Latest update
Claude as director, Codex as illustrator, Seedance as animator: a Chinese creator's AI short-film playbook racks up 54,000 bookmarks in a day
First seen on X 25 hours ago@EianLu ♥ 23,97025 views
"Changing a few words of a prompt costs nothing, but regenerating video is billed by the second." That is the starting point of a long X article Chinese creator Eian (@EianLu) posted Nov. 5, walking through how he makes AI short films with Claude Code, Codex and Seedance. In a day it passed 24,000 likes, 54,000 bookmarks and 3.1 million views.
The setup splits the work across three models. Claude Code is the 'director': it writes the story, storyboard and every prompt, inspects each image and clip, and does post-production in ffmpeg and Python. Codex is the 'illustrator', drawing character sheets and the first and last frame of every shot, and ByteDance's Seedance is the 'animator' that turns those two frames into a few seconds of video. The human picks the direction, approves the art style and signs off on spending. He published the full director prompt, with rules like no drawing before the story is approved, a separate agent must review prompts and frames before any paid video job, and never top up credits. Codex is called via codex exec; Seedance via a Volcano Engine API key, at roughly ¥1.5/second for Seedance 2.5 at 720p and ¥0.5/second for 2.0 mini.
His rule: catch in text what you can, so it never reaches an image; catch in images what you can, so it never reaches video. He says the pre-submit reviewer caught a cup 13% bigger in one frame, a hat brim missing in one image and swapped left and right. Other tips: render at 720p and upscale to 1080p, trim or fix with code before reshooting, and have a fresh agent watch the final cut cold like an audience.
In a follow-up he wrote it hit 1 million views and 22,000 likes in eight hours. One reply called the tutorial "too hardcore." The Seedance endpoint in the article is Volcano Engine's Beijing region.
“Claude code/Codex+Seedance AI视频:从一个想法到成片的全流程 这篇文章只解决一个问题:怎么靠一套多模型工作流和实际操作过程中需要注意的事项,用尽量少的钱、尽量低的抽卡率,做出一条自己满意的AI视频。”
