Something about this week's AI updates feels a little more useful than another impressive demo.

We're seeing more tools that give us something we can actually work with afterward.

Editable vectors instead of flat images. AI assistants that can make changes inside real projects. Image models that think about layout before generating anything.

I like where this is going. Getting a good result from AI is one thing, but being able to take that result and make it your own is where things get interesting.

Here's what caught my attention this week.

Framer lets AI assistants edit your actual website

This one feels like a pretty significant step for web designers.

Framer introduced an official MCP server that lets ChatGPT, Claude, Cursor, and other compatible AI tools connect directly to your projects.

Once connected, an assistant can work with the canvas, update CMS content, create pages, and even publish changes.

The interesting part is that you don't need to keep the editor open or run a separate local connection.

If you maintain client websites or a portfolio, you can probably imagine a few ways this could save time.

I'd still be careful about letting an agent make changes to a live project. Framer supports working on branches, so there's an opportunity to review everything before publishing.

Read the full brief →

Ideogram can now generate actual vector artwork

This one caught my attention immediately.

Ideogram introduced Vector, a model that generates SVG artwork directly from a prompt or converts an existing image into a vector.

The idea is to produce artwork with editable shapes, colors, and gradients rather than a flattened image.

If you've ever tried converting AI-generated artwork into something usable in Illustrator, you know why this matters.

I'm particularly interested in the image-to-vector option. Being able to turn a rough illustration or low-resolution graphic into something editable could be useful.

Of course, there's a difference between an SVG file and a good SVG file.

I'd want to see how clean the paths are, whether elements are grouped logically, and how much manual cleanup is still needed.

See how it works →

Canva's new image model thinks about layout first

Canva has introduced its first image generation model developed entirely in-house.

It's called Craft, and what's interesting is how it approaches the job.

Instead of going directly from a prompt to an image, Craft first creates a plan for the composition.

It considers where elements should go, their relative sizes, the visual hierarchy, and any text that needs to appear.

Then it generates the image.

Canva says this approach also makes the results easier to separate into editable elements using Magic Layers.

If you've ever generated an image with beautiful typography that you couldn't actually edit, you know the frustration.

I haven't seen enough independent testing to know how reliably this works, but the direction makes sense.

Read the brief →

Spline makes it easier to build 3D scenes with AI

Here's another example of AI getting closer to the tools we already use.

Spline moved its MCP server to the cloud, making it easier to connect AI assistants without running the desktop app.

You can ask a connected assistant to build or edit a 3D scene, work with materials and animations, and return a link to the actual Spline file.

I like that the result isn't just a rendered image.

You get a project you can open, inspect, and continue working on.

For designers experimenting with interactive 3D on websites, this could make it easier to get started.

I'd still check the geometry, lighting, and interactions before using anything in production.

Read the full brief →

Worth trying: use AI to draft alt text, then make it useful

This week's workflow is a good reminder that AI can save time without necessarily understanding the job.

Webflow can generate alt text across a website, and Framer has plugins that offer similar functionality.

That's useful, especially when you're working with a site that has hundreds of images.

But there's a catch. AI is pretty good at describing what an image looks like. It isn't necessarily good at understanding why that image is there.

A decorative image might not need a description at all. A linked icon needs to communicate an action. A chart needs to convey information.

I'd try this on a small website and review the results in context.

Start by identifying which images actually need descriptions. Let AI draft the straightforward ones, then go through each result and make sure it communicates what someone needs to know.

It's a simple workflow, but one where human judgment makes a real difference.

See the workflow →

The Pulse: when AI recreates someone's face, who controls the result?

Here's something worth thinking about as AI-generated 3D becomes more accessible.

KeenFace, a new service from KeenTools, can turn a handful of photographs into a textured 3D head with consistent topology and facial blendshapes.

The output is designed to work with existing character animation pipelines, including MetaHuman workflows.

For 3D artists, that could eliminate a lot of repetitive modeling work.

But there's another side to this. The service processes photographs in the cloud, and the available documentation doesn't clearly establish how uploaded images are retained or what commercial rights apply to the generated results.

When the asset is based on a real person's face, this matters.

I think this is something we'll need to pay more attention to as AI tools become capable of producing increasingly realistic digital doubles.

The technology is interesting. Making sure we have permission to use someone's likeness is just as important.

Read the full brief →

One more thing: After Effects makes rotoscoping a lot easier

Adobe's Object Matte tool lets you select a subject in a video with a click, then uses AI to carry that selection through the shot.

It replaced Roto Brush in After Effects 26.2 earlier this year, so this isn't a brand-new feature.

What makes the workflow useful is understanding what happens after that first click. You still need to check the selection, correct frames where it loses the subject, refine difficult edges, and freeze the result before moving on.

The article walks through that process, including a few common problems that can make an otherwise good matte fall apart.

I can see this being particularly useful for putting typography behind moving subjects or isolating elements for compositing.

See the workflow →

That's it for this week.

I'll keep watching what's changing, testing what looks useful, and filtering out the stuff that probably isn't worth your time.

The Only