# lanshu-awesome-ai-video-kit: 543 AI video prompts, 15 model profiles, 7 Claude skills

> A text-first toolkit for people writing AI video prompts across vendors, with a required source field on every prompt, per-model formulas instead of one house style, seven Claude skills and a weekly endpoint monitor. The model table is where the value is and where the weakest claims sit, because its grades are assertions rather than measurements.

**cclank/lanshu-awesome-ai-video-kit** — 做企业 AI 视频项目逼出来的工具包 · 411 prompt · 15 模型 · 7 Claude Skill · 14 篇方法论

- Repository: https://github.com/cclank/lanshu-awesome-ai-video-kit
- Website: https://lanshu-awesome-ai-video-kit.lank.workers.dev
- Stars: 410 · Forks: 94
- Language: HTML
- License: MIT
- Published: 2026-09-15 · Updated: 2026-09-15 · Language: en
- Canonical page: https://hysenlabs.com/projects/cclank-lanshu-awesome-ai-video-kit

## Every prompt carries a source field, and 32 endpoints get walked each Monday

The kit attacks three named failures in prompt collections: entries with no author, formulas that stopped working on newer models, and house style smeared across vendors. Each failure gets a mechanism. Provenance is handled by a required `source` field on every prompt, linking back to official documentation or a review blog. Staleness is handled by `.github/workflows/model-version-monitor.yml`, which inspects 32 official endpoints every Monday at 09:00 Beijing time and opens an issue when a version changes. Mixing is handled per model, so Sora entries follow a Shot List, Veo entries follow an 8 element structure and Kling entries follow a 5 layer scheme, each taken from that vendor's own formula.

Two things follow for a reader. Any single line can be traced back to a named document, which is unusual in this genre. And the freshness signal lives in the issue tracker rather than in the commit log, so a quiet git history tells you nothing about whether the model table still matches the vendors.

## The counts are printed in three places and they already disagree

Version v0.9.0 is dated 2026-05 in the data block, which itemises 543 prompts split into 433 single model best practices and 110 cross model comparison rows, 15 models as 11 commercial flagships plus 4 open source or open source friendly entries, 7 Claude skills, 21 methodology SOPs, 29 scenario categories, 32 monitored endpoints, 3 web tools, and 55 or more files at roughly 640 KB. Those figures do not agree everywhere in the file. The hero image at the top is generated with the caption 16 Models and 543 Prompts, so the banner says sixteen where the tables say fifteen. Model tables carry their own stamp, a May 2026 version line, and the note that data is reviewed by hand monthly alongside the weekly walk.

Quote the data block rather than the banner, and expect to recheck it, since these numbers are duplicated in the image caption, the summary sentence and the data block at once.

## serve.py adds a Markdown redirect that python3 -m http.server never had

Getting the site running locally takes three commands. Clone the repository, move into it, and hand `serve.py` a port:

```bash
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/ afterwards and the repository's Markdown renders in the browser instead of downloading as text. That redirect is the reason the wrapper exists. Three differences from `python3 -m http.server` are named: a `.md` request is redirected to the viewer rendering, the UTF-8 charset is forced, and caching is disabled during development. There is no build step, no package manifest and no dependency install in those three lines, which suits a kit that is mostly text. Port 8000 is the only argument the quick start shows, so serving anywhere other than localhost is not covered, and the script itself is not opened up for other options.

## The skill installer is a hardcoded symlink loop, not a resolver

Each of the 7 skills is its own directory under `skills/` with YAML frontmatter, written to the Anthropic SKILL.md format. Installation means linking those directories into your Claude skills folder:

```bash
for s in seedance-prompter seedance-storyboard seedance-debugger \
         happyhorse-prompter kling-prompter \
         model-selector prompt-translator; do
  ln -s "$(pwd)/skills/$s" ~/.claude/skills/$s
done
```

Two of the seven carry the cross model work. `model-selector` answers which model to use across all 15, and `prompt-translator` rewrites a prompt written for one model into another model's formula, returning a field mapping table plus a lookup against the 110 row comparison matrix. The remaining five are narrower: structured Seedance prompting, splitting a plot into 3 to 5 storyboard shots, a debugger that sorts 12 classes of failure, and single model prompters for HappyHorse and Kling.

The loop spells out its seven directory names, so a skill added to the repository later is not picked up by running it again. Links land in your home directory rather than a managed location, and they outlive the clone.

## Model profiles grade Chinese, audio and physics with stars

Each of the 11 commercial models gets a row of vendor, signature strength, clip duration and three star columns for Chinese, audio and physical behaviour. What lands in those columns is judgement, not measurement. Kling 3.0 is marked S tier with 2 minute clips, 48fps 1080p and lip sync. Veo 3.1 takes the highest audio grade and a 148 second chained duration. Sora 2 takes the highest physics grade and carries a warning that its web and app were stopped on 2026-04-26, which is the one row in the table that admits a model has gone away.

The four open source entries are graded by licence instead: LTX-Video 0.9.7, Mochi 1 and CogVideoX are all Apache 2.0, Higgsfield Soul is marked partially open source. Hunyuan Video 1.5 is listed at 13B and 8.3B sizes with LoRA support and a note that it runs on an RTX 4090. No method is given for turning a generated clip into a star, so a graded row cannot be audited from this repository, and the commercial access terms behind the top eleven rows are not covered at all.

## The SOP set runs from document 01 to 21 and the README runs out at 18

The methodology directory holds 21 SOPs across nine themes, with the ninth reserved for collections that keep growing. Documents 01 to 08 cover general practice and the Seedance system: the base formula, an 8 element advanced structure, storyboard timing, an emotion externalisation table, a camera movement dictionary, constraint words, special character rules and a 12 question pitfall list. Documents 09 to 12 give Kling's three writing styles, a five model comparison, the Sora 2 Shot List and the Veo 3.1 eight elements. Document 13 collects six commercial models in one file, document 14 does the same for the four open source ones and adds a selection decision tree.

Heaviest is the last group, built from video rather than from documentation. Document 15 draws on 10 YouTube tutorials with more than 500K views, document 16 builds a Gemini Omni formula on five official Google prompting tips plus ten official templates, document 17 uses 14 HappyHorse tutorials, and document 18 uses 25 Kling channels with 6M views to cover a 5 element formula, a Constraint Sandwich and six shots per 15 seconds. The README text runs out inside that last table. Above it, the contents list already links sections on the monitor, the directory layout, contributing and the licence that do not follow, so check those files directly rather than trusting the contents page.

## Three single-file web tools and a demo on a personal Workers subdomain

Three browser tools ship with the kit, described as Liquid Glass style, zero dependency, single file HTML. A live demo runs at lanshu-awesome-ai-video-kit.lank.workers.dev, a Cloudflare Workers address rather than GitHub Pages, and the v0.9.0 homepage is described as a mesh gradient with a floating capsule navigation bar and both dark and light themes. The repository root holds `index.html`, `viewer.html` and `favicon.svg`, and `viewer.html` is what the Markdown redirect lands on. Everything else is spread across `assets/`, `docs/`, `scripts/`, `tools/`, `methodology/`, `prompts/` and `skills/`.

Applying one visual style across three separate single file pages means maintaining that look three times over, which is the price of the zero dependency rule. The hosted demo is the other loose end: it is a deployment of its own on a personal subdomain, and the project does not say who redeploys it or how often.

## No GitHub releases, so the version number lives in README.md and CHANGELOG.md

Two facts frame the maintenance picture. There are no GitHub releases on this repository, which means the v0.9.0 marker is maintained by hand in `README.md` and `CHANGELOG.md` rather than produced by tagging. And the last push landed on 2026-06-01, while the data block it describes is dated 2026-05.

Neither fact means the project stopped. The weekly endpoint walk opens issues instead of committing, and the monthly review is described as manual, so a quiet commit log is exactly what the advertised mechanism would produce. It does mean the only freshness evidence available to a reader is the issue tracker and the dates printed in the tables. Licence is MIT, with `CONTRIBUTING.md`, `RESOURCES.md` and an `awesome.md` submission draft sitting beside the kit.

## Conclusion

Adopt this kit if you write AI video prompts across several vendors and want one place to see each vendor's own formula, because the per-model collection and the source field are the parts nothing else here duplicates. Skip it if you need access instructions, billing details or anything executable, since the kit carries none of that, and read its model grades as editorial claims rather than measurements. Verify first that the endpoint behind a prompt is still the one listed: 32 endpoints are inspected on Mondays, the last push to this repository was 2026-06-01, and one row already records a model whose web and app stopped on 2026-04-26.

## FAQ

### How do I run the lanshu-awesome-ai-video-kit site locally?

Clone the repository, run `python3 serve.py 8000` from its root, then open http://localhost:8000/. The script redirects `.md` requests to the viewer rendering, forces the UTF-8 charset and disables caching, none of which plain `python3 -m http.server` does.

### How does the kit decide whether a prompt is still current?

A GitHub Action inspects 32 official endpoints every Monday at 09:00 Beijing time and opens an issue when a version changes. The prompt set itself is reviewed by hand once a month.

### Does the model table still cover Sora 2?

It carries a warning that the web and app were stopped on 2026-04-26, alongside a 25s Pro duration. The kit documents no alternative access route for that model.

### Which models in the kit can run on hardware I own?

Hunyuan Video 1.5 is listed at 13B and 8.3B sizes with LoRA support and noted as runnable on an RTX 4090. In the open source group, LTX-Video 0.9.7, Mochi 1 and CogVideoX are all Apache 2.0.

### How are the 7 Claude skills installed?

A bash loop runs `ln -s` for each of the seven skill directories into `~/.claude/skills/`. The names are written into the loop, so a skill added later in the repository is not linked by rerunning it.

## Sources

- [cclank/lanshu-awesome-ai-video-kit on GitHub](https://github.com/cclank/lanshu-awesome-ai-video-kit)
- [Issues](https://github.com/cclank/lanshu-awesome-ai-video-kit/issues)
- [License: MIT](https://github.com/cclank/lanshu-awesome-ai-video-kit/blob/main/LICENSE)
- [Project website](https://lanshu-awesome-ai-video-kit.lank.workers.dev)
- [README](https://github.com/cclank/lanshu-awesome-ai-video-kit/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/cclank-lanshu-awesome-ai-video-kit