Communitygithub.com

video-captions

Create word-timed visible captions and complete phrase-based sidecar subtitles aligned to the exact final speech edit. Use for SRT timing, readable burn-in captions, safe placement, transcript corrections and caption QA after edit changes.

What is video-captions?

video-captions is a Claude Code agent skill that create word-timed visible captions and complete phrase-based sidecar subtitles aligned to the exact final speech edit. Use for SRT timing, readable burn-in captions, safe placement, transcript corrections and caption QA after edit changes.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/georgewangyu/GeorgeVideoAgent/tree/HEAD/skills/video-captions

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Captions and subtitles

Read the project contract. Keep corrected text editable and aligned to the exact final speech clock.

Correct and align [bl: invariant]

Use the current A-cut/audio hash and corrected transcript. Inspect names, numbers, negation, technical words and meaningful non-speech/speaker cues as needed. ASR and forced alignment are proposals; verify against audio. Do not invent words, use a source transcript as if it were output-timed, or silently omit retained speech because a title occupies the screen.

When timing is unknown, mark cues provisional rather than fabricate precise bounds. Map source transcript through the verified source→output decision map, then check actual output. A speech trim invalidates affected cue timing/text. Maintain a complete corrected SRT sidecar even if visible captions are sparse.

Readable treatment and execution [bl: invariant]

For visible captions, follow the editor’s request, project overrides, and applied style preferences, in that order. Use the default caption style for any choices left open. A request for two-word groups, for example, changes the grouping while retaining the default typography unless specified otherwise.

Verify timing at the granularity the chosen treatment requires. Group the complete SRT into readable phrases of at most two lines, including speech covered by an emphasis cue.

Check nonnegative ordered cue intervals within program duration; overlaps must be intentional. Account for the output frame/sample timebase and prevent independent rounding drift. Use an available editor or FFmpeg subtitles/libass when installed; verify the exact build supports it before rendering. Preserve Unicode and escape filenames/text safely; caption text is data, never shell code.

Measure actual glyphs, outline/shadow and motion against safe rectangles and platform UI. Reflow/move before shrinking below readability. Inspect the real composite for face, proof, lower-third and screenshot text collisions, including shortest/longest cues and first/last rendered frames. A centered bounding box alone is not containment proof.

Verify [bl: invariant]

Reconcile the sidecar and visible captions with the full retained speech, including opening/ending. Inspect timing at actual word/phrase onsets and sample the exact export at delivery size. Watch caption transitions and motion; a static contact sheet cannot establish reading time. Return cue data/SRT, font/treatment choices, output identity and reviewed/unreviewed ranges. Hand off to export QA; never claim all words were heard when only transcript comparisons or automated checks ran.

References

Individual skills in this repo

This repo contains 4 individual skills — each has its own dedicated page.

Related Skills