lanshu-awesome-ai-video-kit: a prompt library with a version monitor attached

What problem the kit actually solves

Every vendor ships its own prompt grammar. Sora wants a shot list, Veo wants eight elements, Kling wants five layers. When a team runs several models in one production, that grammar knowledge lives in scattered notes and in whoever last read the vendor PDF. The kit's stated premise is that most "AI video prompt collections" fail in three ways: the source is unmarked, the entries go stale, and the formulas from different vendors get mixed together.

The repository addresses each one structurally. Prompts carry a source field that links to official documentation or a review post. Each model gets its own section written to that vendor's published formula rather than a generic template. And a GitHub Action is described as checking 32 official endpoints every Monday, opening an issue when a version changes.

The audience is narrow and identifiable: people running AI video projects inside a company, working in Chinese, and switching between commercial and open-weight models in the same pipeline. The README calls the kit something that was forced out of enterprise AI video work, and the structure reads that way. It is a working reference, not a tutorial for someone who has never generated a clip.

How the content is organized: prompts, skills, methodology

The README reports 543 prompts split as 433 single-model best practices plus 110 entries in a cross-model comparison matrix, spread across 29 scene categories such as product, dialogue, physics, image-to-video, wuxia and pets. The 15 models divide into 11 commercial and 4 open or open-friendly. Commercial entries include Seedance 2.0, Kling 3.0, Veo 3.1, Sora 2, Runway Gen-4.5, Pika 2.5, Hailuo 02, Hunyuan Video 1.5, Wan 2.7 and 即梦 AI. The open side lists LTX-Video 0.9.7, Mochi 1, CogVideoX and Higgsfield Soul / DoP.

On top of the prompt files sit 7 Claude Skills and 21 methodology SOPs. Two skills are described as cross-model core: model-selector and prompt-translator. The methodology directory holds documents like 13-六大模型公式速查.md, a formula cheat sheet for six models.

Three web tools are shipped as zero-dependency single-file HTML, which is why the repository's primary language is HTML. A viewer.html renders Markdown, and index.html is the landing page shown in docs/screenshots/home.webp. The whole tree is described as 55+ files at roughly 640 KB.

Installing it and running a first prompt lookup

There is no package to install. The README's five-second start is a clone plus a local server. The custom serve.py script is worth using over the stdlib server because it redirects .md requests to the viewer, forces the UTF-8 charset, and disables caching during development.

git clone https://github.com/cclank/lanshu-awesome-ai-video-kit
cd lanshu-awesome-ai-video-kit
python3 serve.py 8000
# open http://localhost:8000/ in a browser

The Python badge in the README lists Python 3.11+, so check your interpreter version before running serve.py. Once the page loads, the landing page is the entry point to the prompt categories.

The machine-readable half of the library is prompts/data/all-prompts.json, which the README references for the 29-category count. If you want to pull prompts into your own tooling rather than browse them, read that file directly instead of scraping the HTML. The schema is not documented in the README, so inspect a few records first to confirm the field names you will depend on.

For model selection, the README points to skills/model-selector/SKILL.md, a Claude Skill rather than a script. Using it means loading the skill into Claude Code, not calling a CLI. The README gives no command for invoking a skill, so treat the skill directory as the source of truth for how it is wired up.

The weekly endpoint monitor and where it stops

The most distinctive piece is .github/workflows/model-version-monitor.yml. The README states it checks 32 official endpoints every Monday at 09:00 Beijing time and opens an issue automatically when a version changes. That is a reasonable answer to the staleness problem in prompt libraries: prompts do not decay on their own, the models underneath them do.

What the workflow can detect is a version bump at a URL. What it cannot do is rewrite the affected prompts. A new model release still means a human reads the diff, decides whether the eight-element formula changed, and edits the prompt entries. The README does not describe any automatic prompt regeneration, and the version table is labeled as a May 2026 snapshot with monthly manual review on top of the weekly check. So the monitor is a triage signal, not maintenance.

There is also a coverage gap worth naming. The 32 monitored endpoints cover commercial vendors with stable documentation pages. The four open-weight models in the table are versioned through model cards and repositories, which move differently. The README does not claim the monitor covers them in the same way.

Where the kit is the wrong tool

The documentation is Chinese-first. README.en.md exists, but the prompt content, the methodology SOPs and the skill descriptions are authored in Chinese, and the README presents the Chinese file as the default. A team that needs English prompt text to hand to a client will be translating constantly, and prompt-translator exists precisely because that is a real step in the workflow.

Two entries in the model table also argue against treating the kit as a stable dependency. Sora 2 carries a warning that the web and app surfaces stopped on 2026-04-26, and Higgsfield is listed as only partially open. A library that tracks vendor surfaces will inherit those discontinuities.

Finally, the kit is not an API, an SDK or a runtime. Nothing in the repository generates video. If what you need is a programmatic client for one vendor, this is the wrong layer entirely. It is reference material plus a local viewer, and the README does not present it as anything more.

The maintenance signal is mixed in one specific way. The repository is not archived, and the last push was on 2026-06-01. That is more than three months before today, which puts the weekly-monitor claim and the actual commit history in slight tension. The Action may well still be running on schedule; the content snapshot is what has not moved since June.

How it differs from a general awesome-list

The obvious alternative is an awesome-list of AI video tools. Those are curated link collections: a row per project, a one-line description, no prompt text. The difference in approach is that this kit stores the artifact rather than a pointer to it. A link list tells you Veo exists. This repository tells you what an eight-element Veo prompt looks like and where that formula came from.

A second alternative is a vendor's own prompt guide. Those are authoritative for one model and silent about the others. The kit's value proposition is the cross-model matrix: 110 entries that compare how the same intent is expressed across models. That comparison is the part a single vendor will never publish.

The cost of the choice is maintenance burden. A link list survives a model deprecation with one dead row. A prompt library with 543 entries and per-model formulas has to be re-checked whenever a formula changes, which is exactly the work the weekly Action is meant to surface and a human still has to do.

Licence, forks and what to verify

The repository is MIT licensed, which permits commercial use and modification with attribution and without warranty. The README does not discuss third-party rights in the prompt text or in the screenshots and banners it embeds, and MIT covers the repository's own code and content as the licensor holds it. If you plan to redistribute the prompt library inside a commercial product, that is the question to raise with your own counsel rather than assume MIT settles it.

Upgrade cost is low by design. There is no dependency graph to reconcile: the web tools are zero-dependency single-file HTML, and the only script is serve.py. Pulling a new version is a git pull, and the risk is that prompt files you edited locally conflict. The README does not document a fork-friendly override path, so if you customize prompt entries, keep them outside the tracked files.

Verify three things before adopting. Read one prompt record in prompts/data/all-prompts.json to confirm the source field is populated the way the README claims. Read methodology/13-六大模型公式速查.md to judge whether the formula coverage matches the models you actually use. And check the model table against your own vendor accounts, because the table is a May 2026 snapshot and at least one listed surface has already been discontinued.

Editorial conclusion

Adopt it if your team already writes prompts per model and you keep losing track of which formula belongs to which release; the per-model structure and the weekly endpoint check are the parts that save time. Skip it if you need English-first documentation, a published API, or a package you can pin as a dependency. Before committing, open methodology/13-六大模型公式速查.md and one prompt JSON in prompts/data/all-prompts.json to confirm the schema and the source field match how your pipeline stores prompts.

Frequently asked questions

Official sources

1. cclank/lanshu-awesome-ai-video-kit on GitHub
2. Issues
3. License: MIT
4. Project website
5. README