Build Beautiful Diagrams with Claude Code (Full Workflow)

Source: YouTube — Cole Medin, published 2026-03-02 Link: https://www.youtube.com/watch?v=m3fqyXZ4k4I

Summary

Cole Medin shares his Excalidraw diagram skill for Claude Code — a packaged workflow for generating high-quality visual diagrams from natural language. The key innovation is the visual validation loop: the agent generates a diagram, renders it as PNG, reviews the screenshot, and iterates on imperfections before returning control to the human.

Core Pattern: Argue Visually

The skill’s design philosophy: structure and labels should convey the concept even if all explanatory text is stripped. Two self-check questions: (1) Does the visual structure mirror the concept’s behavior? (2) Could someone learn something concrete from this diagram?

Workflow

  1. Assess depth — simple diagrams built at once; complex ones built section-by-section (32k token output limit)
  2. Map pattern — determine shapes, text, layout
  3. Build JSON — generate Excalidraw-compatible JSON
  4. Visual validation loop — render PNG via Python script → agent reviews image → makes direct edits to JSON → repeats 2-4 times
  5. Human review — typically 1-3 additional directed iterations to reach final quality

Installation

Clone GitHub repo → copy to .claude/skills/excal-diagram/ → skill.md provides the instruction set. Works with any coding agent. Renders via excalidraw.com (free) or Obsidian Excalidraw plugin.

Connection to AutoResearch

Nick Saraev used a diagram generator skill as his AutoResearch example (improving from 32/40 to 39/40 eval score). Cole’s skill represents the manual design approach; AutoResearch represents the automated improvement approach. Complementary.

Pages Created or Updated

See Also