Skill

"AI Agent Becomes a Director: How a GitHub project allowed Claude Code to automatically produce film-sensitive product videos"

If you ask AI to make a product promotional video, you may think of this process: Give AI a product introduction → AI generates several screens → […]

25 August, 2026159 Reading21 minutes of reading
article directories

If you ask AI to make a product promotional video, you may think of this process:

Give AI a product introduction → AI generates several pictures → match music → export videos.

But people who have actually done product videos will find that"Being able to generate videos" and "being able to make a decent product promotional video" are two different things.

The really difficult part of product video is not just animation, but how to organize product pages, lens language, visual design, rhythm, sound and post-production into a complete production process.

Recently seen open source projects video-shotcraftIt is precisely solving this problem.

It is not a simple video generator, but a "movie product video production skills library" for AI Agents such as Claude Code and Codex: it organizes a large number of mature lens movement methods into a reusable Shot Recipe, and then cooperates with Remotion implementation, complete video templates, sound effects materials and production processes, allowing AI Agents to not only "write code and animate", but to complete product videos according to a relatively complete set of film and television production methods.

As of now, this project has included 152 lens recipe cards, 209 dynamic previews, 209 sports styles, and a complete set of runnable promo templates。The GitHub repository currently has more than 6200 Stars and 500+ Forks.

video-shotcraft GitHub repository

1. What problems did it solve?

Traditional product video production can be roughly divided into several stages:

Product understanding → creativity → Styleframe → lens splitting → material collection → animation → sound effects → editing → rendering → QA

If any link is not done well, the final effect will be compromised.

For example:

  • The product screenshots are fake and the audience can see it at a glance;
  • The animation is very cool, but it does not highlight the product functions;
  • There is no rhythm between shots;
  • There are no stuck points in BGM and pictures;
  • SFX is added indiscriminately, and the sound becomes cheap;
  • All elements are moving, there is no "pause" or "breathing";
  • The fonts, colors and the Design System of the product itself are completely inconsistent.

So the real question is not:

"Can AI generate an animation?"

Instead:

Can AI understand the product and choose the appropriate lens language to string the entire video production process?

The idea of video-shotcraft is very interesting:Don't let the Agent invent animation from scratch every time, but give it a "shot language" that can be retrieved, combined, and reused.

2. Turn "video production experience" into a Shot Recipe

The core design of video-shotcraft is Shot Recipe.

The project currently has 152 lens formula cards, each card not simply telling the AI:

"Make a fly-in animation."

Instead, we structure the key information needed for a shot, including:

  • Lens use
  • mood and energy
  • Recommended duration
  • animation parameters
  • implementations
  • precautions
  • Existing Motion Demo
  • dynamic preview

This is actually doing a very important thing:

Turn the "motion design knowledge" that originally relied on the designer's experience into a knowledge base that Agents can call.

For example, a product function display lens may require:

Page appears → Cards fly in → Arrangement → Lens advances → Subject focuses → Information stays

If LLM were completely left to play on its own, it could easily appear:

Zoom + Rotate + Glow + Particle + Bounce + Blur + Camera Shake...

It looks "very AI", but it may not be advanced.

The role of Shot Recipe is to tell the Agent:

What movement syntax should be used for this scene and how the movement should be implemented.

This is actually very similar to a programmer calling a component library.

Previously:

Write animation from scratch.

Now it becomes:

Select a sports language from the lens library → adjust parameters → adapt product materials.

3. Why did it choose Remotion?

The underlying video implementation of video-shotcraft uses Remotion。

The core idea of Remmotion is to use React/TypeScript to describe videos.

This is very important for AI agents.

Many operations in traditional video software are:

Timeline → Drag → Adjust parameters → Preview → Adjust again.

For Coding Agents, a more natural way to work is:

Read the code → modify parameters → render → check → modify again.

Therefore, Remotion is very suitable as a video execution layer for AI Agents.

The Demo of video-shotcraft itself is a TSX implementation, and a large number of animations use standardized t(normalized progress) to drive.

Thus, a shot can be abstracted as:

Input material + animation parameters + time schedule → output screen

This makes the lens a programmable asset.

4. The most noteworthy point: the real product page, not the AI hand-rubbed UI

This is one of the most valuable design principles of video-shotcraft.

The project clearly emphasizes:

Real screenshots must be used to reproduce existing pages.

In other words, if the video is introducing a SaaS product, the AI should not be allowed to redraw a Dashboard that "looks like" itself.

The correct way to do this is:

Real product page → browser automation → high-definition screenshots → element-level materials → Remotion mirror

The project's production process will start the local product development server, then use a headless browser to collect pages, and further obtain page texture and element position data.

This has a very obvious benefit:

The product in the video is really the product that the user is using.

This will greatly increase the credibility of product videos.

5. 2.5D: Why can product screenshots make a "movie feel"?

If there is only one webpage screenshot, the common practice may be:

Zoom in → zoom out → move left and right.

It will soon turn into a PowerPoint.

The video-shotcraft method is closer to 2.5D photography.

You can understand a web page as many different visual layers:

Camera
  │
  ├── Background
  │
  ├── Page
  │    ├── Header
  │    ├── Content
  │    └── Cards
  │
  └── Foreground

Then pass:

  • Camera Zoom
  • Pan
  • Orbit
  • Perspective
  • Depth
  • Parallax
  • Rotation

Give the originally flat page a sense of space.

For example:

The camera first sees the entire page → slowly advances → the main area enters the center of vision → the surrounding content produces slight parallax → finally stops at the core functions of the product.

At this time, what the audience sees is no longer a "webpage screenshot", but more like a camera is shooting a real product space.

6. The true "movie feeling" is not to add more special effects

There is one sentence in the design principles of the project that deserves attention:

The sense of film comes from the combination of mirror movement, light and shadow, rhythm and sound, rather than dazzling animation.

This is an easy pit for many AI videos.

It's easy for AI to think:

Advanced = more animations.

So:

  • Every element flies;
  • Every element shines;
  • Play every word;
  • Add Blur for each transition;
  • Every camera Shake.

The result is often the opposite.

video-shotcraft emphasizes:

One shot only talks about one dynamic effect

For example, if a shot selects "fly in", then fly in is the protagonist of this shot.

Instead of:

Fly in + spin + glow + particle + bounce + Shake.

This restraint is actually very close to the traditional Motion Design approach.

The sense of advanced quality often comes from "what to do less" rather than "what can be added."

7. Another key: Let the picture "stop"

AI has a natural tendency to generate animations:

Keep moving.

Because for models,"movement" seems to mean "content".

But professional video requires a lot of Hold.

Video-shotcraft has very clear rules for this:

  • It takes at least about 1 second to Hold after the brand sign is settled;
  • Leave a still time after the batch animation ends;
  • Enough time should be given for the opening main action;
  • Reserve a break frame in advance on the timeline.

The reason is simple.

If the picture keeps moving:

Audiences don't have time to understand the information.

The real product video rhythm should be:

Action → settle → understand → next action

Instead of:

Action → Action…

So "pause" itself is also a design.

8. BGM is not just background music, but a Timeline

Another very interesting part is video-shotcraft's handling of Beat Sync.

The so-called "stuck points" of many videos are actually just:

Animation every few frames.

This is not a real music jam.

The idea of the project is to first analyze the music itself, including:

  • BPM
  • Phase
  • Kick
  • Snare
  • Hi-hat
  • transient

This information is then mapped to the video Timeline.

In the end, the camera time is no longer simply written:

0s
2s
4s
6s

And closer to:

beatF(1)
beatF(2)
beatF(4)
beatF(8)

This allows:

Lens switching, element appearance, Camera Move, Impact, SFX

It really happens at the rhythm node of the music.

The project also proposed the idea of remeasuring the tangent point error after rendering, with the goal of controlling the error within a small frame number.

This actually explains one thing:

Video rhythm can also be engineered.

9. Sound design: SFX is not a "last-minute addition"

Many automatically generated videos ignore sound.

In fact, the same picture:

no sound

and

Precise Impact / Riser / Sparkle / UI Click

There are completely two feelings.

Video-shotcraft even classifies SFX by scene and material.

The current project contains 149 SFXs and follows:

  • transition
  • impact
  • riser
  • camera
  • UI
  • text
  • paper
  • film
  • light
  • data
  • sci-fi
  • mech
  • glass
  • fluid
  • crowd
  • counter

and other categories of organization.

This classification method is very similar to a designer's own "sound dictionary".

For example:

UI element appears → UI sound
Paper sound
Lens impact → impact
Light effect appears → spark/ light
Lens advance → rise

So sound also became part of the lens language.

10. The 152 lens cards are just "vocabulary" and not a complete lens division

This is a very important point in understanding this project.

Shot Recipe does not equal a full video.

It's more like:

Words in film language.

For example:

  • Fly-in
  • Stack
  • Orbit
  • Zoom
  • Flip
  • Reveal
  • Spotlight
  • Parallax

These are "words".

The real video is:

Words → Sentences → Paragraphs → Stories

So video-shotcraft also provides a complete Production Pipeline.

It wants the Agent to understand the product first:

产品是什么？
      ↓
最值得展示什么？
      ↓
视觉方向是什么？
      ↓
哪些功能适合哪些镜头？
      ↓
完整 Storyboard
      ↓
素材采集
      ↓
动画实现
      ↓
声音设计
      ↓
渲染
      ↓
QA

This is a level higher than simply "letting AI write Remotion code."

11. It actually provides three creative models

The project did not force all videos to follow the same process, but designed three models.

1. Ink Press templates

This is the most suitable route for quick release.

The project provides a complete set of verified promo templates:

  • 36.2 seconds
  • 1920 × 1080
  • 30fps
  • 10 shots
  • 2.5D page mirror operation
  • Title Card
  • transition
  • SFX

Users only need to tell the Agent:

Use video-shotcraft's Ink Press template to make a promotional video for my product.

Then replace:

  • product screenshot
  • Product Copywriting
  • Logo
  • brand colors

You can quickly get a fully structured video.

2. Independent and free creation

If you don't want to use a fixed template, you can let the Agent decide for himself:

  • video structure
  • visual direction
  • Lens selection
  • Material processing
  • audio
  • rhythm
  • finally realize

This is more like:

AI director mode.

Users give products and goals, and Agents complete the entire creation chain by themselves.

3. co-authoring

If you want to participate in creative decisions yourself, you can adopt a co-creation model.

The Agent stops at several key nodes:

Product Briefing → Requirements Decision → Visual Direction → Shot Mapping → Storyboard

and then confirmed by someone.

After that:

Material collection → Removal implementation → SFX → Render → QA

Then hand it over to the Agent for execution.

This actually solves a very real problem:

AI not all decisions should be made by itself.

Creative directions and business goals often require human participation.

Specific execution can be highly automated.

12. Gallery is another very clever design for this project

It would be very painful to ask users to directly read 152 Markdown lens cards.

So the project provides an online Gallery.

Users can:

  • Browse the camera;
  • View dynamic Preview;
  • search;
  • Filter by category;
  • View different variants;
  • Copy the Shot Card name.

Online Motion Gallery

This design actually puts:

"What kind of animation do I want?"

becomes a question of visual choice.

Users do not need to describe:

"I want one to have a little 3D spatial feel, then quickly enter from the right, and have a slightly slow-moving card stack..."

Just need:

"I want this."

Then hand over the Shot Card name to the Agent.

This is very important for people to collaborate with AI.

13. From "Prompt Engineering" to "Asset Engineering"

I think the most noteworthy thing about video-shotcraft is actually not the 152 lens cards.

But the method behind it:

Don't just optimize Prompt, prepare AI's callable assets.

Traditional AI workflow:

Prompt
  ↓
LLM
  ↓
代码
  ↓
结果

And video-shotcraft is closer:

Product
   ↓
Knowledge / Rules
   ↓
Shot Recipes
   ↓
Reference Implementations
   ↓
Assets
   ↓
Agent
   ↓
Remotion
   ↓
Render
   ↓
QA

This is a typical Agent Skill Engineering。

AI's capabilities no longer rely entirely on the model itself.

Instead:

Models + tools + knowledge + templates + reusable assets + acceptance rules

Together form the ultimate capability.

14. Why may this model be more practical than "AI one-click generation video"?

Many AI video tools now emphasize:

Enter a sentence → Output a video.

This approach is very suitable for:

  • Creative exploration;
  • Concept video;
  • social media content;
  • Atmosphere films;
  • Unstructured content.

But product videos have one special requirement:

Must be accurate.

The product page cannot be changed at will.

The Logo cannot change.

Functions cannot be compiled.

Buttons that do not exist cannot appear in the UI.

Words cannot be scribbled.

So product videos are actually more suitable:

Agent + Code + Real Assets

Rather than relying entirely on:

Text → Generative Video

The route of video-shotcraft is essentially:

Let AI be responsible for creative and engineering organization, let real product assets be responsible for factual accuracy, and let Remotion be responsible for deterministic picture generation.

This is especially valuable for SaaS, developer tools, desktop software, and AI products.

15. You can also see a bigger trend from this project

If video-shotcraft is put into the entire AI Agent ecosystem, it may not represent just "AI makes video."

But in a larger direction:

Agent Skill is changing from a "prompt word" to a "professional workflow"

A mature Agent Skill may include:

Skill
├── Rules
├── Knowledge
├── Templates
├── Recipes
├── Components
├── Assets
├── Tools
├── Examples
└── QA

This is very similar to traditional software engineering.

In the past, we would give programmers:

API + SDK + Component Library

Now you can also give agents:

Skill + Recipe + Template + Asset Library + Toolchain

Then let the Agent build on these capabilities to complete the work.

Video-shotcraft is a very intuitive case.

16. How to start using it?

The easiest way is to directly give the warehouse address to Claude Code or Codex:

Install this skill for me:
https://github.com/Vincentwei1021/video-shotcraft

You can also use skills CLI:

npx skills add Vincentwei1021/video-shotcraft

Or directly Clone:

git clone https://github.com/Vincentwei1021/video-shotcraft.git

cd video-shotcraft

Then establish a Skill link based on the Agent used.

After the installation is complete, you can directly ask for requirements like this:

Use video-shotcraft to create a promo for my desktop product.

Or specify a lens:

Use the deck-deal-flyin and row-embed shot cards
to present this feature.

You can even say directly:

Use video-shotcraft to make a promo for my product
with the Ink Press template.

The complete installation and use methods given in the project README are also relatively clear.

17. What scenario is it most suitable for at present?

Judging from the positioning of the project itself, it is most suitable for:

SaaS products

For example:

  • AI tools
  • project management software
  • CRM
  • data analysis platform
  • collaboration tools

developer tools

For example:

  • IDE
  • API platform
  • CLI
  • DevTool
  • AI Coding products

desktop application

It is especially suitable for products with a large number of real UIs and need to highlight the interactive experience.

Product release video

For example:

New Feature Launch
Product Update
Version Release
Product Demo
Landing Page Hero Video

Moreover, a single Shot Recipe itself is not limited to product promotional videos, but can also be applied in:

  • Brand Film
  • Feature Demo
  • Launch Video
  • Social Video
  • UI Motion

Wait for the scene.

18. What I think is really valuable

If you only think of video-shotcraft as:

"A Remotion project with 152 animations"

Actually underestimated it.

What's really interesting about it is breaking up an area that used to rely heavily on human experience into a production system that can be executed by agents:

设计经验
   ↓
镜头配方
   ↓
代码实现
   ↓
模板
   ↓
素材库
   ↓
声音设计
   ↓
生产流程
   ↓
自动化 QA

This is a very typical one:

Knowledge → Workflow → Agent → Production

path.

What AI Agents really change in the end may not be "allowing AI to do more things", but:

Gradually turn the implicit workflow in a professional's mind into an explicit system that machines can execute.

Conclusion: AI videos in the future may not be "generated" but "produced"

In the past, when we discussed AI videos, we often focused on one question:

"Which model generates the most realistic video?"

But the really important question for the product team may be:

"Can AI start from my product and complete a video that can be released?"

These two issues are completely different.

The former focuses on Generation。

The latter focuses on Production。

Video-shotcraft chose the second route.

Instead of trying to get AI to imagine a beautiful video out of thin air, it gives the Agent a set of lens language, real product materials, Remotion implementations, sound assets, templates and quality standards.

Eventually let AI move from one:

"Animation Code Generator"

Gradually become:

"Product video production assistant who understands camera language."

This may also be a very noteworthy development direction for Agent Skills:

The future of professional AI is not necessarily a larger model, but may be a carefully designed "model + workflow + tools + assets + acceptance system."

And video-shotcraft provides a very specific and beautiful case.

Project homepage: Vincentwei1021/video-shotcraft