An awesome list that audits its own hype

What it does

This is an awesome-list of 394 source posts showing videos, animations and workflows made with Claude 5.5 models (Opus, Sonnet, Haiku, plus creator-reported Fable testing). Every entry links back to the original post, records which model did what, and notes the evidence behind each claim. It also catalogs production routes — code-drawn 2D, Manim explainers, WebGL, footage editing, external video models — and points to four open-source production systems you can actually reuse.

The interesting bit

Most AI video roundups are screenshots with vibes. This one keeps a unified CSV of every batch, runs nine-frame preview sampling on 194 posts, and explicitly states what it cannot verify: backend identity, full motion, costs, original music. The “What we learned” section even quotes a creator admitting his tower sequence is invented, not engineering-calculated. That level of self-skepticism in a hype-prone genre is rare.

Key highlights

• 394 distinct source posts, browsable by model: 282 Opus, 45 Sonnet, 15 Haiku, 65 Fable (creator-reported, including suspected routing)
• Seven documented production routes, from ant-colony Canvas animations to Opus directing the Seedance video model
• Four linked open-source toolkits: Lemo-Opuscar, Papermotion, ClaudeAnimationBase, and a motion-graphics music video skill
• Real workflow data: repeated rejection of intermediate results, five-version iteration records, and per-task cost comparisons
• Frozen snapshot of 1,401 files with classification stats (games/ads/AI-about-AI at 57.7%)

Caveats

• Fable entries are creator reports, not verified — the README itself flags “suspected routing”
• Nine-frame sampling can’t confirm full motion or which model actually ran; original MP4s aren’t distributed
• Several deep-dive docs (methodology, prompt playbook) are Chinese-language only

Verdict

Worth a browse if you want honest, source-linked examples of what LLM-driven video workflows look like in practice — and reusable tooling to try one yourself. Skip it if you want verified benchmarks; the collection is upfront that it isn’t that.