Back to Blog
GuideNovember 18, 2026•9 min read

Podcast to Shorts: How to Pick the Best Clips (What the Ranking Looks At)

What an AI clipper's ranking reads, from RemixViral's code, and what it cannot see.

Short answer

  • The ranking reads text, not video. RemixViral transcribes your upload with Whisper, numbers the sentences, and asks one language-model call (default gpt-4o-mini) for the best standalone moments. The model never sees frames or audio, so it cannot judge tone, laughter, a guest's reaction or what is on screen.
  • The model picks sentence numbers; code sets the seconds. Code snaps every pick to sentence edges, stretches or trims it into your length window, drops overlaps and keeps the top N by the model's own 0 to 100 score. That score ranks clips inside one video. It is not a prediction of views, and we make no claim about views here.
  • A transparent rules-based version is free to try in the browser. It scores a pasted transcript with fixed weights and shows its reasons. It is a different, simpler ranker, and our code says it is not a stand-in for the product's.

We build RemixViral. This post describes what its code does, plus counts from the one 18-minute job in an earlier post. Nothing here says the ranking finds clips that get more views.

Two rankers, not one

Free Clip Moment FinderClip project (credits, one free project per account)
Where it runsYour browser, on text you paste; nothing is uploadedRemixViral's worker, on a video you upload
InputAn SRT, WebVTT or timestamped transcriptThe video: a Whisper transcript with word times
Who ranksFixed rules in code; every score comes with reasonsOne language-model call, then code
Sees video or audioNoNo
Sourcesrc/lib/clip-moments/find-moments.tsworker/src/lib/highlight-selector.ts in our shared worker repository

The worker is shared with a sister product, so its code lives in a sibling repository; the RemixViral web app only sends the job (src/lib/clips/worker-client.ts).

What the clip project's ranking asks for

The prompt tells the model it is "a short-form video editor" reading one numbered line per sentence with its start time, and lists what a moment needs. To qualify, a moment must meet all of these:

  1. It stands alone. No "as I said before", no unresolved pronoun, no answer to a question asked outside the clip.
  2. Its first line is a hook: a claim, a question, a number, a confession or a contradiction. Never filler, a greeting, an ad read or the tail of the previous thought.
  3. It resolves: the point lands before the clip ends.
  4. It fits your window (the default is 20 to 90 seconds).
  5. It does not overlap another pick.

Two more instructions shape the list: spread the picks across the whole video instead of clustering at the start, and give an honest score from 0 to 100 using the full range. If you typed a focus (up to 200 characters, for example "the part about pricing"), the model is told to prefer matching moments and to give any non-matching extras clearly lower scores. Your text goes into the prompt as a quoted topic after control characters are stripped.

Each pick comes back with a title (60 characters at most), a one-sentence reason and the score. The reason is what you see in the dashboard next to each clip.

What the code does after the model answers

The model is never asked for a timestamp. The worker's own comment explains why: models confidently invent times like "00:14:32", but a line number either exists in the transcript or it does not.

StepWhat the code does
RejectA pick whose line numbers do not exist is dropped
Snap the startWalk back up to 8 segments to a segment that follows a sentence end
Snap the endWalk forward up to 8 segments to a sentence end
Fit the windowGrow forward until the clip reaches the minimum, back the start up if still short, trim the end if over the maximum
OverlapA pick that overlaps an accepted one by more than 0.25 s is dropped
RankSort by score, highest first, then by start time; keep the top N
TimingStart at the first word's start, end at the last word's end

A worker comment says why snapping exists: Whisper segments are about 5 seconds long, not sentences, so roughly half begin mid-clause, and on one real 32-minute source 5 of the model's 10 picks started mid-sentence before snapping. That is a comment; we cannot show you the data.

A worked example, tested. I ran the worker's real materializeClips function on 12 synthetic segments of 5 seconds each (no model call), with five hand-written picks and a 12 to 25 second window. On 2026-11-18 it returned:

Pick (score)What I fed itWhat came out
"Raised price" (91)starts on segment 4, which begins mid-clause15 s to 30 s: the start snapped back one segment to the sentence start
"Overlaps pick 1" (80)overlaps the first pickdropped
"Why cheap plans churn" (74)segments 7 to 935 s to 50 s
"bad index" (99)segment numbers that do not existdropped
"Greeting" (20)a single 5-second hellogrown to 0 s to 15 s and kept as the third clip

The last row is the limit: nothing in the code rejects a greeting; it relies on the prompt and the low score. Ask for more clips than the video has good moments and the list fills with weaker ones. That the model then scores them low is our inference from the prompt, not a measurement.

What the ranking cannot see

  • Audio and picture. A funny pause, a laugh, a guest leaning in or a slide on screen is invisible. A transcript-only ranking will undervalue a moment that works because of delivery.
  • Who is speaking. The ranking prompt has no speaker labels. Layout (fill the frame or stack two speakers) is chosen later, at render time.
  • Your audience. There is no input from your channel, your past clips or any platform metric.
  • Bad transcripts. If Whisper returns nothing, the job fails with "source may have no speech". A noisy file gives a noisy transcript and a worse ranking.
  • Very long transcripts. The numbered transcript pasted into the prompt is capped at 120,000 characters. Beyond that, neighbouring segments are merged pairwise until it fits, which makes the model's view coarser. We have not measured which source length reaches this.

Picking clips from a podcast, step by step

  1. Set the window to the platform. Presets in the product are under 30 s (15 to 30), 30 to 60 s, 60 to 90 s, or Auto (20 to 90 s). Check each platform's current limit before choosing.
  2. Type a focus if you know the topic. A guest's story or a pricing section is easier to find with 200 characters of direction than by scrolling. It is a topic, not an instruction, so check the results.
  3. Ask for the count you will post. The model is asked for N picks; 20 from a 30-minute episode invites filler.
  4. Read the reason and the score together. A one-sentence reason that sounds generic is a sign the moment is weak. Compare scores only inside the same video.
  5. Watch the first line. Even with snapping, the model can choose a weak opener. If you need the sentence before it, extend the start.
  6. Watch with sound. The ranking never heard the clip.
  7. Optionally tighten. A per-job option cuts filler words and shrinks pauses longer than 0.7 s to 0.25 s. It does nothing when it would remove more than 40% of a clip.

What our one measured job looked like

From the earlier timing post: an 18-minute upload (1,078 s), transcribed in 2 Whisper chunks, returned 10 ranked moments in 3 min 42 s. They ran 20 to 57 s (median 30 s), covered 30% of the source and carried scores from 75 to 85. That is one job on one video. It shows the shape of an output, not how good the picks are.

Try the free rules-based ranker

The free Clip Moment Finder scores pasted text with fixed weights. A window must start after a sentence end and end on one, needs a minimum length of at least 5 seconds, and may not overlap another pick.

SignalPoints
Opens with a question / a direct claim+14 / +12
A number in the first line+4
Questions, numbers, contrast words (but, however, actually), emphasis words+4, +3, +3, +2 each, capped at 2, 3, 3, 3
Steady pace, 1.3 to 2.6 words a secondup to +12
Clean start, clean end+8, +8
Length close to the middle of your windowup to +8
Filler words above 2% of the wordsup to -15
Pauses longer than 2 s-3 each, up to -12

On its built-in sample (a 1 minute 25 second kitchen-workshop transcript, window 20 to 45 s, 3 requested) it returned two moments: 0:04 to 0:41 scored 77 ("opens with a question", 4 numbers, 3 contrast words) and 0:52 to 1:19 scored 51. The "um, so, you know" stretch between them was left out. Run it on your own transcript and read the reasons.

Frequently asked questions

Does the ranking predict which clips will get views?

No. The score is the model's own estimate of how clip-worthy a moment is inside one video, and we measured nothing about views or retention.

Does it use video or audio, or only the transcript?

Only the transcript with segment start times. It cannot hear tone or see the picture.

Can I steer it?

Yes: a length window, a clip count and a focus of up to 200 characters. The focus is treated as a topic, and moments that do not match it should score lower.

Which model does it use?

The worker's default is gpt-4o-mini at temperature 0.4 with JSON output. We did not test other models.

Does the free project rank differently?

It runs the same pipeline with a source capped at 10 minutes and keeps the best 3 clips by score.

Methodology and data

  • Code read on 2026-11-18: RemixViral origin/main at e32c8aa7f (src/lib/clip-moments/find-moments.ts, src/lib/clips/job-options.ts, src/lib/clips/types.ts, src/app/api/cron/process-clip-jobs/route.ts) and the shared worker repository's origin/main at 90bab6de3 (worker/src/lib/highlight-selector.ts, clip-ingest.ts, clip-tighten.ts). Both repositories are private.
  • Tested: materializeClips with the five picks above (synthetic transcript, no model call) and findMoments on the built-in sample, both run with bun on 2026-11-18.
  • Not tested: any live model call, how good or consistent the model's picks are, whether scores spread across the 0 to 100 range, behaviour near the 120,000-character prompt cap, and any effect on views or retention.
  • Measured job: the 18-minute job (10 moments, scores 75 to 85) comes from the production query in the timing post (2026-11-17). Its n is 1.

We make RemixViral.

Your first clip project is free, no card needed: sign in with Google, upload one video you own (up to 10 minutes) and get up to 3 ranked clips.

See pricing and the free clip project →

Written by RemixViral Team

From the same team