Skill Captions

Skill-字幕

Separate subtitle timing from subtitle appearance:

• Let ra-audio-to-subtitles own ASR, phrase grouping, timestamps, SRT/VTT, and caption-qc.json.
• Let this skill own typography, panel geometry, fixed anchoring, preview, overlay rendering, burn-in, and render QC.
• Never estimate production timing from character count, scene duration, or TTS segment duration.

Required workflow

1. Lock the final narration audio or final merged video.
2. Run ra-audio-to-subtitles unless the project already has the canonical caption artifacts described in references/timing-contract.md.
3. Require caption-qc.json to report status: pass.
4. Select a style from references/style-registry.json. Use anchor-dark when the contract does not specify another style.
5. Generate a representative still preview before rendering a full video.
6. Inspect the preview against references/layout-qc.md.
7. Render the captioned derivative. Preserve the clean video and portable SRT/VTT.
8. Run scripts/validate_caption_render.py and archive its JSON report.

Registered styles

anchor-dark reproduces the approved Kimi K3 subtitle treatment: STHeiti Medium, warm white text, a tight charcoal translucent rounded panel, no text stroke, no text shadow, no large panel shadow, and a fixed bottom-center anchor. Read references/style-anchor-dark.md before changing its values.

anchor-light keeps the exact same typography, size, padding, width, bottom anchor, line fitting, and resolution scaling as anchor-dark. Only its surface changes: dark ink on a warm-white translucent panel with a fine border and subtle soft shadow. Read references/style-anchor-light.md before changing its values. This is the second registered style; anchor-dark remains the default.

Scale every 1080p metric by target_height / 1080. Render 4K captions at 2× their 1080p measurements; never upscale a rasterized 1080p subtitle layer.

Commands

Preview a QC-passed project:

python3 .claude/skills/skill-captions/scripts/render_captions.py \
  final.mp4 media/captions/captions.json \
  --qc media/captions/caption-qc.json \
  --preview qc/caption-preview.png \
  --preview-only

Burn the default style:

python3 .claude/skills/skill-captions/scripts/render_captions.py \
  final-clean.mp4 media/captions/captions.json \
  --qc media/captions/caption-qc.json \
  --out renders/final-captioned.mp4 \
  --preview qc/caption-preview.png

Burn the light translucent style:

python3 .claude/skills/skill-captions/scripts/render_captions.py \
  final-clean.mp4 media/captions/captions.json \
  --qc media/captions/caption-qc.json \
  --style anchor-light \
  --out renders/final-captioned.mp4 \
  --preview qc/caption-preview.png

Skip an interval that already contains burned captions:

python3 .claude/skills/skill-captions/scripts/render_captions.py \
  final-clean.mp4 media/captions/captions.json \
  --qc media/captions/caption-qc.json \
  --exclude 0:23.936 \
  --out renders/final-captioned.mp4

Create a reusable transparent caption band:

python3 .claude/skills/skill-captions/scripts/render_captions.py \
  final-clean.mp4 media/captions/captions.json \
  --qc media/captions/caption-qc.json \
  --overlay-out renders/caption-band.mov

Validate the result:

python3 .claude/skills/skill-captions/scripts/validate_caption_render.py \
  --source final-clean.mp4 \
  --render renders/final-captioned.mp4 \
  --captions media/captions/captions.json \
  --qc media/captions/caption-qc.json \
  --out qc/caption-render-qc.json

Hard rules

• Use the final audio/video as the only timing source.
• Consume captions.json start/end values directly.
• Refuse final burn-in when caption QC is missing or failed.
• Keep a fixed bottom anchor. Let multiline panels grow upward.
• Keep captions inside the frame and clear of titles, UI, faces, and PiP.
• Keep English model/product tokens intact.
• Preserve one canonical caption timeline across renderers.
• Keep the clean master; a burned-caption MP4 is a derivative.
• For a new or materially changed style, show a short preview before the full render.
• Do not add karaoke, behind-person typography, VFX captions, or animated hero words unless the user explicitly requests the separate embedded-captions lane.

Integration contract

Video orchestrators should call this skill after ra-audio-to-subtitles. Handoff contracts may set caption_style to anchor-dark or anchor-light; otherwise use anchor-dark. Renderers may reproduce the registered CSS in Remotion or HyperFrames, but must preserve the style metrics and fixed-anchor behavior.