DEV Community

neo xia
neo xia

Posted on

Repurposing long recordings: what it takes to get 3–5 usable vertical clips

Most creators already own the raw material for short-form video. It is sitting in a webinar recording, a podcast episode, or a 40-minute interview. The hard part is not filming more. It is finding the eight or ten seconds inside forty minutes that will actually hold attention on a vertical feed.

I have been building FreeClipping around that problem, and this post is about the pipeline decisions that mattered more than the model choices.

The real constraint is selection, not transcription

Transcription is a solved-ish commodity now. Whisper-class models give you word-level timestamps for a fraction of a cent per minute, and the quality on clear English speech is good enough that captioning is no longer the bottleneck.

What breaks a clipping tool is choosing which moments to cut. Two failure modes show up constantly:

  1. Overlapping cuts. Score every 15-second window independently and your top ten windows will all sit on top of each other. You ship one good clip five times.
  2. Truncated thoughts. A window that starts mid-sentence and ends mid-sentence looks fine in a transcript and terrible on screen.

The fix I settled on is boring: score candidate windows, then do constrained non-overlapping selection over the ranked list, and snap boundaries to sentence edges from the transcript before rendering. The clips get slightly shorter than the naive window, and dramatically more watchable.

Vertical framing is a product decision, not a CSS transform

Cropping a 16:9 frame to 9:16 is lossy by definition. There is no crop that keeps a two-person interview, a slide, and burned-in captions all legible. Every tool picks a lane:

  • face-tracking crops that follow the speaker,
  • letterboxed clips with a blurred background,
  • or a deliberate reframe that assumes one subject.

FreeClipping renders 9:16 MP4 with burned-in captions on a simple background. Captions are burned in rather than delivered as a sidecar file because the target platforms reward on-screen text and because a caption track that fails to load is worse than no caption at all.

Audio-only sources get the same treatment: the waveform becomes a captioned video on a plain background. Podcasters get vertical clips without ever turning on a camera.

Processing has to be observable

Video jobs take minutes, not milliseconds. If the only feedback is a spinner, users close the tab and assume it failed.

A persistent job URL solved this better than email notifications did. The job page shows status, and when it finishes the same URL is the download page for each clip. Users bookmark it, come back on their phone, and download. No account required, no magic link.

Retention is explicit rather than implied: sources and cached transcripts expire after seven days, generated clips after thirty. Saying that up front removes a whole category of support questions.

What the free tier actually costs you

The first recording is free up to 10 minutes without an account or a card. That sounds generous until you notice it is one recording — enough to evaluate whether the highlight selection is any good on your own material, not enough to run a weekly show.

Later Free usage requires sign-in and includes 20 source minutes per calendar month per verified account, with watermarked 720p exports. Paid tiers are $9/month for 150 source minutes and $19/month for 500, with 1080p exports and no watermark.

The sign-in requirement exists because per-account minutes need an account to attach to. It is not a growth funnel trick, and I would rather say that plainly than pretend otherwise.

Things that are still rough

  • Non-English speech is not supported yet. The highlight scorer is tuned on English transcripts and quality drops fast outside that.
  • Clips are 15–60 seconds. Longer cuts out of a single source are possible but the selection quality degrades.
  • There is no automatic publishing to TikTok, Reels, or Shorts. The output is a file you review and post yourself, which is also the honest workflow: nobody should auto-post an AI cut without watching it.

What I would tell the next person building this

Get the non-overlapping selection logic right before you touch framing. Ship a persistent job URL before you ship notifications. Make retention limits visible in the UI. And measure the thing that matters — how many of the generated clips a creator actually posts — instead of how many clips the model generated.

If you want to try the pipeline on your own recording, FreeClipping is the product; the first ten minutes cost nothing and no account is needed.

Disclosure: I work on FreeClipping. The copy in this post was drafted with AI assistance and reviewed before publishing.

Top comments (0)