relies on jump cuts hitting the exact millisecond an acoustic tail ends, such as the hard consonant at the end of a word like "talking."
  • Brooke Lacey replaced manual editing software by pairing Claude Code with local terminal tools Whisper and FFmpeg.
  • Lacey prompts Claude to detect verbal editing cues like "cut that out," discard bad takes, and select the best delivery automatically.
  • Framing shifts, like switching between wide shots and close zooms, can be automated alongside audio inflection tracking.
  • Lacey maintains her audience while posting just once a week, avoiding the daily production grind through precise local automation.

The Method

Brooke Lacey joined TikTok in 2020 and built an audience balancing tech education with comedy. She quickly ran into the central bottleneck of short-form video: comic timing lives or dies on cuts that feel instantaneous.

“The trick there is those jump cuts. They're very difficult,” Lacey said. “It's not funny unless you are jumped at the very exact moment.”

Standard editing tools require dragging cut points across a timeline for hours. Standard automated AI video tools are blunt; they cut on silence thresholds rather than linguistic cadences, creating awkward micro-pauses that ruin jokes. Lacey wanted cuts timed to the exact acoustic tail of speech.

“The sound ends before you think it does,” Lacey explained. “It has to hit right after that word ends. We don't think about that when we're talking. Like right there, the G in talking.”

Instead of paying for clunky cloud software, Lacey wired together three local tools: Whisper for transcription, FFmpeg for video manipulation, and Claude Code to drive the workflow.

First, Lacey records long, messy takes directly on camera. She does not stop recording when she stumbles. If a line lands flat, she repeats it immediately. If she makes an outright mistake, she speaks an audio cue directly into the mic: "cut that out."

Second, she runs Whisper locally to extract accurate, word-level timestamps across the entire recording. Whisper maps the exact start and end times down to fractions of a second.

Third, she feeds the transcript and audio cadence instructions into Claude Code. Lacey trained the model on what her comedic delivery sounds like. When she introduces a clip with a phrase like “I've been in tech for 25 years and I've never been more scared for people,” she tells Claude to identify the slight laugh and vocal inflection in her voice to preserve the hook.

Claude inspects the raw timestamps, flags repetitions, and evaluates cadence. “It's listening and goes, 'Oh, I just heard you say cut that out.' And then it also was like, 'You repeated this a couple of times and I took the one that was the best cadence,'” Lacey said.

Finally, Claude scripts FFmpeg commands to slice the video at the exact millisecond each chosen word concludes. Host Claire Vo highlighted how this workflow handles visual shifts as well: “Jump cuts between times, clip to words. I love this idea of three frames... frame wide and close cut because if you all don't know, you got to do like this when you're on TikTok.” FFmpeg punches in for close zooms during punchlines and returns to wide frames for exposition, delivering a finished video locally in seconds.

Where This Breaks Down

This method requires clean acoustic separation. If you record in an echoey room or beneath loud street noise, speech-to-text models blur the boundaries of words. The acoustic tail of a trailing consonant bleeds into ambient noise, causing FFmpeg to slice off the end of your words.

It also requires you to understand your own comic delivery well enough to articulate it in prompt instructions. If you do not know why a delivery works, Claude cannot guess your timing preferences for you.

What to Do With This

Pick a 60-second video draft with multiple retakes and export the raw audio to a WAV file. Run local Whisper to spit out word-level JSON timestamps, then ask Claude Code to write an FFmpeg bash script that trims out any take immediately followed by the phrase "cut that out." Test the resulting cut against your manual timeline.