learn¶

Reverse engineer D2 code from existing SVG diagrams.

Usage¶

d2vision learn <file.svg> [flags]

Flags¶

Flag	Default	Description
`-f, --format`	`toon`	Output format: toon, json, yaml
`--d2`	`false`	Output D2 code instead of spec

What It Does¶

The learn command analyzes a D2-generated SVG and produces either:

A DiagramSpec (TOON/JSON/YAML) that can be modified and regenerated
D2 code that recreates the diagram

This helps AI assistants:

Understand existing diagram patterns
Learn from examples
Modify existing diagrams

Examples¶

Get D2 Code¶

d2vision learn diagram.svg --d2

Output:

grid-columns: 2

cluster1: Cluster 1 {
  direction: down
  service1: Service 1
  db1: Database { shape: cylinder }
  service1 -> db1
}

Get TOON Spec¶

d2vision learn diagram.svg

Round-Trip Editing¶

# Learn from existing diagram
d2vision learn original.svg > spec.toon

# Modify the spec (add nodes, change labels, etc.)
# ... edit spec.toon ...

# Regenerate
d2vision generate spec.toon | d2 - modified.svg

Compare Original vs Learned¶

# Generate D2 from learned spec
d2vision learn original.svg --d2 > learned.d2

# Render the learned version
d2 learned.d2 learned.svg

# Compare
d2vision diff original.svg learned.svg

What Gets Detected¶

Layout¶

grid-columns from horizontal node alignment
direction from child node positions (variance analysis)

Structure¶

Container hierarchy
Nested containers
Node shapes and labels
Edges (internal and cross-container)

Shapes¶

Detected shapes:

rectangle
circle
oval
cylinder
diamond
hexagon

Limitations¶

Styling (colors, fonts) may not be fully preserved
Complex path-based shapes may default to rectangle
Layout engine hints (like near) are not detected
Some edge routing details may be lost