My AI workflow

Over the last year or so AI has become an essential part of my daily work. For quite some time I have seen it as a kind of short hype that doesn’t really produce anything significant and would eventually go away. But the unprecedented pace of progress in this area can’t be neglected anymore.

Designers have become builders, enhancing their toolset with personal or brand-specific software. This thread lists some nice examples. Dan Hollick gives a detailed insight into his custom Figma plugins. Wild’s lovely microsite also was put together with helper tools made with AI. The designer-developer handoff process has also entered a new chapter. In principle, you could now skip traditional design tools altogether and design directly in code. Alex Kehr calls it direct design.

Paul Goldsmith-Pinkham wrote an excellent overview on the different levels of AI coding. For me levels 1-4 are the most relevant to my work. As the sole product designer in a media company this enables me to achieve things much faster than before. Here’s a little breakdown of how I use individual tools to help me get things done.

Chatbots

I use ChatGPT specifically for these three areas:

UX and strategy feedback

Challenging UX/UI ideas either by describing them and uploading screenshots out of Figma (ok, so this is possible directly in Figma now - but I like having everything in one place). It’s like having a partner helping you iterate on things and pointing out weak points you might not have seen.

Web development advice

Generating or analysing short snippets for existing stuff where bringing in agents directly would be overkill or simply not possible. A good example would be tedious HTML email markup generation for SaaS tools like HubSpot or Piano, where template code is fragmented and has to be copy-pasted into a web UI. This works extremely well and is a big time-saver.

Image generation

GPT-Image works really well when used inside a custom GPT with an exact, detailed description of the desired outcome (which I again developed with the help of GPT). I currently use it for generating variations of our brand mascot at Kleine Zeitung. The big challenge here is achieving style consistency also over longer conversations. Even with a carefully defined prompt, some degree of style drift is almost inevitable. For brand work, where consistency matters more than simply producing a good-looking image, this is still a significant limitation. I often have to start a new conversation to get back to the desired visual language.

Agentic coding

Codex or Claude Code is coming into play for larger tasks. It lets me try out ideas that would have been too complicated or time-consuming before. A custom web app for mascot variant generation is such an example. Instead of spending hours working through an API’s documentation before writing the first line of code, I can get a basic version working surprisingly quickly and then iterate from there. Figma MCP is great for directly manipulating or generating layouts on the canvas, effectively allowing an AI agent to work with Figma as part of the design process.

Figma AI

I know Figma is using existing models under the hood but lets summarize it as Figma AI for the sake of convenience. For one, little things like being able to quickly remove an image background directly inside your design tool is really handy. Who doesn’t remember the tedious process of doing this manually? But also Figma Make or now Weave enable new workflows that were not possible before. With Make I can create enhanced prototypes out of Figma layouts that would be too complicated to set up with the traditional prototyping functionality.

Node-based AI tools like Weave allow me to use various models inside a single workflow. For example, I can connect a video model to animate a static image of our mascot. Figma’s built-in agent also let’s you create plugins easily. My current project is a generator that lets me quickly remix our podcast cover layout pattern based on title text, colors and the host’s image.

Conclusion

I’m still far from replacing traditional design tools or development workflows. But the boundaries between designing, prototyping and building are becoming increasingly blurry and that’s probably the part of AI that interests me most.