Docs: a Douyin case study, a landscape variant for knowledge films, ideation prompts

[ZLHad]

What and why

Ten AI-made Douyin films were studied for what is worth learning: nine posted 2026-09-25 to 10-03, most tagged #vibe知识大赏, seven of them knowledge explainers, plus one April film as a contrast. The maintainer asked for four items: the reverse-engineering method, a case study, a Douyin landscape variant, and "a bit on ideation". They asked for no style restrictions and nothing else. At their request, the case does not name the creators or say how the films were found.

• cases/douyin-vibe-knowledge.md (new) covers the ten films: post date, likes (a 10-04 snapshot), length, frame and opening. It also records:
  • a few things the top four share, each labelled with which films do it;
  • caption sizes measured on two films (about 42–56 px at 1080p);
  • one creator's 15 films in a week, with likes from 2,246 to 427,108 and no attribution of the gap;
  • two films on the same topic (the 37% rule).
    It is linked from cases/README.md, type 02 and the CLAUDE.md route table (AGENTS.md regenerated).
• video-types/02-knowledge-short.md adds a "抖音横屏中视频（2–5 分钟）" variant:
  • start it with bin/vh new short <slug> --aspect 16:9;
  • it uses the desktop floor. It says why this departs from playbook/01's "landscape on Douyin = feed" mapping (measured captions; viewers going full screen is marked as an inference), and says when to use --watch feed instead. playbook/01 points to the variant.
  • feed-size hook text is an option;
  • it covers the beat sheet, captions, post copy and series titles.
• playbook/12-ideation.md: two prompt-table rows ("找一个历史现场", "交还给观众") and the real 37% pair.
• playbook/07-reverse-engineer.md: how to get Douyin frames without logging in:
  • short link to id and post time;
  • the mobile share page's description has the caption, date and likes;
  • the playwm URL plays as a media document;
  • the creator post-list request gives every film's likes.
    Also: X at 1080p, measuring caption size, one browser tab per agent.

How it was checked

• The data was gathered in the built-in browser on 2026-10-04. About 150 frames went into contact sheets, two openings were viewed second by second, and caption sizes were measured against a 10 px grid. No downloads; no frames committed.
• Three fix commits followed review. One restores two CHANGELOG entries from Review desk: tick, comment and submit each stop; the agent wakes on submit (bin/vh desk) #54 that a fix commit had dropped; git diff main...HEAD -- CHANGELOG.md now shows only added lines.
• An independent reviewer (fresh context) found one must-fix: the Watch-on floor contradicted playbook/01. It also found should-fixes on sample scope, causal claims and rule-like wording. All are fixed in the second commit, and a re-review was requested.
• tools/ci.sh --committed passes.
• A second independent review (approve with nits) led to one more commit: sample scope stated everywhere, landscape Captions / Safe box / Cover lines for the TYPE block, the oneshot-five count, one consistent paraphrase, an eyeballed size marked as such.