video-render

name	video-render
description	Assemble the final short from visuals, timestamps, audio, and caption treatment, using the repo's Remotion stack without losing short-form visual intent.
allowedTools	["shell","read","write"]
model	inherit
argumentHint	{"visualsPath":"output/visuals.json","timestampsPath":"output/timestamps.json","audioPath":"output/audio.wav","outputPath":"output/content-machine/render/video.mp4","orientation":"portrait","fps":30,"downloadAssets":true}
entrypoint	node --import tsx scripts/harness/video-render.ts
inputs	[{"name":"visualsPath","description":"Path to visuals.json.","required":true},{"name":"timestampsPath","description":"Path to timestamps.json with word-level timing.","required":true},{"name":"audioPath","description":"Path to generated or source voiceover audio.","required":true},{"name":"outputPath","description":"Path that will receive the rendered MP4.","required":false},{"name":"captionPreset","description":"Optional caption preset override.","required":false}]
outputs	[{"name":"video.mp4","description":"Final rendered short."},{"name":"render.json","description":"Render metadata."},{"name":"captions.remotion.json","description":"Remotion-compatible caption export."},{"name":"captions.srt","description":"Plain SRT caption sidecar."},{"name":"captions.ass","description":"ASS karaoke-style caption sidecar."}]

Video Render

Use When

• The user already has visuals.json, timestamps.json, and audio.wav and wants the final video assembly step.
• The main question is no longer script or visual selection but how the final short should feel once motion, timing, and captions are combined.
• You need to render a real MP4 while preserving caption intent, safe zones, and short-form pacing.

What This Skill Owns

• Final composition decisions in collaboration with short-form-production-playbook.
• Final caption treatment selection in collaboration with short-form-captions.
• Composition choices such as orientation, pacing feel, overlays, and safe render defaults.
• Final MP4 quality hygiene: clean sources, readable captions, correct dimensions, and reviewable metadata.

Inputs

• visuals.json
• timestamps.json
• audio.wav
• optional caption preset or caption overrides
• optional overlays, audio mix, browser override, or template override

Render Approach

1. Confirm the output target is portrait 9:16, normally 1080x1920, for TikTok, Instagram Reels, and Shorts.
2. Confirm the visuals are caption-clean. Rendering is not the place to hide pre-existing text burned into source footage.
3. Confirm music, SFX, ambience, or extracted audio already has source evidence and Content ID risk notes. Rendering is not the place to turn unclear YouTube downloads into a public demo asset.
   • If source/audio evidence is missing, stop and route back to source-media-review or audio-source-policy.
4. If the edit still needs judgment on pacing, hook support, shot variation, or assembly choices, read short-form-production-playbook before touching render options.
5. Choose the closest caption family from short-form-captions.
6. Keep layout readable before adding more styling. Chunk size, line count, and placement matter more than decorative motion.
7. Keep caption overlays inside the practical social safe box: x=96..984, y=220..1540 on a 1080x1920 render.
8. For animation-heavy cards, diagrams, SVG/HTML assets, or Remotion scene work, use motion-design-coder before approving the render plan.
9. Use the repo caption presets as the baseline and then override only the specific fields the brief actually needs.
10. Render and inspect the actual MP4. Do not trust config alone.

Split-Screen Rules

• For Reddit-story or gameplay-backed split-screen work, default to a true 50/50 vertical split unless the brief has a stronger reason not to.
• Honor the upstream fill plan from reframe-vertical: crop-fill, contained-blur, or fit-pad.
• Never let a default contain render create accidental plain black gutters. If containment is needed, use a designed/blurred fill behind it unless the format intentionally calls for black.
• For pure reddit-post-over-gameplay, render full-screen gameplay only with card/caption overlays. Do not create a top lane or request support clips.
• If captions belong on the seam between the two lanes, overlay them at the midpoint instead of inserting a dedicated caption band.
• Treat hook cards and opener assets as their own upper-band layer, not as ordinary subtitles.
• A thin divider, seam gradient, or center badge is a valid readability aid for split archetypes; use it as a reusable layout token, not ad hoc decoration.
• For OCR-reviewable exports, prefer full-line active-word highlighting over progressive reveal burns when you need stable rendered caption verification.
• If a fallback ASS burn is used, set captionAssStyle.maxCharsPerLine and captionAssStyle.maxLines. Never rely on raw positionY alone; ASS absolute positioning bypasses margins unless the export code clamps and wraps it.

Isolated Project Rule

• In fresh installed-pack projects, do not assume the renderer can fetch a browser on demand. If the environment already has a cached Remotion browser, pass browserExecutable and chromeMode explicitly in the render request.
• Treat that browser override as part of the portable render contract for offline or sandboxed harness runs.
• If the only path to success is a manual fallback render outside the shipped runtime, record that as a runtime gap instead of pretending the packaged path worked.

Technical Surface

• Main implementation: src/render/service.ts
• Caption stack: src/render/captions/*
• Optional runtime entrypoint when you want repo-side execution from a coding-agent CLI: node --import tsx scripts/harness/video-render.ts

Outputs

• one final video.mp4
• sibling render metadata, usually render.json
• caption sidecars when exportCaptions is enabled: captions.remotion.json, captions.srt, and captions.ass
• caption quality fields in the JSON result: captionQualityPassed and captionQualityScore
• optional review bundle if this skill is being used under a larger generate-and-review pass

Validation Checklist

• The output is a real playable MP4 with the expected aspect ratio.
• Caption sidecars exist when exportCaptions is enabled.
• Captions remain readable on a phone-sized viewport.
• Captions stay inside the 1080x1920 frame and social safe box.
• The active-word treatment feels synced, not merely mathematically aligned.
• Motion is frame-driven and deterministic; no browser-time animation clock or unseeded randomness drives the final Remotion render.
• Cards, diagrams, captions, and support footage do not compete for attention during the same beat.
• No source-text collision was introduced by the footage choice.
• The final frame composition still leaves room for platform chrome.
• Pure gameplay modes contain no unrelated video layers.
• Split-screen lanes use the planned fill strategy and do not show accidental gutters or cropped-off subjects.