Miro – Board generation

Early on, I built This or That, a small voting app that shows two frames and asks which one is better, changing one thing at a time. I shared it with the product and design orgs to get a quick read beyond my own taste. The clearest results were about basic craft: generous margins, a big left-aligned title, clear labels and soft colour, and a spacious frame beat a packed one.

From there I wrote a short set of design rules for boards, building on how Miro's own templates are designed. The board title sits alone at the top, sections read left to right, and headings go inside their frames. Each frame uses two or three colours, and colour shows meaning or grouping, not decoration. Dot voting and timers sit inside the frame they belong to. And when Miro has a real widget for something, like cards for a kanban, the AI should use it instead of drawing one out of shapes.

In the first version of board generation, the prompts mixed design guidance with instructions for how to build a board, and the style was described as "Refined Neobrutalism", with bold black borders. I moved all the design guidance into one file, built on the new rules, and rewrote it with Claude. With AI, it's hard to tell whether a change made things better, because every run comes out different. So I ran every change against the same test prompt, "a retro board for my team", and compared screenshots round after round. The style moved to bright, readable colour on calm white surfaces with plenty of space.

For the second version, I went back to the style. When the model was given exact colour values, it used those colours everywhere and rarely anything else. Describing a style instead gave it room to be more creative, while the boards still felt like they belonged together. The style has six colours in four shades each, flat surfaces, titles set as text rather than in coloured blocks, and colour only where it means something. We tested it over dozens of runs, and it went into the prompt that runs in Miro. It's now the default when an organisation hasn't set up a brand theme. When it has, the brand theme from Miro's Brand center is used instead.

The same idea of quality had to be written twice. Once so people could follow it, and once so the AI could.

For people, I wrote a rulebook organised by content, composition and visual, with reasons and examples. One idea kept coming up. Every area of a board has a job: bringing in context, doing the work, or showing the outcome.

I built a journey map to put the rulebook into practice. Each stage has evidence, signals, an emotion curve, insights and opportunities. It became the reference for the second version of board generation.

Writing for an AI turned out to be its own craft. Rules worked better as defaults than as bans. If you tell the model something is never allowed, it refuses even when the user asks for it, and a warning about one thing can make it drop a whole kind of element. When the prompt said some lines could break a board, the model stopped drawing lines altogether. Examples worked better than rules, but the model copies everything in them, mistakes included. A few black labels in one example were enough to fill every board with black shapes, so in the end the examples only showed where things go, and colour was left to the style.

Models also have their own taste. Some reached for black sticky notes and dark colours whenever they could, and the prompt had to work against that. Some problems no prompt could fix, like where things go on the board and how big they are. The engineers solved those in how the AI builds boards, so the prompt could focus on style.

I also built demos where the AI works on the board step by step with the user.

Board generation launched as part of Miro's Intelligent Canvas in October 2026. It was a team effort across product, design and engineering, and my part was how the boards look and feel. Joe McLean wrote about the wider effort, including the engineering work behind it, in How Miro fixed AI canvas editing.

The style was never meant to be the finished look. It's a good starting point for every board, and people make it their own with their brand, their templates and their own edits.

I worked with Joe McLean, Arturo Arranz Mateo, Ivan Achille, Mark Boyes-Smith, Monika Pawlak, Nina Simonoska and many others.