/review-video

name: review-video
description: Review a Remotion video with AI vision. Starts with a cheap local check of rendered still frames, then has TwelveLabs watch the full render and score it against a JSON rubric (pacing, composition, text legibility, audio sync and more) with timestamped fixes, applies the fixes and re-reviews. The feedback loop that closes the production circle.

Review Video — AI-Powered Video Feedback Loop

You are reviewing a Remotion video and providing professional feedback: first from a few rendered frames, then by having TwelveLabs "watch" the full render. This closes the production loop — preview, review, improve, repeat.

**Load the `remotion-production` skill** for the `video-analysis` rule. Rendering follows the official `remotion-render` skill.

**Rendering is allowed here.** Invoking `/review-video` is an explicit request to render for review, so you don't need to ask again.

Workflow

1. Quick Frame Check (cheap, local)

Before uploading anything, render a handful of still frames and look at them yourself, as official `remotion-render` shows:

npx remotion render [CompositionId] out/review-frames --frames=0,45,90,150,240 --image-format=png

Pick frames that cover the hook, each scene's midpoint, the busiest moments (captions, overlays) and the last frame. Read the PNGs and check:

• Blank or empty frames, missing assets (a missing `staticFile()` shows as nothing or an error)
• Text readability, overflow and safe areas (official `remotion-create/video-layout.md`)
• Caption placement against the platform UI (the `captions-workflow` rule)

If these frames show obvious bugs, fix them first (see step 6) and repeat this step. That's much cheaper than a full render plus a TwelveLabs upload.

2. Render the Full Video

Look for an existing render:

ls -t out/ 2>/dev/null | grep -E '\.(mp4|webm|mov)$'

If there's none, or the code changed since it was made, render it now:

npx remotion render [CompositionId] out/video.mp4

3. Upload the Render to TwelveLabs

create-asset:
  file: [absolute path to the render, e.g. /path/to/project/out/video.mp4]

Keep the returned asset ID. No index is needed. If the analyse call says the asset isn't ready, wait a few seconds and retry.

4. Analyze — The Director's Review

Run **one** `sync-analyse-video` call that returns the whole review as JSON (renders over 1 hour: `async-analyse-video` with the same `prompt` and `jsonSchema`, then poll `get-analyse-task`):

sync-analyse-video:
  assetId: [asset ID]
  modelName: "pegasus1.5"
  temperature: 0.2
  prompt: "You are a video director reviewing a [duration]s [platform] video at [WxH]. Intent: [one-line brief]. Script: [narration, per scene]. Score each dimension 1-10 and list concrete fixes with timestamps in seconds. Check: framing and composition; whether on-screen text is readable and on screen long enough; transitions; blank, frozen or glitchy frames; pacing (scenes too long or too short); whether the voiceover matches what's on screen; SFX timing; music mood and level under the voice; whether the first 3 seconds hook; whether the CTA is clear."
  jsonSchema: [the rubric below]

Always pass `modelName: "pegasus1.5"`: the server's default model has been removed from the API.

**Rubric (`jsonSchema`):**

{
  "type": "object",
  "properties": {
    "overall_score": { "type": "integer", "description": "1-10" },
    "scores": {
      "type": "object",
      "properties": {
        "hook": { "type": "integer", "description": "1-10, first 3 seconds" },
        "pacing": { "type": "integer", "description": "1-10" },
        "composition": { "type": "integer", "description": "1-10, framing and layout" },
        "text_legibility": { "type": "integer", "description": "1-10, size, contrast, time on screen" },
        "transitions": { "type": "integer", "description": "1-10" },
        "audio_sync": { "type": "integer", "description": "1-10, voiceover and SFX against visuals" },
        "audio_mix": { "type": "integer", "description": "1-10, voice clarity, music level" },
        "cta": { "type": "integer", "description": "1-10, clarity of the ending" }
      },
      "required": ["hook", "pacing", "composition", "text_legibility", "transitions", "audio_sync", "audio_mix", "cta"]
    },
    "working": { "type": "array", "items": { "type": "string" } },
    "fixes": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "start": { "type": "number", "description": "seconds" },
          "end": { "type": "number", "description": "seconds" },
          "dimension": { "type": "string", "enum": ["hook", "pacing", "composition", "text_legibility", "transitions", "audio_sync", "audio_mix", "cta", "glitch"] },
          "severity": { "type": "string", "enum": ["critical", "major", "minor"] },
          "issue": { "type": "string" },
          "fix": { "type": "string" }
        },
        "required": ["start", "end", "dimension", "severity", "issue", "fix"]
      }
    },
    "priority_fix": { "type": "string" }
  },
  "required": ["overall_score", "scores", "working", "fixes", "priority_fix"]
}

Save the JSON (e.g. `out/review-1.json`) so the next pass can be compared with this one, then clean up:

delete-asset:
  assetId: [asset ID]

Every review uploads a new asset, and assets count against the account's storage.

To look closer at one stretch (say a fix at 12-18 s), run `sync-analyse-video` again with `startTime`/`endTime` around it (the window must be at least 4 s) before deleting the asset.

5. Present the Review

Format the JSON as a professional director's review. Convert each fix's timestamps to frames (`Math.round(seconds * fps)`) and name the scene it falls in, so the fix points at code:

🎬 Video Review: [filename]
Duration: [X]s | Resolution: [WxH]

📊 Overall Score: [overall_score]/10
   Hook [x] · Pacing [x] · Composition [x] · Text [x] · Tr