AI Design Studio

Make anything you can describe

You don't need to draw or own Photoshop. If you can describe it, you can make it. This is a fast 90 minutes, I demo, and the hands-on is your homework (it's all here and in the workbook). We go module by module: what it is, how the tools differ, how to do it.

The five modules

The one rule for a good image prompt

Stay in touch

What's an image model?

An AI you hand words to, and it paints a picture. You're not picking from stock or clip art, you describe what you want and it generates something brand new. Think of it as a wildly fast illustrator who's seen millions of images and will draw anything you describe, but only as well as you describe it. That last part is the whole class.

How the models differ, pick by the job

Also strong for high-volume marketing visuals: Leonardo (now part of Canva), see the More Tools page.

How to prompt

Then turn up the detail with six dials: who, where, what, light, mood, style. The more specific, the closer it lands.

Prompt modifier cheat sheet

When the picture needs words

Posters, logos, menus. The models render text fine now, they just won't guess which words. Write the text first, quote it, use Ideogram or Nano Banana, and proofread. Three real generations made live for this class:

Vague prompt. The model wrote its own copy, dropped the headline, stamped "50% OFF."

Exact words, quoted. "GO BANANAS" + "NOW 50% OFF" + mascot, exactly as asked.

It scales to real copy. A full chalkboard menu, but it quietly changed a price. Always proofread.

Then clean it up

Guardrails & Trust, making images responsibly

Two things every maker needs: how to handle the model's limits, and how to publish so people trust what you make.

When the model says no

An education piece: "adversarial poetry"

How blunt filters are, worth understanding, not copying. In late 2025, researchers showed that rewriting a refused request as a poem made models far more likely to comply: jailbreak success rose from ~8% to ~43% across 25 models, past 90% on some (arXiv 2511.15304). Why: safety training concentrates on plain phrasing, so metaphor no longer "looks like" what the filter catches. The takeaway: knowing the gap exists makes you literate, not clever, it breaks every tool's terms and unlocks the output that gets you sued. Don't depend on it.

When you hit a FALSE no

The line you don't cross: if it's blocking a real person, a celebrity, a brand you don't own, explicit content, or a deceptive fake, that's the guardrail doing its job. Never route around it.

The five Trust rules, before you publish

Let yourself be creative. Be human.

Google pours design tools out of one place, Google Labs (labs.google), its public playground for experimental AI. Worth knowing as a family, because they hand off to each other.

The art-direction prompt you hand NotebookLM

Notice how it specifies palette, grid, layout, type, and mood, that's what turns a generic deck into a branded one:

The thread: sources → NotebookLM structures + designs it → Slides to refine → Vids to make it move.

Claude is more than a writer. It has a full set of design abilities built in: it can make a brand, build an animated prototype, design and ship a website, drive your editing tools, and connect to the apps you already use. This is the conversational hub of your design stack, and where your brand foundation lives.

Claude Design: polished visual work by talking

Animations and interactive prototypes

Ask for an animated prototype and Claude writes a self-contained HTML/CSS/JS page and renders it live, something you can click, not just a picture. It can go as far as code-powered prototypes with motion, video, shaders, 3D, and built-in AI.

Websites and apps

Design a site or app in Claude Design, then hand it off to Claude Code to turn the prototype into working, shippable software. (More options in Module 5.)

Drive your real editing tools

Through its connectors, Claude reaches into the tools you already use: Adobe (Photoshop, Illustrator, Lightroom, Premiere, Firefly), Canva, Figma, Blender, and more. So it can edit images, lay out designs, and help cut video inside your real apps, not just talk about them.

MCP connectivity: one assistant, all your tools

All of this runs on the Model Context Protocol (MCP), the open standard that lets Claude plug into 75+ tools and data sources, read your design system, and operate your apps. You describe the goal; Claude works across your whole stack.

Take it to Canva

Export any design to Canva, PDF, PPTX, or a standalone web page. A Canva Brand Kit then locks your logo, colors, and fonts so every future design pulls from them, and Bulk Create turns a spreadsheet into dozens of on-brand posts at once.

Claude Design is a paid research preview (Claude Pro, Max, Team, Enterprise). On free plans, you can still build brand guides, artifacts, and animated prototypes in regular Claude, and a Brand Kit in Canva.

Past single prompts, the next step is a canvas where you wire AI models together into a pipeline you can rerun, generate, edit, branch, remix, instead of one-shot prompting. This is where AI design becomes a system, not a trick.

Why it matters: you stop prompting one-at-a-time and start building a little design machine you can reuse and hand to a teammate. Claude Design is the conversational flavor of this; Flora and Luma are the node-canvas flavor.

You can now describe a website, or a working app, into existence. You don't hand-build pages; you describe the page and the sections, then refine.

Never accept the first draft, refine section by section

AI sites come out generic ("Welcome to our innovative solutions platform"). Highlight one piece and fix it:

More tools to explore

Worth knowing, beyond the five modules. Scan and try the ones that fit your work.

The cheat sheet

Everything in one scrollable card. Screenshot it. Use it this week.

The banana rule

The five modules

Guardrails & Trust

The full toolkit

A studio where art meets artificial intelligence

Talentless AI helps brands and creators turn AI into a working production stack. Austin-based. Built to ship work, not slide decks.

I'm Steve Mudd, founder of Talentless AI, veteran of IBM Watson and Ogilvy. I run classes like this one and build the pipelines I teach. A hands-on Level-2 workshop (agentic design + MCPs) is in the works.

Work together? Email steve.mudd@talentless.ai, text 720-254-7893, or grab the Substack on the Welcome tab.