Claude Code 技巧：用 Opus 5.5 生成视频前的自动化质检流程

Claude Code tip: before you post a video Opus 5.5 made to X, have it send out 𝟯 𝗿𝗲𝘃𝗶𝗲𝘄𝗲𝗿 𝘀𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 to flag problems by timestamp against 𝘁𝗵𝗲𝘀𝗲 𝟵 𝗰𝗵𝗲𝗰𝗸𝘀, and only bring it to you once all 3 score it 7 or higher.

Both videos I made last week went through 3 reviewer subagents. One took 3 rounds, and its scores went from 𝟱/𝟲/𝟲 𝘁𝗼 𝟴/𝟴/𝟴.

Opening → When X autoplays a video, frame 0 is the thumbnail. Does it open on a blank screen, a black frame or something still fading in? → Watch the first 3 seconds on mute. Can you tell what it's about?

Visuals → Only review frames pulled from the exported MP4. Something that looks fine in the preview can disappear in the render. → Record a demo that loops back and forth on a page, and half of what you capture plays in reverse. Does every clip you cut in play forward? → Is there a black frame or fade-out in the middle that makes people think it's over? → Does the last frame lead back into the first? X plays videos on a loop.

Content → Does every caption and number match the source material? Does a claim like "the first" have a source? → Are captions or key info sitting in the bottom 8%? X's mobile player covers that strip. → Shrink it to phone width. Can you still read the text?

Run all 3 on Opus 5.5 at high effort. After each round of fixes, have the same subagent review the new version. It remembers what it flagged last round and checks those first.

If the main session takes more than 5 minutes to fix and re-render, the subagents' cache expires. When they pick up again, every image from earlier rounds gets processed from scratch, and that counts more against your usage limits. So the prompt gives each of them a 1-hour cache.

For the sound, follow the quoted post. It shows how to measure it.

Send this whole post to Claude Code, and add this line at the end 👇

"Create 3 reviewer subagents in ~/.claude/agents, one for each group above: hook-reviewer, motion-reviewer and content-reviewer. Use model opus and effort high, and give them only Read, Glob and Grep. They review and never edit. In each one's frontmatter, write experimental: and put cacheTtl: 1h on the next line, indented. If my Claude Code is older than 2.1.248, or subagentPromptCacheTtl is set in my settings, tell me first. The first time, ask me where the video project is and which MP4 is the latest export. Install ffmpeg or Pillow first if either is missing. Every time I export a video after that: 1. Pull frames from the exported MP4 and stamp each one with its time (if ffmpeg has no drawtext, use Pillow). hook-reviewer gets frame 0 and the frames from the first 3 seconds. motion-reviewer gets contact sheets at 4 frames per second, plus the first and last frames. content-reviewer gets a full-size frame for every caption, a set at phone width, and the paths to the caption script and source material. 2. List issues as 'time · P0/P1/P2 · issue · fix', and have each one end with a score from 1 to 10. 3. Fix P0s first, then P1s, re-render, and save it as a new version (v1, v2…). Save each round's issues, scores and the 3 subagents' agent IDs in a review/ folder. 4. Use SendMessage with each agent ID to have the same subagent review the new version. Don't start new ones. They check last round's issues first. Once all 3 score 7 or higher, or after 3 rounds of review, stop and hand it to me. Start fresh subagents for the next video. Show me the files you'll create first, and don't write them until I confirm."