Should I vibe code
Finds highlights and turns long video into captioned short clips
A highlight extractor's whole job is removing context. The guest in the clip will have a different word for it.
?
Their verdict, the Pro (month-to-month) price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
The pipeline is unusually well-trodden: transcribe with word-level timestamps, ask a model which sixty seconds are interesting, crop to 9:16, burn in captions, write out with ffmpeg. An agent will produce a working version of that on a Saturday and the first clip will look astonishingly close to the real thing. The distance from there to a tool you would let publish unattended is made of two very different problems. One is craft — cuts that land on breaths rather than on transcript boundaries, reframing that follows the right speaker on a two-shot, captions that get names and numbers right when they are welded into the pixels and cannot be patched. The other is the thing nobody scopes, which is that a highlight extractor is a machine for removing context and then publishing the result. The most clippable forty seconds of any conversation is very often the moment somebody states the position they are about to demolish, and the model reading your transcript has no way to know that. Add the auto-post step OpusClip sells and you have automated the part where a human would have watched it first. Build the pipeline; it is a genuinely good weekend. Just do not wire the last stage to a publish button.
What actually breaks
not "if". the specific failures.
- The cut, first and constantly. Segment-level ASR timestamps drift by a few hundred milliseconds, so every clip starts a syllable late and ends on an inhale, and the fix is word-level alignment rather than a better prompt
- Burned-in captions, which are the one output you cannot patch. A misheard surname or a transposed figure is now in the pixels, and correcting it means a re-render and a re-upload that resets the view count to zero
- Hook selection, because the model is scoring a transcript for quotability with no model of the argument. The forty seconds it likes best is regularly the setup somebody spent the next two minutes disagreeing with
- Auto-reframe on anything other than a single talking head. Face tracking on a two-shot either jitters between speakers once a second or confidently follows the person who is not talking
- Speaker attribution, when colour-coded captions put a sentence in the wrong person's mouth on a clip that has your podcast's name on it
- Rights, which the pipeline has no way to check. It will happily clip a stream you were watching, a film, a conference talk or a track playing in the background, and Content ID does not care that the long-form version was cleared
- Cost and queue, once the input is a two-hour recording rather than a ten-minute test. Transcode plus inference per hour of source is real money and does not fit in a request timeout
- The auto-post step, which is where every other item on this list stops being a draft and becomes something strangers saw
- Storage, quietly. You are now holding every raw upload you ever fed it, including the recordings you meant to delete
Is that you?
the verdict is a default, not a law
- The output is a folder of drafts and a human watches every one before anything is published
- The source is footage you shot or own outright, and nobody else appears in it without knowing what it is for
- It runs locally against files on your disk, with no upload and no queue
- You treat it as a rough-cut generator that saves you scrubbing, not as a publishing system
- It posts on a schedule without a person watching the clip first
- The inputs are other people's videos — streams, talks, films, YouTube downloads — regardless of how the output is credited
- Guests appear in the source and have not agreed to being cut into fragments and captioned
- Captions go out unreviewed on anything where a name, a number or a negation matters
- You are running it for clients, which turns every one of these into someone else's brand rather than yours
If you build it anyway
the checklist, then the prompt that enforces it
- Use word-level timestamps, not segment boundaries, and snap every cut to a silence rather than to a transcript index. This single change is most of the visible quality gap.
- Always keep a few seconds of lead-in on either side of the chosen span, and show the operator the surrounding transcript before export. Context is the thing being destroyed; make it visible at the moment of the decision.
- Render captions as a sidecar subtitle track first and burn them in only at the final export. Once they are in the pixels the only correction is a re-upload.
- Put a human approval step between generation and publication, and make the automated path stop at 'draft'. If a scheduler exists, it schedules approved clips only.
- Record the provenance of every source file — where it came from, who is in it, what you are permitted to do with it — before the first frame is processed, and refuse to run on anything without it.
- Cap and price the render path explicitly: maximum source duration, maximum clips per run, a hard spend limit. A two-hour upload is not a request, it is a job.
- Flag low-confidence transcription rather than rendering it. Names, numbers and negations are where ASR fails and where a wrong caption does actual damage.
- Set a retention policy on uploaded source video at the point you accept it. Nobody ever goes back and deletes the raw files.
I am building a pipeline that turns long video into captioned short clips.
The risky part is not the encoding, it is that the tool removes context and
then publishes. Apply these in order and push back when I ask for otherwise.
1. Before any clipping code, ask me where the source video comes from and who
appears in it. If the answer is anything other than footage I own with
consenting participants, stop and tell me what I am about to republish.
2. Transcribe with word-level timestamps and snap every cut to a detected
silence, not to a transcript segment boundary. Explain the drift I would
otherwise get.
3. Always include lead-in and lead-out padding, and print the surrounding
transcript around each proposed clip so I can see the context being cut off.
4. The default output is a draft folder. Do not wire publishing or scheduling
into the same run as generation, and if I ask you to, make me confirm that
nobody will watch the clip before strangers do.
5. Produce captions as a sidecar subtitle file first. Burn-in happens only at
final export, after I approve the text.
6. Mark low-confidence words instead of silently rendering them, and treat
names, numbers and negations as always requiring review.
7. Set hard limits before the first run: maximum source length, maximum clips
per job, maximum spend. Tell me the per-hour cost of transcode plus
inference for the settings you chose.
8. Store provenance with every source file — origin, participants, permission —
and refuse to process a file that has none.
9. Give uploaded source video a retention deadline at ingest and write the
cleanup job before the review UI.
10. Out of scope unless I ask: face-tracking reframe, virality scoring, and
multi-platform posting. Say so, and tell me OpusClip is $29 a month
month-to-month, which is less than the GPU bill for one bad weekend.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
When the clips are going to be published on a schedule rather than reviewed one at a time. $29 a month buys the two things a weekend build does not have: reframing and caption rendering that thousands of people have already found the ugly edges of, and a queue with an approval step in it. The economics also flip fast — a two-hour podcast run through your own transcode and inference does not cost nothing, and three of them a month is most of the subscription. Build the pipeline for the pleasure of it; buy the thing that decides what goes out.
$29/mo is cheaper than your weekend.
There is nothing to migrate off, which is the nice thing about this category: the artefacts are files. Keep the source video, the transcript with timestamps, and the clip definitions — in point, out point, caption text — as plain data next to the renders, and you can rebuild every output in any tool including the commercial one. The part that needs an actual plan is the published side. Keep a record of which clip came from which source, with the timestamp span, so that when a guest asks you to take one down you can find it, remove it everywhere it was posted, and check whether the same span went out three other times.
Python library that does exactly this job — transcript-driven clip selection plus speaker-aware resizing — and has been quiet since early 2024.
Word-level timestamps and speaker diarisation, which is the difference between cuts that land and cuts that clip a syllable.
Silence and scene-based automatic cutting, actively developed, and the fastest way to see how much of this is solved already.
Questions
Why is context the risk rather than quality?
Because quality failures are visible and context failures are not. A jittery crop or a bad cut makes a clip look amateur and you fix it. A forty-second extract of someone stating a position they immediately went on to reject looks perfect, plays well, and is a misrepresentation of a real person that you published under your own name. The model picked it precisely because it was the most quotable moment, which is the same property that makes it the most dangerous one.
What is actually the problem with clipping someone else's video?
That the pipeline has no idea it is doing it. OpusClip makes you assert you have the rights; your Saturday build asks nothing. In the US the argument you would be relying on is fair use, which is a four-factor defence assessed after the fact rather than a rule you can check in code, and in the EU the quotation and parody exceptions are narrower than people assume. Separately from copyright, downloading source from a platform generally breaches its terms, and background music trips automated matching regardless of what the surrounding footage is.
Is running this locally enough to make it safe?
It removes most of the boring risk — no uploads, no queue, no storage of other people's recordings on a server you administer. It removes none of the interesting risk, because the interesting risk is at the output. A clip rendered on your laptop and posted from your laptop is exactly as out of context as one rendered in a datacentre. Local is the right default; the approval step is the actual control.
- U.S. Copyright Office — Fair Use Index (the four-factor test, decided case by case)
- Directive (EU) 2019/790 on copyright in the Digital Single Market
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
Your cut lands on the nearest keyframe, not where you clicked. Closing that gap is most of what an editor costs.
A browser video editor is a weekend right up until Export. That button is the entire product.
It's ffmpeg with a template. The product is the hundred boring decisions after the first render looks great.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice