Something most marketers have done a thousand times: right-click on a website, hit “Inspect,” poke around the CSS to understand how a component was built. Colors, fonts, spacing, shadows. All code. What you see on the front end of any website or app is a visual representation of code underneath. We just don’t design that way because we’re not computers. We think in visuals.

I started thinking about this a lot when I realized that the hardest thing to do well with LLMs wasn’t writing or strategy or analysis. It was making things look good. Making ads. Making landing pages that didn’t look like a first draft from a template library. And I was deeply motivated to figure it out, because as a solo Growth operator, the place where I most often had to hire people and had the hardest time doing it alone was in design.

Most designers I know work full time. They don’t take calls that sound like, “I need four hours this week, then nothing for three weeks.” So you end up on Fiverr with mixed results, or you’re hoofing it in Canva for four days straight trying to get campaign assets done on time. Three months ago I decided I was going to figure out how to actually make production-quality marketing graphics with AI. What I found was more interesting, more frustrating, and more rewarding than I expected.Why most AI design content is a dead end

The first thing I noticed is that there is a lot of content about AI and design, but much less help with the production problems working marketers face. A demo has to hold your attention; a production workflow has to survive revisions, approvals, and the next campaign. I want to know what happens after the impressive first output.

The second reason is a quality bar mismatch. Some of these tools might work fine for someone making lead-gen pages for their own AI automation business. But I’m making assets for high-visibility clients spending real money. I can’t present the quality level that most of these tools produce out of the box. The output isn’t bad. It’s just not premium or differentiated in-market. And turning a 70% output into a 95% output often takes longer than doing it a different way from the start.

The third reason is that everything changes constantly. The tools change every day. The LLMs themselves change almost every week. Content that was accurate yesterday is a dead end tomorrow. A workflow that worked great with one model might not work at all with the next. You can’t rely on any specific tool recommendation from three months ago, let alone three weeks.

For all three of these reasons, the only thing worth learning right now is a process that’s tool-agnostic and built on how the models actually think.

Taking a first principles approach: why is it so hard to get an LLM to produce creative assets? Here’s the analogy that changed how I think about this. Imagine you ask a chef to cook a four-course turkey dinner. They walk into the kitchen and all they have is potatoes. They’re going to be incredibly creative with those potatoes. They’ll make hash browns, mashed potatoes, an impressive potato soup. But it’s not going to be a four-course turkey dinner, because you can’t make cranberry sauce without cranberries. You can’t make gravy without turkey drippings. You can’t make a turkey if you don’t have a turkey.

This is exactly what happens when most people try to make marketing graphics with AI. They go straight into the tool and say, “Make me Instagram ads for next month’s campaign.” The tool does its best with what it has, which is essentially nothing, and produces something generic that could have been made for anyone.Using AI and producing useful creative are different things. In my own work, the gap often came down to what I had prepared before opening the generation tool.

Manuel Bergin, AI Creative Lead at Superside, said: “AI can generate assets. Someone has to assemble those pieces into a narrative.” Cristian Ginori from the same team described it even more directly: “I’m the brain and AI is the hands. I focus on being very precise about what I need.”

The precise “right ingredients” part is what most people skip. And it’s everything.

When I think about what a (great) designer has when they make great ads for a client, the list is longer than most people realize. They have brand positioning and competitive landscape analysis. They have ICP profiles and user journey maps. They have key messages, value propositions, proof points, and statistical evidence to support every claim. They have voice exemplars that define how the brand sounds. They have visual brand guidelines with color palettes, typography, textures, and iconography. They have product screenshots, photography, B-roll video, and examples of best-in-class ads from competitors. None of this shows up by magic. A designer at an agency gets all of this from an intake process, from the strategy team, from the creative director. It often takes weeks to assemble.

When you’re working with an LLM, you have to create or collect every single one of these pieces yourself before you ask it to produce anything visual. If you skip this, you don’t know if you’ll get a turkey dinner or hibachi chicken.

The process I’ve landed on has five stages.

One thing this process taught me is a deeper respect for designers. I’ve worked alongside brilliant designers for years, but I never fully appreciated how many steps are involved, how much expertise each step requires, and how much time even simple things like masking backgrounds or building icon sets actually take. If you’re a designer reading this: I see you. There is so much more to this than I thought.

In the workflow I was building when I wrote this series, I worked with HTML, CSS, JSON, and SVG before reviewing the result visually. The useful loop was to build in code, inspect the design, make changes, and bring those changes back into the working files.

That isn’t the only way to make creative with AI. My original criticism of Figma was too broad and is now outdated: Figma supports agent-driven canvas editing. The updated Paper review explains that correction and what I would compare today.

If this resonates, here’s where to start. Not with tools, not with production. With documentation.

The biggest barrier to good AI-produced creative isn’t the technology. It’s that marketing teams don’t have a documentation culture the way engineering and product teams do. The strategic context that makes great ads possible, the positioning, the brand personality, the key messages, the visual rules, all of it lives in people’s heads. Not in files. If it’s in your head, the LLM doesn’t know it. And if the LLM doesn’t know it, it’s going to pull from generic information on the open internet to make your graphics. Which is why they look generic.

Superside’s research confirmed this: custom AI systems trained on documented brand context dramatically outperform generic prompts. Daniel Bell, Design Manager at Booking.com, framed it as: “The brand is our north star. If we’re telling the story right, we’re hitting that consistency.” The north star has to be written down for AI to follow it.Start by taking stock of what you already have versus what needs to be created versus what just needs to be written down. Most teams have made positioning decisions, brand aesthetic decisions, messaging decisions. They just haven’t documented them anywhere an LLM can access. Next, translate the design decisions into precise references: colors, typefaces, spacing, and reusable components. Screenshots are useful for showing the intended result; structured values help the model reproduce details. Use both where they help, then inspect what the model makes.

Don’t try to build all of this in a day. Just start with the documentation. Get the strategic underpinnings out of your head and into files. The next time you sit down to make creative with AI, give the LLM all those context files and watch how much better the output gets. Take it from there.

The five-stage overview expands into six practical articles:

I’m Hanna Huffman. Marketer in the Loop is my weekly letter about doing good marketing work as AI changes how we do it. Expect useful ideas, people worth knowing, and things I’m figuring out in my own work.

Published by MultiplAI Growth Partners.