Architecture Diagram

Turn a stabilized design — a list of components and the connections between them — into a polished, self-contained HTML + SVG diagram. Components are boxes, arrows are communication/data flow, dashed boundaries are regions or trust zones. This is the diagramming engine for the whole plugin; the orchestrator and building blocks hand off here when it's time to draw.

Self-contained by design: one HTML file with inline SVG and CSS, a system
monospace font (no web-font fetch), and no dependency to render — it draws
fully offline. The only external calls are two pinned, SRI-protected CDN
scripts that power the optional PNG/PDF export; remove them and the diagram
still renders — only the export buttons are lost.

When to reach for this

At step 3 (high-level design) and step 6 (deep-dive) of the reasoning loop, and any time a design is clear enough to draw: the component list and their connections are known. Also for failure-path diagrams (step 5) — showing a degradation flow is part of the design.

When NOT to

Before the design has stabilized — don't draw boxes that aren't yet justified (that's failure mode #7). A quick ASCII sketch in the conversation is fine for early thinking; render with this skill once the components have earned their place.

How to build a diagram

1. Gather the spec from the design: the list of components (each with a type),

the directed connections (with protocol/label), and any region/boundary groupings. The system-design design-doc already produces this.

1. Copy the template assets/template.html to a working file. It carries the

dark theme, grid background, fonts, the export toolbar, and the summary-card layout — keep those intact.

1. Map each component to a semantic type (color) using the table below.
2. Place boxes and draw arrows following the layout/spacing rules in

references/design-system.md (draw arrows before boxes so they sit behind; mask arrows behind semi-transparent fills; keep ≥40px vertical gaps; legends go outside boundary boxes).

1. Label the flow — number arrows to show request order; use dashed arrows for

async/replication/fallback; dashed boundaries for regions and security groups.

1. Fill the summary cards with the key decisions/trade-offs (ties to the

design doc), update the title/footer.

1. Open it: open <file>.html to view; the toolbar exports PNG/PDF.

Semantic colors (map components to types)

These align with how the building blocks think about a system.

Full styling, arrow markers, masking, spacing, and legend rules are in references/design-system.md.

Output rules

• One self-contained .html file: inline CSS, inline SVG, no external images.
• Keep the two pinned CDN scripts (html2canvas + jsPDF) with their SRI hashes —

they power Copy/PNG/PDF export and nothing else. The diagram renders fully without them; they're only needed for export.

• Avoid SVG <foreignObject> (renders inconsistently in html2canvas) — use plain

<rect>/<text>.

• Keep it to ~8–15 components for a high-level view; zoom into one subsystem for a

deep-dive rather than cramming everything into one picture.

Dos and don'ts

Do:

• Draw only after the design stabilizes; render the component list and connections the design doc already produced.
• Map every box to a semantic type from the color table so the diagram reads at a glance.
• Draw arrows before boxes (so they sit behind), number them for request order, and reserve dashed arrows for async/replication/fallback.
• Keep it to ~8–15 components; zoom into one subsystem for a deep-dive instead of one crowded picture.
• Keep the file self-contained — inline CSS/SVG, system monospace font, no external images — so it renders offline.

Don't:

• Don't embed Mermaid or invent a second diagram convention; this skill is the one engine.
• Don't use SVG <foreignObject> (html2canvas renders it inconsistently) — use plain <rect>/<text>.
• Don't strip or rewrite the pinned CDN scripts' SRI hashes; recompute them only when bumping versions.
• Don't place legends inside boundary boxes or stack components with under-40px gaps — both cause overlaps.
• Don't add a hard dependency on any send-to-Claude transport; the Copy-Prompt button is the self-contained default.

Interactive / collaborative mode (optional)

For iterating with the user, an interactive variant adds layer toggles, connection-type filters, zoom, and click-to-comment that builds a feedback prompt with a Copy button. See references/interactive.md. It degrades gracefully: the Copy-Prompt button always works. A live "Send to Claude" button is possible only if a separate prompt-transport plugin (e.g. playground-sync) is installed — this plugin does not require or bundle one.

Related building blocks

• system-design — invoked by the orchestrator at the "draw it" steps; its diagramming notes feed what to draw at each step.
• Any building block — invoked by a block whose "Diagram" section says "render with architecture-diagram". This skill renders for all of them; it owns diagram generation and they point here rather than ship their own.

References

• references/design-system.md — colors, arrow markers, z-order/masking, spacing, legend placement, the export toolbar internals. Read before drawing.
• references/interactive.md — the optional interactive/commentable diagram pattern and the (optional) send-to-Claude transport. Read only if building an interactive map.

Assets

• assets/template.html — the self-contained dark-theme HTML+SVG starting point with the export toolbar. Copy and customize it.