Virtual whiteboards became more useful than physical ones by adding features such as infinite space, keyboard input, selection, deletion, and copy-paste.
2
Excalidraw's most successful AI integration converts prompts into structured diagrams that become editable Excalidraw components.
3
AI features should fit the product's actual use, and domain-specific languages give language models a structured target that can be converted into application data.
Summary
Christopher Chedeau explains how Excalidraw grew from a tool for making handwritten illustrations into a virtual whiteboard. During the move from physical to virtual whiteboards, Excalidraw improved the experience by removing physical limits and adding features such as infinite space, keyboard input, selection, deletion, and copy-paste. He argues that AI products are at a similar early stage, so teams should avoid adding models simply because they exist. Excalidraw tried image generation and direct diagram generation, but the useful path was to have a language model produce Mermaid syntax, then convert that structured output into editable Excalidraw components. This makes AI a starting point that people can rearrange, recolor, and finish themselves. Chedeau also describes document name generation, planned illustration generation, and a request for better browser-based logo background removal. His final advice is to find a domain-specific language for the problem and let the model target that structure.
Excalidraw began as a way to make handwritten illustrations
While procrastinating on performance reviews at Facebook in January 2020, Chedeau decided to write a blog post. The tool he wanted for handwritten illustrations was unavailable, so he opened CodeSandbox and began reimplementing it with rough.js. He posted the result on Twitter, where the first tweet received about 200 likes and many replies. That experiment became Excalidraw. Chedeau says that if he started the project today, he would probably use a coding agent instead of writing the first version himself.
Virtual whiteboards improved by dropping physical constraints
When COVID sent people home, users began treating Excalidraw as a whiteboard even though Chedeau had built it for blog illustrations. He noticed that many online whiteboards copied physical limitations, such as a fixed board size and the need to draw every letter with a mouse. Excalidraw allowed people to scroll to new space, type into text boxes, select and delete objects, draw circles as shapes, and copy and paste. These virtual-native features made the online version more useful than a literal physical-board recreation.
Excalidraw experimented with a model from Gorak at fal.ai that turned a sketch, such as a palm tree or sun, into a live realistic beach image. The model worked well, but the integration did not fit how people use Excalidraw. Users generally draw diagrams rather than realistic pictures. Chedeau's lesson is that an available model is not automatically a good product feature. Adding AI without a clear fit can make the overall experience worse than leaving AI out.
Structured diagram output works better than direct file generation
The team first tried asking language models to generate diagrams directly, with poor results. They then found that models could produce Mermaid syntax, such as a flowchart describing an AI engineer. Mermaid provides a way to describe many diagram types in text, including flowcharts, sequence diagrams, network diagrams, architecture diagrams, and class diagrams. The remaining problem was that Mermaid produced SVG or canvas output rather than Excalidraw files.
Editable components make AI a starting point for human work
Kato implemented an Excalidraw backend for Mermaid, creating an integration that turns a prompt into an actual Excalidraw component. A user can request a flowchart, an architecture diagram, or another supported diagram and then edit the result. Chedeau shows the human role as changing the layout, moving objects, adding colors, and deciding what the finished diagram should be. He expects future versions to support follow-up instructions such as asking the system to modify the existing result, although that interaction was not available yet.
Language models can remove small document-management tasks
Excalidraw's cloud product can contain many files named Untitled 1, Untitled 2, and so on. The team waits until the user stops working, then asks a language model to suggest names based on the document content. The user chooses a suggestion with a click. Chedeau presents this as a good use of AI because it removes a repetitive task and does not ask the model to control the whole editing workflow.
Chedeau says people use Excalidraw for infographics and illustrations in slides and diagrams, even though they do not generally use it to make realistic paintings. The team was testing an illustration-generation feature in a closed alpha. He said the plan was to open source it and ship it within a few weeks. This feature targets a type of content that users already want to place inside their diagrams.
Domain-specific structure gives AI a reliable handoff
Chedeau compares the current moment in AI product design with the earlier move from physical to virtual whiteboards. He asks teams to resist shipping random models and to choose integrations that help users do what they came to do. His final technical advice is that language models work better when a problem can be represented in a domain-specific language. The model generates that language, and the non-AI part of the application converts it into the system's own output.
"Please please please resist the urge to actually ship all of the random models that exist within your app."15:48
Who should watch
You are adding AI to a design or collaboration product and need a way to decide which model capabilities belong in the workflow.
Your application produces diagrams, documents, or other structured artifacts that users need to edit after generation.
You are considering direct model output and want an example of routing structured language-model output through an existing application backend.
