Steve Ruiz

Steve Ruiz is the founder of tldraw and creator of Perfect Freehand, a library for drawing pressure-sensitive lines. His AI work centers on the canvas as a place to communicate requirements, inspect results, and revise them visually. In 2026, he introduced shared canvases inside agent conversations and described bringing local coding agents into a file-based desktop application.[1][2][4][5][6]

In “Text is not enough”, published March 30, 2026, Ruiz argued that visual adjustments and complex workflows can be difficult to explain through chat alone: “Some ideas just don’t fit in a chat box.” His team’s MCP App placed a tldraw canvas inside conversations, allowing agents to draw, diagram, and wireframe while users edited the same canvas and agents saw those changes. He announced availability in Claude and Cursor at the time.[5]

Ruiz described asking a Cursor agent to explain a proposed change to tldraw’s curve calculations visually. It placed the old approach on the left, the new approach on the right, and annotated the optimization. He said this helped him understand the change and steer the agent. In another example, an agent diagrammed the MCP App’s clients, transport layer, tool dispatch system, and editor instance; after visual adjustments, the diagram went into the documentation.[5]

In his July 16, 2026 tldraw offline launch account, Ruiz recalled starting a desktop prototype during winter break 2024, then shelving it after technical barriers. One mistake was saving the entire document after every change. Returning during winter break 2025, he used newer coding agents to replace that approach with sending and merging only the changed portions, analogous to tldraw’s multiplayer synchronization. He reported that this made the application dramatically faster.[6]

Ruiz wanted agents to interact with the canvas while keeping the application local. He described a consequential execution choice: online, the team constrained AI output to a simplified schema that it translated into code; offline, it opened a port accepting raw JavaScript with access to the editor and window. His rationale was that the absence of a server greatly reduced the risk—a judgment he expressed about the design, rather than a general security guarantee.[6]

The team distinguished one-off execs from persistent scripts. Generating 100 shapes could be a single execution, but a drawn bin that deleted shapes dragged into it needed its behavior restored whenever the file reopened. The new .tldraw format combined canvas data, assets, and scripts in one file, allowing a drawing to retain programmed behavior. The desktop application launched without accounts or a server.[6]

These agent interfaces followed Ruiz’s earlier work on visual prompting. In November 2023, Figma engineer Sawyer Hood demonstrated drawing an interface in tldraw, exporting it as an image, and sending it to GPT-4V to obtain a working website. In his November 18 account, Ruiz credited Hood’s initial demo and described his team forking the repository the next morning.[3]

The team’s first change was to put the generated website back on the canvas. Because tldraw used regular HTML and React components rather than an HTML canvas element, the website could appear as an interactive iframe alongside drawings. Users could resize it to inspect responsive layouts, compare iterations spatially, and draw revision notes over the result.[3][1]

That created an iterative loop: generate a prototype, annotate it, and submit the result with its annotations as a new prompt. Their image exports represented embedded iframe content as empty rectangles, so the team supplied the previous result’s HTML and instructed GPT-4V to fill in the blank box. Ruiz also described sending canvas text strings separately to reduce reliance on the model reading small text from images.[3][1]

In the interview published January 5, 2024, Ruiz compared four kitchen-timer prompts: a labeled box, a UI drawing, a state chart, and the UI plus state chart. The UI-only result placed buttons correctly but did not work; the combined input produced a working timer in the demonstration. Another result counted upward until he annotated it to count downward. For Ruiz, the important capability was revising requirements and outputs rather than finding a perfect initial prompt.[1]

The public Make Real release immediately hit service rate limits. The team added an input for users’ own OpenAI developer API keys, which Ruiz explicitly called a normally terrible idea and security risk. It subsequently released a starter template so other developers could build their own tldraw and GPT-4V experiments.[3]

With Fal.ai’s help, the team also explored rapidly updating generated imagery in drawfast and lens. Lens used a collaborative, vertically scrolling drawing as input. Ruiz recalled that, alongside Cloudflare Workers and CRDT-based collaboration, a laptop in his living room captured and uploaded screenshots; he left his family a note asking them not to close it. He valued the shared experience more than any individual generated image.[1]

Ruiz favored continuing to develop the reusable canvas over committing exclusively to a single AI subscription application. In the 2024 interview, he argued that the refined canvas interactions were the difficult-to-recreate contribution, while models and particular AI features would keep changing. A single application would explore only one direction; an extensible canvas could support many. He did not rule out a future paid application.[1]

The canvas underlying these experiments grew out of Ruiz’s earlier art and design-tool work. He earned a master’s in visual art at the University of Chicago, wrote about contemporary art, organized exhibitions, and made paintings and drawings. After moving to the UK, he continued studio work. He recalled turning toward design at age 30 because he wanted to earn money and work more closely with other people, drawing on experience building WordPress portfolios for himself and other artists.[1]

Ruiz used Framer for technical prototyping and later spent a year creating its educational content, including tutorials and writing. He subsequently worked with Play on a mobile-first iOS design application. His prototyping work addressed interactions whose behavior had to be explored by building them.[1][2]

His open-source projects brought aesthetic judgment into technical tools. Perfect Arrows addressed how to make arrows between arbitrary points and shapes look good, through repeated visual iteration. Perfect Freehand addressed variable-width digital ink using vectors: Ruiz wanted input points converted into a polygon outlining the stroke, suitable for SVG, rather than a sequence of raster stamps. He developed it publicly and recalled receiving help from people with related expertise.[1][2]

Perfect Freehand exposes getStroke, which constructs spline points from input points and then produces polygon outline points. Callers choose how to render the result. Its documentation describes default pressure simulation based on distances between input points; actual stylus pressure can instead be supplied as a third coordinate with simulatePressure disabled. The library is MIT licensed.[4]

Integrating his ink into Excalidraw and building canvas tools to develop and inspect it preceded tldraw. Ruiz recalled arranging to join Adobe but delaying his start for six months to work on his own whiteboard project. He launched tldraw in November 2021, then declined the planned Adobe role after teams wanted to build applications on its canvas. He reoriented the project toward a developer infrastructure SDK, comparing its role to Mapbox providing maps for applications that would not build them themselves. His personal profile dates the startup’s formation to 2022.[1][2]

Ruiz’s founder advice retained a lesson from studio practice: use a sense of what is interesting as a filter for code, features, partners, and problems. In the 2024 interview, he described preserving the ability to prioritize “what is most interesting right now,” alongside working backward from desired outcomes.[1]