Mermaid Architecture Diagrams

Mermaid Architecture Diagrams

Produce diagrams that are correct, legible in print, and consistent as a set — not just syntactically valid Mermaid.

The one non-negotiable rule

Never trust unrendered diagram code. Mermaid layout is emergent: label widths, edge routing, and subgraph sizing only exist after rendering. The loop is:

1. Author the .mmd
2. Render it: scripts/render-mmd.sh diagram.mmd
3. Look at the preview image (the script writes a *-preview.png sized for inspection)
4. Fix clipping, overlaps, label collisions, edge crossings — re-render
5. Repeat until clean (minimum one fix round; diagrams are never right first try)

A diagram you haven't looked at is a draft, not a deliverable.

Workflow

1. Purpose first. One diagram, one purpose, one audience. Write the 5-second story the reader must grasp (what talks to what, through what, where the boundaries are) before drawing. If a diagram needs two stories, make two diagrams.
2. Pick the flow direction. Topology/containment reads top-down (TB); pipelines read left-to-right (LR) — but for print/PDF destinations, tall-and-narrow TB beats wide LR: an LR pipeline scaled to text width prints at ~4–5pt. Verified repeatedly.
3. Apply the design system. Use the semantic tokens, edge grammar, and label rules in DESIGN-SYSTEM.md. Color means something or it isn't used. Default to the glass + elevation look (Look 2, judged 8/10); the flat look (5/10) is the exception — reach for it only when colorblind-safety or compliance is explicitly required. "Unknown audience" or "just show me X" is a Look 2 case. See DESIGN-SYSTEM.md's decision rule; don't pick flat because it's listed first.
4. Author within the verified allowlist. TECHNIQUES.md contains only constructs verified through real render pipelines, each tagged with where it works. The headline trap: never set themeVariables.fontFamily for Quarto-destined diagrams — it clips every label.
5. Render with the destination's pipeline (table below).
6. Self-review adversarially against REVIEW-CHECKLIST.md — legibility, design consistency, comprehension/correctness. Fix and re-render.
7. High stakes? Run a variant competition (sketch in REVIEW-CHECKLIST.md): several styled variants, adversarial judging, recombine winners.

Which pipeline renders the truth?

Preview with whatever will render the diagram in production — otherwise you are tuning against the wrong engine:

render-mmd.sh auto-detects (-p auto, the default): prefers Quarto when installed, falls back to mermaid-cli. Both paths produce a full-resolution PNG plus an inspection preview.

Engine reality check

Quarto bundles its own Mermaid (e.g. 11.6.0) with its own screenshot machinery; mmdc ships current Mermaid (e.g. 11.12). Same major version — different behavior: parser strictness, font handling, and theme defaults all differ. That is why TECHNIQUES.md tags every construct [universal], [quarto], [mmdc], or [version-sensitive], and why the portable strategy is to author inside the intersection: full classDef tokens (never rely on theme defaults), no fontFamily override, filters via themeCSS not classDef, subgraph titles ≤24 characters. Diagrams written that way render near-identically in both pipelines.

When you meet a new pipeline (different SSG, different Mermaid version), spike it first: render TECHNIQUES.md's probe snippets through it and note divergences before authoring real diagrams.

Bundled references

• DESIGN-SYSTEM.md — semantic color tokens (colorblind-safe default palette), edge grammar, layout, label and legend rules
• TECHNIQUES.md — verified syntax allowlist with pipeline tags, gotcha table, probe snippets
• REVIEW-CHECKLIST.md — adversarial self-review lenses + the variant-competition pattern
• scripts/render-mmd.sh — adaptive render + preview loop (Quarto or mermaid-cli)