Most people are visual learners (I definitely am), but coding agents like Claude Code aren’t naturally visual. If you’ve ever asked an AI coding tool to make a diagram, you’ve probably gotten the same messy output: bad spacing, overlapping elements, zero hierarchy. In this video, I walk through the Excalidraw Diagram Skill I built for Claude Code so it can “argue visually” and produce diagrams that actually teach something—clean layout, intentional structure, and a consistent style. The core of the workflow is simple: you give the agent an idea (or a file path—YouTube script, PDF, notes, even a codebase), it generates Excalidraw JSON, and then it validates the diagram visually. That self-validation loop is the secret sauce: it renders the diagram, screenshots it, inspects the PNG, and iterates on imperfections by directly editing the JSON. It’s not perfect on the first pass (there are too many micro-decisions in layout, color, arrows, labels), but it consistently gets you an insane starting point that you can refine in a couple of directed iterations. I also show how I structure patterns, avoid “boxes and boxes,” handle deep diagrams in chunks to stay under token limits, and customize palettes so the output stays on-brand.