Captions plus a hook line
Two layers: a transcript of speech and an authored overlay. Mixing them into one track is the bug.
Two layers: a transcript of the speech, and a hook line you wrote. Mixing them into one track is the bug. Auto captions will not typeset a claim nobody said. An overlay will not follow the mouth.
The first second of a mute feed has to carry a sentence the scroller can finish without waiting for the transcript to catch the greeting. That sentence is authored. The rest of the clip still has to be readable as speech. That is transcribed. Versely keeps the jobs on different tools on purpose: add captions to a video for the soundtrack, add a text overlay for the line you typed. Text overlay is the glossary name for the second object.
Teams merge them because both are "words on the video." The merge looks like a VEED pass on a clip whose first line is a Figma headline, or a single overlay that stays up for 20 seconds while the founder talks. The scroller gets either a late transcript or a poster that lies about what is being said.
Two layers, two tools
Layer A — transcript. Speech in, timed captions out. 165 language codes. 21 basic looks at 1× credits, 9 dynamic at 2×. Caption Studio's nine families if you are drawing with the CSS catalog. The words must have been said. Names and numbers need a read-back. Placement follows the mouth and the platform chrome.
Layer B — hook. One line you wrote: the claim, the price, the "wait for it," the paradox. Position is top, centre, or bottom, required. Font, colour, optional background block. It does not change with the waveform unless you switch to timed text overlays. Nobody has to have said it. If they did say it, you still may want it as overlay so it can sit large in the first second, off the chrome, while the transcript starts on "hey guys."
Run A on the locked cut. Run B as a second burn, or in the same session as a second tool call. Do not paste the hook into the transcript and hope the captioner treats it as a title. The captioner does not read your brief. It reads the audio.
Do not ask a video model to paint the hook into the generate. You will get uneditable glyphs, often misspelled, always in the wrong band after a recrop. Type is these two tools.
Do not run A on a silent plate to "get a hook." There is no speech. You will spend a caption render on nothing. Write B.
What the hook is for
The hook is the sentence a mute scroller has to get before they have a reason to stay. Hook rate is that job as a metric. The overlay is the drawing of it.
A good hook line on a talking-head is rarely the first spoken sentence. People start with a greeting, a name, a context clause. The claim is in sentence two. Overlay the claim at frame one. Let the transcript follow the actual words, including the greeting, or trim the greeting so speech and hook agree. If you leave the greeting in and the hook says the claim, the viewer reads two different first lines. That is allowed if they do not fight. "I wasted $12k on ads" as overlay while the mouth says "what's up team" is a pattern interrupt. "50% off" as overlay while the mouth says "we would never discount this" is a lie in two layers.
Keep it short. A hook that wraps into a paragraph is a second caption block. You now have two transcripts. One line, large, stroke or card, top of frame on 9:16 so it does not sit in TikTok's composer field or Shorts' Subscribe row. Headline-family energy is fine here because it is a title. Do not also run Headline as the speech family.
The hook can stay for two seconds or for the whole clip. A price that remains true can stay. A "wait for it" that remains after the payoff is a spoiler. Timed overlays exist for hooks that must leave.
Do not karaoke the hook. Snap and Reveal need spoken tokens. A line you wrote has no word-level timestamps from the soundtrack. You can fake a timed overlay of authored words; that is still layer B, not a caption family.
What the transcript is for
The transcript is how the rest of the watch stays readable with the sound off. It is also accessibility for viewers who cannot hear the mix. It should say what was said, not what the thumbnail promised.
If the hook already states the claim, the transcript should not reprint it in the same band at the same time. Two copies of one sentence is how both sit on the face. Let the overlay own the claim at the top; let Classic / Paper / Ink own the speech in the mid-lower band. When the spoken line is the claim, you can drop the overlay and let the first caption be large — or keep a short overlay for the scroller who will not read a two-line cue.
Reading speed applies to both layers. A hook that vanishes at 0:00.4 did not hook. Hold it. Speech cues that flash because the talker is fast still need splitting. The overlay hold and the caption hold are different clocks. Do not time the hook to the waveform of a word it does not contain.
Language: the transcript uses the 165-code pass that matches the audio. The hook is copy you localise by writing. A French overlay on an English transcript is a choice (claim in the market language, speech as original). A French language code on English audio is a mistake. Do not "caption the hook" by pointing the transcriber at a silent title card.
Recaptioning speech must not delete the hook. That is the practical reason they are different tools. Fix a name in the transcript, burn A again, leave B. If the hook lived as a fake first cue in the caption file, it is gone.
The mix-up
The failure modes are specific.
Hook in the captioner. You type "STOP SCROLLING" into a prompt and run auto captions on a clip that starts with a laugh. The laugh becomes the first cue. The hook never appears. Or you add a silent slate with the hook as burned title, then caption the whole file and the slate has no speech so the generator skips it. Write an overlay.
Transcript in the overlay tool. You burn one sentence for the whole 30-second clip. The founder says four other things. Mute viewers get the hook and then a person mouthing silence. Caption the speech.
Both in one band. Top overlay and top Headline captions. Or bottom overlay and bottom Classic, under the composer field. Chrome eats the stack. Split vertical: hook top, speech mid-lower, face visible, right rail empty.
Same sentence twice. Overlay "This serum faded my dark spots in 11 days" plus a Snap caption of the same line, popping on the mouth. Pick one performance. Keep the other as a quiet transcript or delete it.
Legal line in the transcript. "Not a medical claim" said once at 0:18 as speech, captioned, then recaptioned away in a style pass. Put durable legal copy in an overlay (or on-screen card) that a recaption cannot wipe.
The AI caption generator is layer A. The overlay task is layer B. Burned-in captions can mean either once they are pixels; the job still has two names. Call them separately in the brief: "Classic captions of the voice, plus a top hook that says X." If the brief says "add captions" and you wanted a headline, you will get a transcript.
A mute scroller gives you about a second. That second is the overlay. The rest of the watch is the transcript. Two layers. Two tools. One file.
Stack the next cut in Versely.
FAQ
If the speaker says the hook, do I still need an overlay?
Only if the spoken hook is late or quiet. A claim in sentence one, captioned large, can be enough. A claim in sentence two needs an overlay at frame one.
Can timed overlays replace speech captions?
They can replace them on a silent B-roll listicle. On a talking-head they will not follow the mouth. You would be authoring a parallel script. That is a different video.
Why did my hook disappear after I recaptioned?
It was in the caption file. Recaption rewrites that file. Put the hook in the overlay tool so a transcript fix cannot delete it.