Anthropic Style Diagram

Anthropic Style Diagram

Author a semantic SVG, let the browser do the rest. scripts/render.mjs supplies the design system, resolves the theme, audits the layout, crops the viewBox to the real ink bounds, bakes every computed style into literal attributes, and writes the PNG.

You draw geometry and meaning. The script owns colour, theme, viewBox and export. Never hardcode a hex value, never write a <style> block, never compute the final viewBox height by hand.

Workflow

1. Plan before drawing. Count the nodes and group them by kind — one colour ramp per kind, never one per row. Decide the diagram type (see below). If the user named 6+ components, split into an overview plus one diagram per sub-flow rather than cramming one canvas.
2. Write the SVG to a working path (scratchpad or /tmp), not the delivery path. Follow the authoring contract below; references/example-structural.svg and references/example-cycle.svg are the canonical shapes to imitate.
3. Render: node scripts/render.mjs <working.svg> -o <target.png>.
4. Clear the audit. Layout warnings print on stdout. Fix every one and re-render until the run is clean. A clean run only means nothing collides — that is the floor, not the bar.
5. Review the render yourself against the checklist below. Any failure sends you back to step 2. This step is not optional: the audit measures geometry, and almost everything that makes a diagram wrong lives outside geometry.
6. Report the saved path, and say what the review turned up.

The review checklist

Read the rendered image, not your source. The source is what you meant; the image is what the reader gets, and only one of those can be wrong.

1. Read it cold. Set aside what you set out to draw and say what the picture actually claims. If that differs from the request, the diagram is wrong — not the reader.
2. Trace every arrow. Each must start and end on the shape it means, and its direction must assert something true. An arrow down the page claims the lower node is used by the upper one; say that claim out loud and check it. Placing a component below another is an assertion, not a layout convenience.
3. Find the loop. If the subject has feedback — an agent, a control system, a retry path, a training cycle — locate the return edge in the image. If you cannot see it, it is not there, and a one-way drawing of a loop is simply false.
4. Read every label. Accurate, ≤5 words, sentence case, unclipped, and not claiming more than the system does.
5. Check the colour. Same colour must mean same category. If two nodes share a ramp only because they were next in the list, re-assign.
6. Scan the whitespace. Crowded regions and dead regions are both defects, and a uniform grid of same-sized boxes means the layout is carrying no information.

Rendering

node scripts/render.mjs diagram.svg -o ~/Downloads/architecture.png

The audit catches the three mechanical failure modes: text spilling out of its shape, nodes overlapping, and connectors slicing through unrelated nodes. Round nodes are measured as ellipses rather than as their bounding box, so ring layouts are checked as strictly as grids. A clean run prints only the output paths. It says nothing about whether the diagram is well composed or even correct — see "Composition" below, and never treat a clean audit as approval.

First run needs Playwright. It is picked up from the skill, the project, or the global npm root; if missing, npm install in this directory.

Authoring contract

<svg viewBox="0 0 680 H" role="img">
  <title>Short name</title>
  <desc>One sentence describing what the diagram shows.</desc>
  <defs>
    <marker id="arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="6" markerHeight="6" orient="auto-start-reverse">
      <path d="M2 1L8 5L2 9" fill="none" stroke="context-stroke" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/>
    </marker>
  </defs>
  ...
</svg>

• role="img" with <title> and <desc> as the first two children. Always.
• Canvas width 680 by default — the text-width table below is calibrated to it. Widen to at most ~1000 when the content genuinely needs more columns; never shrink to hug narrow content, centre it instead.
• Height is a rough guess. The renderer crops to the true bounds, so overestimate rather than clip.
• No <script>, no onclick, no sendPrompt, no <a href>, no gradients, no filters, no drop shadows, no emoji, no rotated text, no <!-- comments -->.
• <defs> holds the arrow marker and nothing else (plus a <clipPath> or one <linearGradient> in illustrative diagrams).

Classes — the only styling you write

Put c-* on the group that directly holds the shape and text. It uses direct-child selectors: an extra nested <g> breaks the fill.

Every <text> needs t, ts or th — an unclassed one renders as raw black.

Colour the nodes, not just the containers

Every node carries a ramp by default. Nodes of the same kind share one; a diagram where the boxes are all neutral and only the containers are tinted looks drained and flat — that is the single most common way these diagrams go ugly.

class="box" is neutral on purpose, and it has exactly one job: homogeneous sub-parts inside one coloured container, where the container already carries the category and tinting the children would invent distinctions that do not exist. Do not generalise that pattern to a whole diagram.

One ramp per category, not per row — colour must never track sequence. A six-category diagram legitimately uses six ramps; a three-step flow uses one or two. c-gray for neutral, structural or external nodes. Prefer purple, teal, coral and pink for generic categories; reserve blue, green, amber and red for genuinely informational, successful, warning or error concepts.

Composition — decide these before you draw

The four judgements the audit will never make for you. Get them right here, then confirm them against the rendered image with the review checklist above:

• Uniform grids are a failure mode. If every tier holds the same number of identically sized boxes, the layout carries no information and reads as a spreadsheet. Vary box widths to match their content and vary the count per tier. Convergence, branching and a full-width emphasis row are what make a diagram look like a structure rather than a table.
• Find the feedback edges. Anything with a loop in it — an agent, a control system, a retry path, a training cycle — drawn as a one-way stack is not simplified, it is wrong. Route the return edge as a leader path around the outside, or lay the stages in a ring when the cycle has no natural entry point; references/diagram-types.md carries the polar geometry for that.
• Not everything is a stack. Before placing a box below another, say the dependency out loud: "the lower one is used by the upper one." A component the core talks to bidirectionally belongs beside it with a two-headed arrow (marker-start and marker-end), not underneath it.
• Find the one thing the diagram is about and give it more weight — a container, a wider box, the only saturated colour in a neutral field. A diagram where everything is equally prominent has no subject.

Geometry

• Box width from the longest label: max(title_chars × 8, subtitle_chars × 7) + 24. At 14px a character is ~8px; at 12px ~7px. CJK glyphs are ~14px. Formulas, subscripts and symbols run 30–50% wider — pad generously.
• Heights: 44px single line, 56px title + subtitle, +20px per extra subtitle line. Keep every box in a tier the same height.
• Spacing: ≥20px between boxes in a tier, 40–60px between tiers, ≥20px padding inside a container, 10px between an arrowhead and the box it points at.
• Tier packing: sum the widths and gaps before placing. Four 130px boxes with three 20px gaps is 580px — fits. Four 160px boxes is 640px — does not.
• Centre text: x = rect_x + w/2, y = the centre of its own line, with text-anchor="middle" dominant-baseline="central".
• <text> never wraps. Each line is its own <text> or a <tspan> with an explicit x and dy. If a subtitle needs wrapping it is too long.
• Connectors: a straight line that would cross an unrelated box gets an L-bend instead — M x1 y1 L x1 ymid L x2 ymid L x2 y2 with fill="none". Feedback loops route around the outside as a leader path; never draw the cycle as a ring.
• Prefer one dominant direction — all top-down or all left-right.

Text budget

Subtitles are ≤5 words. Detail belongs in your prose reply, not inside the box. Sentence case everywhere — never Title Case, never ALL CAPS. Only two sizes exist: 14px and 12px. Avoid floating labels: every <text> sits inside a box or is a leader callout.

Diagram types

• Flowchart — steps, decisions, transformations. "what are the steps", "what happens when…", "what's the flow".
• Structural — containment and architecture; things inside things. "what's the architecture", "how is this organised", "where does X live".
• Illustrative — a spatial metaphor that builds intuition. "how does X actually work", "I don't get X", "give me an intuition for". Freeform shapes, colour encodes intensity rather than category.

Route on the verb, not the noun: "transformer architecture" is structural, "how does attention work" is illustrative. references/diagram-types.md has the full playbook and copy-ready snippets; references/design-system.md has the ramp table and the detailed spacing and typography rules.

Database schemas and ERDs are not this skill's job — emit mermaid erDiagram instead.