Captions deserve their own review gate - not a checkbox inside final delivery - because caption errors are the errors viewers actually see: a name spelled wrong sits on screen in 40-point type while the audio it misheard has already moved on. The workflow that works has five passes: generate the AI draft, correct it in the transcript, review it on the frame with fresh eyes, approve captions as an explicit gate, then export per deliverable.
Why captions get their own gate
- They are watched, not heard. The majority of social viewing is sound-off, and accessibility audiences depend on captions entirely - an error rate a viewer would forgive in audio is unforgivable in text.
- AI drafts are strong, not finished. Auto-transcription nails common words and misses exactly the load-bearing ones: names, jargon, numbers, product terms. A caption pass exists to fix the words the model could not know.
- Caption quality is per-platform. Burned-in captions for social deliverables, sidecar files for players that caption themselves, safe-zone and line-length rules that differ per channel - the export-format split is in what format to export for client review.
The five passes
- Generate the draft. Auto-transcribe the cut; in Frooty the transcript is generated and editable, which matters because the correction pass happens in the text, not in a separate tool. The same transcript feeds the shorts pipeline in AI shorts from long video.
- Correct the text. Proper nouns, brand names, numbers, technical terms, homophones - work from the script, the brief, or the SME, not from your ear. The SME pass is exactly what course teams run; the loop is in the course creators guide.
- Review on the frame, with fresh eyes. This is the pass teams skip and regret. Someone who did not write the corrections watches the cut with captions visible and checks the reading experience: sync at cuts, line breaks mid-word, captions that outrun the shot, text in the title-safe zone. Frame-pinned comments make each problem a note at the exact second - "caption reads 'wait' at 0:41, audio says 'weight'". The comment mechanics are in timecoded comments vs frame-accurate annotations.
- Approve captions explicitly. "Captions approved" is a separate statement from "cut approved" - a reviewer signs the caption pass on the exact version, so the record shows both gates closed. The approval-chain pattern is in the approval workflow guide.
- Export per deliverable. Burned-in for social where captions are part of the look, sidecar (SRT/VTT) for YouTube and players, both for mixed campaigns - then version caption fixes like any other round so the history stacks instead of forks. The campaign-scale version of this is in the marketing teams workflow.
What the caption pass checks
| Check | What good looks like | Typical failure |
|---|---|---|
| Names & terms | Spelled per script/brief, verified once at source | Phonetic guesses of names and product terms |
| Numbers | Digits match the audio exactly | "45" vs "forty-five" inconsistency, misheard figures |
| Sync | Captions appear with the line, survive cuts | Lag after edits; captions straddling shot changes |
| Reading load | Two lines max, readable at platform pace | Walls of text that outrun fast dialogue |
| Placement | Inside safe zones, off faces and lower-third graphics | Captions colliding with burned-in titles |
| Style | Consistent casing, punctuation, speaker labels | Style drifts after the correction pass |
Who reviews captions
Not the person who made them. The correction pass is blind to its own typos, and the editor has heard the dialogue forty times - they no longer read captions, they remember them. The right reviewer is a fresh reader with the script in hand, and the tooling requirement is simple: captions visible during playback, frame-accurate commenting to pin each problem to its second, and versions so caption fixes stack as rounds instead of overwriting. Frooty's review checklist covers where this pass slots into the whole round - it is in the video review checklist, and the caption generation side is on the AI captions page.
FAQ
Do AI-generated captions still need human review?
Yes - the draft is strong on common words and weak on exactly the words that matter: names, numbers, jargon. The correction pass fixes what the model could not know, and the on-frame pass catches the reading-experience problems no transcript view shows.
Should captions be burned in or sidecar for review?
Both, in sequence: burned-in (or rendered) captions during the review pass, because reviewers must judge the reading experience on the frame - then export the sidecar SRT/VTT for delivery per platform. Reviewing captions from a separate text file hides every placement and sync problem.
Who should review captions?
A fresh reader who did not write the corrections, with the script or brief in hand. The editor who has heard the dialogue forty times stops reading captions and starts recalling them - which is exactly how "wait/weight" survives to delivery.
Do captions need a separate approval from the cut?
Yes - a caption approval is a different statement about a different artifact, and recording it separately on the version means delivery disputes ("the captions were wrong") have a clean answer about which gate failed and when. The explicit-approval pattern is in the approval workflow guide.