Tutorials
Claude Opus 5.5: Repair Product Video Subtitle Timing
Before asking Claude Opus 5.5 to retime subtitles, determine whether they have a constant offset, progressive drift or isolated bad cues.
Before asking Claude Opus 5.5 to retime subtitles, determine whether they have a constant offset, progressive drift or isolated bad cues. Each requires a different repair. Shifting every cue can fix a uniform delay while making an already correct opening worse when the real problem is drift.
The workflow below uses invented timing examples, not measured Opus output. Provide the final video version, editable caption file, approved transcript and manually checked speech anchors. If the model cannot access the audio, it can audit timestamp structure but cannot honestly confirm synchronization.
Measure the error at several anchors
Choose clearly audible words near the beginning, middle and end. For each, record the actual speech time and its subtitle time. Compare the difference in seconds using the same timeline origin. A cropped intro or a newly inserted logo animation often changes the origin; a transcript from a different edit can require cue-by-cue reconstruction.
| Pattern | Example | Repair to investigate |
|---|---|---|
| Constant offset | Every anchor is 0.4 s late | Shift all affected cues by −0.4 s |
| Progressive drift | Aligned at start, 1.5 s late near end | Check edit version and time scaling |
| Local error | Only one sentence appears late | Repair those cue boundaries |
| Bad wording | Timing matches, product name is wrong | Correct text without global retiming |
Editors discussing automatic subtitles describe cleanup of wording and timestamps after transcription. That is a reader problem, not evidence that a particular model solves it reliably.
A worked offset and reading-load example
In a hypothetical 12-second ad, speech anchors occur at 1.0, 5.0 and 9.0 seconds. Their cues appear at 1.4, 5.4 and 9.4. Subtracting 0.4 seconds aligns those three anchors. A cue from 1.4–3.4 becomes 1.0–3.0; duration stays two seconds. Recheck every cue for negative starts, reversed intervals and accidental overlap before accepting the change.
Reading load is a separate issue. An illustrative 42-character cue shown for two seconds is 21 characters per second; the same cue shown for three seconds is 14. Here the count includes spaces and punctuation but excludes line breaks and markup. These are calculated examples, not universal thresholds. Choose the delivery specification for the language, audience and platform; do not impose an English limit on every script.
Extending a cue into the next spoken sentence can create confusion. If there is no room, improve segmentation or revise the voiceover and edit. Do not silently remove a qualifier such as “up to” merely to make a product claim fit.
Ask for a change list before a new file
WebVTT's documented cue structure separates start, end and text. Opus's prompting guidance can help define the review output. Preserve the actual file format: WebVTT and SRT are not interchangeable by renaming the extension.
Audit this caption file against the supplied final video and approved transcript.
Use these manually checked speech anchors: 1.0 s, 5.0 s, 9.0 s.
Corresponding original subtitle anchors: 1.4 s, 5.4 s, 9.4 s.
First classify constant offset, progressive drift or local errors.
Propose changes in a table with cue ID, old/new start and end, text changes and reason.
Keep product names, numbers, negation and claim qualifiers intact.
Calculate duration and reading load using the stated counting method; ask for the delivery limit if absent.
Do not invent speech timings when audio is unavailable. Mark those checks unverified.
Preserve the input subtitle format and encoding. Return a revised file only after the change list, with no negative or reversed times.
Flag overlapping cues for review rather than deleting meaningful simultaneous speakers.Watch the delivered version twice
First listen with captions visible to check synchronization, wording and speaker changes. Then watch muted at the intended phone size: can the subtitle be read while the product action remains visible? Inspect line breaks around names and quantities, UI overlays, contrast and the closing CTA. Captions and promotional headline graphics serve different jobs.
Import the revised file into the real editor or player and export a short proof. Test the beginning, middle, end and every changed cue again. An internally valid timestamp file can still be attached to the wrong video revision. Keep video and caption version identifiers together in the delivery manifest.
Prepare the visual sequence in Panelly
Use Panelly Studio for the product scenes and accompanying campaign visuals, then finish timed captions in your video editor. The article does not assume Panelly performs Opus transcription or caption synchronization. For overall pacing see product storyboard timing; for a screen-based promo see the screenshot video workflow.
Should every word become a separate caption?
Not automatically. Word-by-word animation may suit a particular creative treatment, but it can fragment meaning and distract from the product. Compare a stable phrase-based version at delivery size before choosing the effect.
