Communitygithub.com

ai-video-review

Review and deliver a short video: per-shot frame checks, technical acceptance, six failure modes, independent review gate, and the full delivery checklist. Runs last in the ai-video-skills pipeline; red light means no delivery.

What is ai-video-review?

ai-video-review is a Claude Code agent skill that review and deliver a short video: per-shot frame checks, technical acceptance, six failure modes, independent review gate, and the full delivery checklist. Runs last in the ai-video-skills pipeline; red light means no delivery.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/LaomaiAI/ai-video-skills/tree/HEAD/skills/05-review

Ask in your favorite AI

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

Documentation


name: ai-video-review description: Review and deliver a short video: per-shot frame checks, technical acceptance, six failure modes, independent review gate, and the full delivery checklist. Runs last in the ai-video-skills pipeline; red light means no delivery.

ai-video-review: 审核与交付

When to use

The mixed film exists (picture plus voiceover plus subtitles). This skill is the last gate before delivery. Nothing ships while this skill reports red.

1. Frame checks

Watch the middle of every shot, the frames before and after every transition, and the head and tail of the film. For each shot check: text overflow, overlap, subject size, on-screen duration, and whether the picture matches the storyboard row.

Check the whole film against the six failure modes. Any hit is a defect to fix, not a style choice:

  1. Slide-flipping PPT: shots that only change text on a static page.
  2. Copying demos: composition, characters, or assets taken as-is from a sample.
  3. No concrete object: a shot with nothing specific to look at.
  4. Dead background: an empty frame with no secondary motion.
  5. Code drawing what real footage should show: a realistic scene faked with shapes.
  6. Obvious AI flavor: blue-purple gradients, generic glow, stock phrasing on screen.

Also check: one style and one palette for the whole film; frames from different shots must not look like they came from different films.

2. Technical acceptance (measured, not judged)

Run and record each check; a check that was not run is written as "not run".

  • Dimensions and frame rate match film.json (1920x1080 or 1080x1920, 30 fps).
  • Frame count equals duration times fps.
  • Full decode passes with no errors.
  • Audio peak within limits; mix at -14 LUFS.
  • Determinism: rendering the same timestamp twice produces the same frame.
  • Richness: a visible change every 2 to 3 seconds across the film.
  • Subtitles match the audio within the stated error bound.

3. Independent review gate

Before delivery, hand the finished film to an agent that did not take part in making it. That agent watches only the finished film and writes its problems into review/*.md (worst 3 problems with timestamps, scored against the craft rules). Fix, re-render, and re-review until every item scores 8 or higher, recording each round in qa/review_log.md.

Red light means no delivery. Yellow or green means deliver, with the known weaknesses written down.

4. Deliverables

  • final.mp4 (picture plus voiceover plus subtitles).
  • brief.md, script.md, storyboard.md, film.json, captions.srt, timeline.json.
  • Render and media check reports, the review log, and the exact commands to re-run the whole pipeline.
  • A short note: key creative decisions, known weaknesses, render time, and any uncertain facts listed separately for the user to check.

Rules

  • Do not deliver a film that has not passed independent review.
  • No upload, no deployment, no publish without release authorization.
  • A failed check cannot be covered by delivering an HTML page, a silent intermediate file, or a prompt instead of the film.

Individual skills in this repo

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

Related Skills