blueprint

Render an interactive, self-contained HTML blueprint app from a JSON spec, using a frozen renderer (tokens + layout + emitter) so the drafting-paper look, generated OKLCH palette, swimlane headers, ph

SKILL.md

GitHub

Sequence diagram

Render real sequence diagrams as one self-contained interactive index.html. Never hand-draw ASCII sequence diagrams — they collide once labels have real text. The renderer is frozen: scripts/tokens.py (the single source of truth for every colour/size — never hardcode style elsewhere) → scripts/layout.py (pure geometry) → scripts/render_html.py (SVG/HTML emitter) + scripts/assets/ (app.css/app.js/page.html). You only ever write JSON data, so the look is identical every run, on every project. Design legend: STYLE.md. PNG is the in-browser ⤓ export (no separate PNG renderer).

The output is a small app: a clean header bar with a scenario dropdown (when there are 2+), the active diagram full-bleed (sticky swimlane header, the body scrolls), and a bottom toolbar of colour lenses + a PNG export. Click any message for its detail.

Two modes

A. Ad-hoc — one diagram, right now. Write a spec (schema below) and render:

python3 ~/.claude/skills/blueprint/scripts/render_html.py spec.json index.html

A spec is either a single scenario {title,actors,messages,…} or a master {project,scenarios:[…]}. Copy examples/web-request.json (single) or examples/master-backdrop.json (master). Open the HTML to view; deliver the file.

B. Project mode — the SET, cached + self-freshening. One master file holds every scenario and renders to one index.html (a card per scenario, switchable via the dropdown). Discovery is a dynamic fan-out: list the flows once, let the user pick, then spawn ONE agent per chosen flow (count scales to the list). A git check means you research once, not every time, and adding scenarios later never touches the ones already drawn.

Project-mode runtime flow (discover → choose → fan out → merge)

  1. Locate the master (convention: agent_docs/diagrams/architecture.json, else docs/diagrams/).
  2. Build the candidate flow list:
    • No master (first build): spawn ONE agent with enumerate-prompt.md at the repo (pass the existing ids = none) → a JSON list of every flow.
    • Master exists: run the git freshness check AND spawn ONE enumerate-prompt.md agent to spot brand-new flows (pass it the existing scenario ids so it returns only what's missing). Candidates = changed-known + new.
      python3 scripts/freshness.py master.json [repo]   # fresh | stale+list | no_provenance
      
      If freshness is fresh AND enumerate finds no new flows → skip to step 6 (render only).
  3. Choose — the gate (default ON): show the candidates as an AskUserQuestion multi-select with a "draw all" option; draw only the ticked ones. Override: if the user's request already said "one shot" / "draw all" / "everything", skip the picker and take the whole list.
  4. Fan out — the dynamic workflow: use scripts/draw_workflow.js as the template — bake the chosen flows (+ repo/masterPath/skillDir) into it as literals, then write the filled-in script to a file and launch it with Workflow({scriptPath}). It spawns ONE agent per flow (each reads only that flow's source_paths + reuses the master's actor vocabulary) and returns {scenarios:[…]} — one v3 scenario object each. ⚠️ Bake the flows in as literals in the script body — do NOT pass them via the Workflow args field (observed: args does not reach the script, so it runs with 0 flows). And prefer a scriptPath file over passing the whole script inline — inline embedding can mangle the prompt strings and trip the parser. (Tiny repo / no workflow wanted? discovery-prompt.md still does list+draw in one agent.)
  5. Merge: add/replace those scenarios in the master by id; leave every other scenario byte-for-byte. (Workflow scripts have no filesystem — the main loop writes the file.)
  6. Validate → stamp → render:
    python3 scripts/validate_master.py master.json   # must pass
    python3 scripts/stamp_provenance.py master.json [repo]
    python3 scripts/render_html.py master.json agent_docs/diagrams/index.html
    
  7. Open the HTML (or screenshot it headless) before sending; commit the JSON + HTML.

Spec schema (v3 — actors are PER SCENARIO)

{ "project": "Acme",
  "provenance": { "git_sha": "GIT_SHA_AT_BUILD" },
  "scenarios": [
    { "id": "capture", "title": "…", "subtitle": "…",
      "meta": { "created": "2026-06-26", "updated": "2026-06-26" },
      "actors": [
        { "id": "user", "label": "User", "zone": "user", "sub_caption": "person" },
        { "id": "phone", "label": "Phone", "kind": "internal", "zone": "client",
          "sub": [ {"id":"ui","label":"Camera UI"}, {"id":"up","label":"Upload"} ] },
        { "id": "srv", "label": "Server", "zone": "server", "sub_caption": "/enrich" },
        { "id": "claude", "label": "Claude", "zone": "llm", "sub_caption": "Haiku" }
      ],
      "messages": [
        { "phase": "CAPTURE" },
        { "from": "user", "to": "phone.ui", "text": "press & hold", "paths": ["capture"] },
        { "self": "phone.ui", "text": "freeze still", "caption": "→ base64 JPEG" },
        { "phase": "ENRICH" },
        { "from": "phone.up", "to": "srv", "text": "POST /enrich", "caption": "+ ID token",
          "via": "api", "paths": ["enrich"], "src": "EnrichClient.swift → routes.ts" },
        { "from": "srv", "to": "claude", "text": "vision", "caption": "narration + image",
          "via": "api", "paths": ["enrich"], "metrics": { "cost": 0.011, "latency_ms": 900 } },
        { "from": "claude", "to": "srv", "text": "enrichment", "caption": "title, blurb, OCR",
          "kind": "ret", "metrics": { "cost": 0.011, "latency_ms": 900 } },
        { "from": "srv", "to": "phone.ui", "text": "push ready", "kind": "async", "via": "api" },
        { "note": "srv", "text": "enqueue wiki compile" }
      ],
      "fragments": [ { "kind": "opt", "label": "voice note present", "range": [1, 2] } ]
    }
  ] }
  • actors (per scenario): {id,label,zone} (+ optional kind, sub_caption, sub:[{id,label}]). Left→right order is computed automatically from zone role (user → our client → our server → external third-party, busiest-first) — author them in any order. sub = sub-services, each a lifeline under the parent band (cap ~3). sub_caption is a one-line caption under a no-sub actor.
  • zone = ROLE, drives ordering + the generated colour: user; client/frontend/mobile/ ui; server/api/worker/gateway; everything else (vendors, datastores) = external. kind:"internal" also sits in the server tier. Colour is positional OKLCH — no per-zone hex.
  • messages (ordered):
    • {phase:"LABEL"} — a section divider (vertical label in the left rail). No text/refs.
    • {from,to,text} — solid request (filled head = sync call). "kind":"ret" = dashed + open head (return); "kind":"async" = solid + open head (fire-and-forget / event). This is the UML arrowhead vocabulary — the mechanics channel. Address a sub-service as actorId.subId.
    • "via":"api"|"io" — the interaction category → a small DRAWN marker (no marker = a plain in-process call). api = crosses a process/network/trust boundary (HTTP/RPC/IPC/3rd-party SDK); io = a data-store read/write (DB/cache/file/blob). Defined by the boundary crossed, not the tech — so it means the same on web, mobile, embedded, a compiler, an ETL job. Composes with kind (a store-read return = kind:"ret" + via:"io").
    • "caption":"…" — an optional soft secondary line UNDER the primary; put params / state / thresholds / model names here so the primary stays one short clause.
    • {self,text} — self-loop (= on-device process).
    • {note,text} — a collapsed note: a folded-page icon + a SHORT tag (text, ≤ ~24 chars); put the long explanation / file refs in src (revealed on click). For a genuine caveat the arrows can't show (a scope boundary, "CPU-only path") — use sparingly, never as a scratchpad.
    • "metrics":{"cost":0.011,"latency_ms":900} — enables the $ cost / ⏱ latency lenses (a perceptual graphite ramp; absent data → that lens isn't offered).
    • "paths":["enrich"] — tags the message to a named flow → a ↪ enrich highlight lens.
    • "src":"File.swift → route.ts" — the caller→callee files; shown in the click-detail footer.
    • "detail":{…} — the AI-prefilled body of the click-detail card (everything the arrow can't show). All six keys optional; emit only for load-bearing messages (api/io/metric'd/path-tagged or otherwise non-obvious) and omit any key you're unsure of — blank is correct, never guess. Keep each a tight clause (≤ ~140 chars). The card already shows the readable route, step N/M, phase and any fragment guard from the spec — so detail must NOT restate the label, the participants, or the on-arrow caption. Keys:
      • why — the card's headline: why this call exists, in plain language (NOT a paraphrase of the label). E.g. label DELETE /account → why Erase the user; token-auth'd; cascade-safe.
      • effects — what it mutates/emits as a result ("deletes 3 Firestore subcollections, then auth").
      • fails — behaviour on error (retry, fallback, throw, partial-state safety) — the 3am question.
      • sends · auth · ordering — short chips, not sentences: the payload/credential token (Bearer ID token), the identity that crosses (Firebase ID token, verified first), and any sequence constraint (runs LAST). Put the ordering/auth fact in its chip — don't also repeat it in fails/effects prose.
  • fragments (per scenario): [{kind,label,range:[a,b],segments?:[{guard}]}]. kind ∈ opt|alt|loop|par|break|critical|ref. range indexes the NON-phase message order (a,b inclusive). segments adds divider guards (for alt/par). A box must wrap ≥2 messages — for a single message, fold the condition into that message's caption instead (a 1-message loop, which asserts repetition, is the only exception). The validator warns on a degenerate box.
  • meta (per scenario, optional): {created,updated,status} — informational.

Conventions — label grammar (keeps every diagram readable on any codebase)

  • One short clause per primary text (≤ ~30 chars / ~4 words → renders as ONE line, no wrap). Detail (params, state, thresholds, model names) goes in caption. Never put ·/— or math in a primary — the validator rejects ·/— there and warns past the length cap.
  • Phrase the primary by category so types read at a glance:
    • in-process call → name() or verb noun (render cells)
    • via:"api" → VERB target (POST /enrich, Auth.Login)
    • via:"io" → read/write/query X (read blob)
    • async/event → emit/on/enqueue X
    • return (kind:"ret") → the value as a noun (JPEG bytes, 200 OK)
    • self/process → an imperative phrase (build quadtree)
  • Tag category with via, direction with kind; they compose. Pair every request that returns with kind:"ret". Put metrics on the paid/heavy call (and its return) — it's the metric, not a boolean. Use phase to chunk a long flow; paths for flows worth tracing.
  • Aim for ≤ ~7 lifelines and ≤ ~20 messages per scenario; split if larger.
  • Only diagram real behaviour — read the code, don't invent.

Extending (keep the style frozen)

Every visual feature is a JSON field consumed by render_html.py, never caller-side drawing code. New token/colour → tokens.py (then STYLE.md). New behaviour → layout.py geometry + render_html.py emit. The author side never changes how things look.