Diagrams your agent
must prove rendered.
DesignCore turns a coordinate-free graph spec into a verified draw.io, Mermaid, or Excalidraw diagram. The model never places a shape. Nothing counts as finished until it has actually rendered.
Judgment for the model, geometry for the machine
LLMs are good at structure and bad at pixels. DesignCore splits the job along exactly that line.
1 · The model never writes coordinates
x, y, width, height, position are rejected at parse time. All geometry comes from Graphviz layout.
2 · No render, no diagram
A missing backend raises an error naming its install command. A spec never quietly degrades into an unverified picture.
Verified, honestly defined
Verified means: compiled cleanly, every backend rendered, deterministic structural/density/geometry checks passed. Not visual taste — that stays human.
One spec, three destinations
Write the graph once. Render it wherever the document needs it.
| Format | Best for |
|---|---|
| Excalidraw (default) | Editable sketches and architecture drawings your team keeps iterating on |
| Mermaid | Diagrams inside markdown — renders natively on GitHub, diffs as text |
| draw.io | Conventional architecture artifacts for orgs already living in drawio |
A real output, not a mockup
This diagram of DesignCore's own pipeline was produced by DesignCore: a coordinate-free spec, Graphviz geometry, Excalidraw render, lint-clean.
Spec in, verified diagram out
Requires Python 3.11+, uv,
and the render backends you use (designcore doctor names each install command).
# Install the CLI uv tool install designcore # Check which backends you have designcore doctor # Describe the graph — no coordinates, ever designcore new system-context --kind context # Compile, lay out, render, lint — or fail loudly designcore render system-context # Prove the manifest matches disk designcore check
Judgment layer ships as skills
Three ready-made agent skills carry the conventions so any AI CLI
can drive the pipeline well: architecture-diagram,
flow-diagram, and concept-sketch.