Interactive Architecture Diagrams

Build single-file, drop-in HTML pages that let workshop attendees, clients, or new team members click through how a system works — step by step, with animated data packets flowing between nodes, payload details on a side panel, and toggleable modes (offline/online, dev/prod, v1/v2). The aesthetic is dark and didactic-first: bounded nodes, gentle quadratic wires, packets that glow only on the active step — no rainbow gradients. Rebrand via the CSS tokens in assets/css-tokens.css.

When to reach for this skill

The differentiator is interactivity + sequenced data flow: if the value is "click through it and watch what happens", use this skill; if you need a static image, a slide, or an inline diagram, use Mermaid instead.

The mental model

Every diagram has four things:

1. Nodes — services, datastores, users, queues, external systems. Each has a role (color) and metadata (tech stack, port, deployment target).
2. Flows — named scenarios the user can pick (e.g. RAG Query, Direct Query, Ingest, Auth). Each flow is an ordered list of steps.
3. Steps — {from, to, color, title, route, payload, desc, chips}. Each step lights up one wire and one target node. (color and title are required — a missing color leaves the wire uncolored, a missing title leaves the panel header blank.)
4. Modes — orthogonal toggle (offline/online, dev/prod, v1/v2). Modes can:

- hide/show entire nodes (e.g. Seed only exists offline) - rename a node (Qdrant → BigQuery) - swap payload bodies, ports, auth headers, latency chips

Modes are NOT alternative flows. Flows describe scenarios ("user asks a question"); modes describe deployment shape ("on Docker" vs "on Cloud Run").

Workflow when invoked

Step 1 — Capture intent

If the user has uploaded a SOLUTION.md, README, OpenAPI spec, or any architecture description, read it first before asking questions. Extract:

• Service names + tech stack + ports
• Deployment modes (if any)
• Named endpoints/operations and what they do (these become flows)
• For each operation, the chain of internal service calls (these become steps)

Then ask ONLY the gaps. Don't ask things already in the doc.

If there's no doc, ask the user briefly:

• "What system are we drawing?"
• "What are the main flows you want to show? (e.g. login, search, checkout)"
• "Any mode toggles? (offline/online, dev/prod, v1/v2)"
• "Who's the audience? Workshop attendees? Engineers? Clients?"

Step 2 — Plan the topology on paper first

Before writing code, sketch on paper (literally in your head or a scratch file) WHICH services exist, WHERE they sit relative to each other, and WHICH flows connect WHICH nodes. Avoid wire crossings — this is the single biggest readability win. See references/flow-design-patterns.md for layout heuristics.

Quick rules:

• Orchestrator/API gateway in the middle, dependencies fanning out
• User on the far left (entry point), datastores on the right (exit), services in between
• Read-only services above, write-heavy services below (intuitive: data flows down into them)
• Optional/one-shot services (seed jobs, cron) tucked in a corner

Step 3 — Build from the template

Copy assets/template.html to the working directory. It's a fully working drop-in: dark theme, side panel, player controls (play/pause/step), keyboard shortcuts, animated SVG packets, mode toggle. Edit only these regions (the template's top EDIT THESE THINGS comment is the authoritative list):

1. <title> + <h1> — replace {{SYSTEM_NAME}} and the header text.
2. .modepick buttons — mode toggle labels (or remove if single-mode).
3. .flowtabs — one .flowtab per scenario; data-flow must match a flows key.
4. .node divs in .stage — one per service: data-id, data-role, style="left/top", label, tech, port.
5. The flows object in JS — one entry per scenario, each with an ordered steps[] array.

Conditional: update the .legend only if your roles differ from the defaults; recolor :root CSS variables only for a restyle. After editing, search the file for {{ to catch any stray {{SYSTEMNAME}}/{{FLOW*}} placeholders.

Don't reinvent the CSS, the player, the SVG wire renderer, or the side panel — they all work.

Step 4 — Validate before showing the user

After generating, do this self-check:

• Open the file mentally: does every flow's step sequence make narrative sense? (Step 1 sets up, last step delivers a result to the user.)
• Do wires cross unnecessarily? If yes, reposition nodes (try moving the late-arriving dependency away from the early ones).
• Are there flows where a single node is the to of two consecutive steps? That's fine if it's a loop, but check the description explains why.
• Mode swap: pick the most different node between modes (usually the database) and verify ALL its mode-dependent fields are filled in (label, tech, port, payload, chips).
• Did you generate the companion architecture.md? (Required — see Step 6.)

If puppeteer is available in the environment, take a screenshot to spot-check layout. Code in references/screenshot.js.

Step 5 — Output the files

Always two files:

• architecture.html — the interactive diagram
• architecture.md — text description (components, flows step-by-step, mode differences, scenarios for the workshop)

Save to the current working directory. Platform-specific overrides:

• Claude.ai — use /mnt/user-data/outputs/ instead
• Claude Code / Gemini CLI / OpenCode / Copilot CLI — save to the current working directory (default)

Both should be self-sufficient — someone can read the markdown without opening the HTML, and vice versa.

Step 6 — Present the output

Tell the user where the files were saved and how to open them (open architecture.html on macOS, xdg-open architecture.html on Linux, or just open in browser). If present_files is available (Claude.ai), call it to display the HTML first, then the markdown.

File reference

**The examples/*.json files are planning specs, not drop-in code** — JSON outlines you translate into the template's HTML nodes + inline JS flows object. See references/flow-design-patterns.md → "Example JSON → template mapping" for the field-by-field translation.

Quick reference — node roles and colors

Hardcoded in CSS variables in template.html. Pick one per node:

If you need more roles, add CSS variables. Keep the count ≤ 6 — past that, the legend becomes a colorblind nightmare.

If your system has e.g. two distinct datastores (Postgres + Redis), still use vector for both but distinguish via the icon and label, not by inventing a 7th color.

Visual states — what the viewer sees on each step

This is the most important didactic mechanic in the diagram. Understand it before modifying the JS or CSS, or you'll break the "click a flow tab, see the whole path" intuition.

When a user picks a flow tab and lands on step N, every node and wire is classified into one of these states:

Why participant matters (and the JS/CSS implementation gotchas) is detailed in references/flow-design-patterns.md → "The participant state". Short version: it lets a viewer see the whole path on the first click, instead of stepping through to discover which nodes a flow uses — that's the entire point of the diagram. If you add node states, keep applyStep and the CSS (.node.participant, .wire.preview) consistent.

Interaction model

The diagram supports five ways to interact, in increasing order of viewer expertise:

Node click vs drag. A 5px movement threshold distinguishes a click (jump to step) from a drag (reposition); pointer events handle mouse and touch alike.

Node click is the easiest to miss when modifying the skill. Three things must be in place:

1. cursor: pointer on .node (visual affordance — "this is clickable")
2. :hover state that reveals the role color (confirms interactivity)
3. The JS handler that does findIndex on the current flow's steps and jumps there

If a node isn't part of the current flow (.dimmed state), clicking it shakes the node instead of doing nothing. Silent ignore is bad UX — the viewer thinks the page is broken.

Keyboard accessibility: nodes have tabIndex=0, Enter/Space triggers the same action as click.

Don't break this when adding features. It's the only way for a viewer to think "wait, where is the LLM used?" and find out instantly. Without it, they'd have to step through all 8 steps just to discover one node's role.

Quick reference — payload format

Payloads in the side panel are monospace + minimal syntax highlight. Keep them realistic but trimmed — show the shape, omit boring fields:

{
  payloadOffline:
`POST /ask
{
  "query": "Ile kosztuje parking?",
  "limit": 3,
  "temperature": 0.7
}`
  // avoid: full HTTP headers, User-Agent, 50-line bodies — show only the shape
}

Use 4-space indent, no tabs. Use // comments for inline annotations.

If a step is the same in both modes, use payload (single key). If they differ, use payloadOffline + payloadOnline (or whatever the mode names are).

Quick reference — chips (metadata badges)

Chips below the payload are short metadata labels. Two semantic types auto-style themselves:

• Latency (matches /\d+\s*(?:ms|s|min)\b/) — gets a stopwatch icon, e.g. "~150ms", "5-30s", "timeout 120s"
• Size/count (matches /dim|×|loaded|rules|count/) — gets a database icon, e.g. "768 dim", "38 loaded", "×N"
• Other — neutral, e.g. "COSINE", "baseline", "idempotent", "UUID5"

Maximum 3 chips per step. They're hints, not specifications.

Naming conventions

Common pitfalls

</script> in payloads breaks everything. If a payload template literal contains the string </script> (e.g. showing an HTML snippet), the browser's HTML parser treats it as the closing tag for the main <script> block — killing all JS. Escape it as <\/script> or </script> inside template literals. This also applies to any </style> or similar closing tags in payload strings.

Source node looks dead. If a step's source node renders dimmed instead of active-from, you regressed the participant logic — see "Visual states".

Wire crossings. Caused by node placement, not code — swap node positions to fix. (The template auto-curves wires in alternating directions per step index, which helps but won't fix a bad layout.)

Too many flows. More than 6 flows clutters the tab bar. If you have 8+, group them: one diagram per group (e.g. "read flows" vs "write flows" as two separate HTML files).

Modes that aren't really modes. If the toggle changes the entire topology (different services, different protocols), it's not a mode — it's a second system. Build two HTML files.

Static screenshots beat live diagrams sometimes. If the diagram will end up in a PDF report, a slide deck, or a printed handout, this skill is the wrong choice — produce a static diagram instead. This skill is for interactive consumption in a browser.

Forgetting the .md. The HTML is the showpiece, but the .md is what people read on GitHub, paste into Slack, and grep for endpoint names. Always generate both.

Polish vs English. Match the workshop language. Bilingual settings typically use Polish for internal workshops and English for client deliverables. Ask if unsure.

Extending the template

If a user asks for something the template doesn't support, here's the difficulty ranking:

Don't promise features the template doesn't support without flagging the work involved.

Iteration tips

When the user gives feedback like "the diagram is too cramped" or "step 5 is confusing":

• "Too cramped" → spread nodes wider; you have 100% width to work with. Don't go below 12% horizontal gap between adjacent nodes.
• "Confusing step" → the step description is doing too much. Split into two steps OR rewrite to focus on ONE handoff (X sends Y to Z, that's it).
• "It doesn't match how it really works" → ask for the actual code path. Don't guess; the diagram has zero value if it lies about reality.
• "Make it pop more" → resist. Didactic clarity > visual punch. Maybe one accent element (a glow on the result node when flow completes) but stop there.
• "Add a legend" → already there. If they want it bigger, increase the .legend font-size and gap.