vanh

Builder and Content Creator

[vanh @VanhDesign]

@SouranyPhomhome thanks, it was fun creating it

[vanh @VanhDesign]

I've been testing Claude Opus 4.8, and here are 2 things I made

a WebGL ink-fluid you can smear with your cursor, and a physics card deck you can throw and tear

Here's where Opus 4.8 actually shines, from building with it:
WebGL shaders, fluid sims, particle systems, real physics
The biggest jump isn't speed, it's honesty. It flags its own bugs instead of papering over them.

Opus is great for coding, reasoning and also more expensive in terms of usage so use it where it really matters.

Will have a video on Opus 4.8 soon

[vanh @VanhDesign]

The new Dynamic Workflows in Opus 4.8 is so powerful, I've used it in many of my personal projects

Claude plans the work, then writes an orchestration script that fans the job out across parallel subagents. Up to 16 running at once, 1,000 total per run. They tackle the problem from different angles, check each other's work, and only the verified final result comes back to you. The agent also have adversary approaches to break down all implementation from other agents and push it further.

The important part for me: The plan lives in the script, not in Claude's context window. So your session stays responsive while the heavy lifting happens in the background.

The example Anthropic gave says it all. Codebase-scale migrations across hundreds of thousands of lines of code.

A few things worth knowing before you try it:
It's a research preview, on Team, Max, and Enterprise plans. Runs are resumable, so an interruption doesn't send you back to zero. And it burns through tokens faster than a normal session, so keep an eye on usage.
This is the shift a lot of people have been waiting for.

[vanh @VanhDesign]

Was playing with animation and this one turned out really clean and nice

Hexagons floating into a framework grid. Subtle, slow, just enough motion without being too much. What I love most is that it comes with the full DESIGN.md attached.

So the entire system, the typography, the color palette, the spacing, the grid behavior, is right there.

Drop it into any project and your AI tool builds in the same visual language without you re-explaining anything. This is what makes DESIGN.md so useful.

[vanh @VanhDesign]

@ErichZiebarth Thank you Erich, here is the full video link: https://t.co/86JUV9iMdD

[vanh @VanhDesign]

I just made over an hour long video on how to extract a DESIGN.md from any website you want to use as inspiration.

Drop a comment if you want the link and I'll send it over.

The problem:
How do you get a DESIGN.md out of a website that doesn't already have one?

So I built the workflow and recorded the whole thing. Three methods, stacked from easiest to most powerful:

1. Screenshot only (great when that's all you've got)
2. Live website using DevTools and computed styles (the real source of truth)
3. Fully automated with Claude Code (start to finish, including the Tailwind config export)

Each method teaches you how to catch the errors the next one builds on. By the end you have a portable DESIGN.md you can drop into any project, any AI tool, and stop re-explaining your design system in every chat.

[vanh @VanhDesign]

Here's a workflow to create UGC content all in one place.

I tested Dreamcut this week with a quick 6 second clip to see how the pipeline holds up. You can use this workflow for your own product.

The flow:
Generated the images and script in ChatGPT Images 2.0 (you can also generate images inside Dreamcut directly)
Brought them into Dreamcut and used Seedance 2.0 to turn the images into video
Edited and exported in the same canvas

What makes this different:
You can do all of it inside one app. Images, video, screen recording, video editing. Loom records. CapCut edits. You end up running multiple processes to ship one video.

So you can create contents really fast, if you're using either a screen recorder or video editor in your workflow, this is worth trying. I created this video using the actual tool

[vanh @VanhDesign]

This workflow is suprisingly good

ChatGPT image to Claude design:
Generate the visual inspiration in ChatGPT.
Feed the image straight to Claude.
Ask it to recreate the design.

Pretty accurate on the first try.
Cleaned up the rest with a few small follow up prompts.

ai images as inspiration could be a thing now

[vanh @VanhDesign]

@aksonvady You re welcome, the DESIGN.md architecture is such an important topic

[vanh @VanhDesign]

Google just open-sourced DESIGN.md.

Before this, every team had its own way to ship a design system to AI.

Some used design-tokens.json. Some used Style Dictionary. Some pasted a Figma export. Some wrote a custom markdown file. A lot of teams just hoped the AI would figure it out from a screenshot.

Result: every prompt, every team, every AI tool produced a different design system for the same brand.

DESIGN.md changes that. One open-source format that defines colors, typography, spacing, rounding, and components in a standard way.

Once Claude Design, Cursor, Lovable, and everything else read DESIGN.md the same way, your design system becomes portable. Move your file between tools and the output stays consistent. Share it with a teammate and they get the same result. Hand it to a different AI six months from now and the structure still holds.

Just shipped a full 34 minutes workflow video on how to use DESIGN.md with Claude Design in production. Link in the comments.

[vanh @VanhDesign]

full video: https://t.co/EhbTDKq8OA

[vanh @VanhDesign]

full article: https://t.co/piaBr8uuSm

[vanh @VanhDesign]

Crazy! If you have been wiring Claude into Blender or Photoshop with community MCP servers, today is a big day

Anthropic just shipped Claude for Creative Work. Official connectors, built with the actual app developers, for:
Adobe Creative Cloud
Blender
Ableton Live and Push
Autodesk
Fusion
Splice
SketchUp
Affinity by Canva
Resolume Arena and Wire

The Blender one was built by the Blender devs themselves. Anthropic also joined the Blender Development Fund as a corporate patron, funding the Python API the connector depends on.

The part that matters most for creators: it is all still MCP. Open standard. Your custom servers for Houdini, Cinema 4D, or whatever niche tool you use still plug in the same way.

Full breakdown in the comments or check out the newsletter

[vanh @VanhDesign]

Link: https://t.co/KX7lQQOqom

[vanh @VanhDesign]

This is big, Neuform is now featured inside Google Stitch

Click Start with a DESIGN.md, then From the web, and Neuform shows up right there in the list.

The workflow:
Pick a design you love on Neuform. Download the design.md. Upload it as reference in Stitch.
Your generated designs now inherit the tokens, typography, and component rules from a template you actually like.

Google open-sourced the DESIGN.md spec a day ago. It now works across any AI design or coding tool that adopts it.

The design.md format is about to change how every AI tool reads your brand. Neuform was already built around this format.

Link in the comments

[vanh @VanhDesign]

I've been building animated web components lately and the workflow is just so easy

Love to play with animations and interactions, grab a component you like. Remix it. Export the design system so you can use it anywhere.

Start with the components that has the styling you're looking for, then build the rest of the section and landing page around it.

[vanh @VanhDesign]

An open-weight model just beat GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro.

Moonshot AI shipped Kimi K2.6 today. Scores:
Kimi K2.6 = 58.6
GPT-5.4 = 57.7
Claude Opus 4.6 = 53.4

First place is open source.

A few specs worth noting:
1T parameters (MoE, 32B active per token)
256K context window
300 parallel sub-agents per run
4,000+ coordinated tool calls
12+ hours of documented autonomous operation API pricing: $0.60 / $2.50 per million tokens. Meaningfully cheaper than closed frontier models at the same tier.

The license is modified MIT. You only credit Kimi if your product hits 100M MAU or $20M monthly revenue. For almost every team reading this, it's effectively free commercial use.

What would you automate if you could let 300 agents work for 12 hours straight? I wanna know because I will def test it, stay tuned! (maybe not 300 agents)

Full deep dive on CreativeAINews. Link in comments.

[vanh @VanhDesign]

Link to the full video: https://t.co/6ZNeL56SSc Link to the tool: https://t.co/KX7lQQOqom

[vanh @VanhDesign]

Build and iterate UI components at lightning speed, and with style

You describe what you need, a hero section, a pricing card, a CTA block, and it generates fully working, production-ready code ready to drop into any project.

You can run 6 generations in parallel, remix community designs in a few clicks, and use Skills to inject curated styles into your prompts automatically.

The remix feature in particular gives you creative directions and animations that you genuinely wouldn't have thought of on your own, which makes it as useful for inspiration as it is for building.

I just published a full quick start walkthrough covering: The homepage, projects and how to organize your work The prompt bar, model selector and Skills feature The community page and how to remix any design Iterate vs Remix vs Remix with Prompt and when to use each

Link in the comments