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: