Brooke Lacey joins Claire Vo on How I AI for a 36-minute walkthrough of her short-form video workflow. Claude Code coordinates existing tools; Lacey supplies the performance and editing judgment.

Why precise cuts matter

  • Lacey says editing one apparently casual video could take hours. Previous attempts to delegate her comic timing fell short.
  • Her example: trimming a two-character sketch from about 65 to 55 seconds made the joke land.
  • Her first experiment used roughly 25–30 minutes of continuous phone footage, including repeated lines and spoken “cut that out” cues.

The editing chain

  1. Transfer the take: AirDrop puts the phone recording on her MacBook for processing with FFmpeg, Whisper, and Python.
  2. Find the boundaries: silence detection locates speech; Whisper’s word timing helps place clean cuts.
  3. Specify the shots: face tracking supports the crop, with wide, mid, and close framing assigned per cut.
  4. Write the edit: a TSV, or tab-separated table, records each selection’s start, end, shot size, and label.
  5. Finish the video: add app screenshots, graphics beside the speaker, and white captions with black outlines.

What the prepared demo shows

  • The walkthrough reports a five-minute take becoming a roughly 15-second vertical ad, with the chain running in about 15 seconds.
  • It describes eight cuts, 57 caption words, and 441 output frames at 30 frames per second.
  • App imagery occupies 40 frames; graphics fit the empty wall beside Lacey, measured separately for each cut.
  • The agent reports zero audio/video offset and support for additional durations and landscape output; those exports are described, not separately demonstrated.

Teach a style with examples

  • Lacey points Claude toward existing videos and explains the little laugh, pauses, and cadence that make her opening hook work.
  • Vo contrasts this with her own unsuccessful instruction to simply make a funny video.
  • The same style knowledge informs cuts and new scripts, but Lacey says her two-character comedy format remains unfinished work.

Pair the video with its audience

  • Lacey manually downloads TikToks on her phone and puts them in Google Drive; an authenticated browser supplies the associated comments.
  • Claude connects the performance to audience questions, including one about AI in hospital billing, to suggest follow-up content.
  • This is a workflow with manual handoffs and browser corrections, not evidence of reliable unattended TikTok scraping.

Keep the business claims in proportion

  • Lacey posts about once a week, supported by years of audience-building and older videos that continue circulating.
  • Livestreams and TikTok ads are next experiments. The interview supplies no conversion results or controlled growth comparison.
  • Her Gen X perspective comes from earlier shifts: discovering Windows 3.1, starting college in 1994, and working on Y2K fixes.

Make corrections persist

  • Her personal “lane” operating system separates project memory, write restrictions, status Markdown files, and global instructions.
  • When Claude drifts, she sends it back to the instructions and asks for the underlying correction to reach the central system.

“That file is the whole edit.” — Brooke Lacey, about 13:15

Caveat: this is a prepared creator demo, not an independent speed test or privacy audit. Local video processing does not establish that all surrounding AI interactions stay local.