Arcgram

Arcgram — Interactive Architectural Diagram (v2)

For any AI agent — Arcgram is agent-neutral, not a Claude-only skill. The block above is ordinary YAML frontmatter: Claude's skill system reads name/description to auto-trigger it, and any other agent (GPT · Gemini · Cline · Continue.dev · Cursor · Aider · local) just reads the plain-markdown body — no rejection, no Claude dependency. This is the v2 skill — engine template-v2.html, fields in schema.md, per-platform install in USAGE.md.

What this is

Arcgram is a human↔AI collaboration substrate: the AI externalizes its reasoning as a typed node-and-edge graph the human audits and adjusts, converging on ground truth that survives across turns, sessions, and agents. The AI draws, the human audits. (Full positioning: README.md.)

Three guards keep the drawing honest — internalize them before you build:

• A node is a diagram element — a component, step, decision, artifact, or tool. It is not an HTML/DOM/React node, and not an argument-map "claim." An edge carries flow or dependency — not "supports/contradicts."
• The gap signal: if a step can't be drawn as explicit nodes and edges, that's reasoning with a gap to expose — not a formality to skip.
• Drawing exposes gaps; it does not certify correctness. A graph you can draw can still be wrong.

When to use

Reach for Arcgram when the structure of a thing — or of your reasoning about it — gets hard to hold in plain text:

• a system map, dependency graph, topology, architecture, or data/work flow
• "show me how X connects to Y" — relationships or lifecycles hard to track in prose
• a decision / thinking flow (branches, forks) you want checkable, not just described
• several people or agents need one shared, point-at-able source of truth

Do NOT use for

Charts (use a chart tool) · sequence diagrams (use a sequence-diagram tool) · mind maps · auto-layout graphs with hundreds of nodes (this skill assumes deliberate manual placement).

Files in this skill bundle

Choose the layout first (pin before building — it's hard to undo)

Pin the layout before you draw — later edits build on it. There are two decisions, and the agent should derive a recommendation from the flow's logic and help the user confirm it rather than leave them to guess:

1. Shape. Do the nodes fall into parallel categories (subsystems, tracks, layers that coexist and wire across each other)? Yes → group them: COLUMNS (vertical) or BANDS (horizontal). No (one causal/branching chain) → free DAG: leave COLUMNS/BANDS empty, each node's role on its cat.
2. Decisions. Does it branch on yes/no forks? Yes → add diamonds (kind:'diamond') — an overlay on any shape, not a separate option. See schema.md § Decision diamond.

Orientation (V vs H) applies only when grouped and is the user's preference — so the agent proposes shape + decisions and asks the user the reading direction (default V): top→bottom story (lifecycles, pipelines) → COLUMNS; left→right stages/tracks → BANDS (layout-tips.md §10b). If the platform has a selection UI (AskUserQuestion), use it; otherwise ask inline and wait. ⚠ Never force a no-category chain into COLUMNS/BANDS — the tags float off their nodes. Keep the same mode on later edits unless the user asks to switch.

Workflow

Hard gate — read before you draw. Before you place a single node, read schema.md (every field) and layout-tips.md (routing + positioning), and open one file in examples/. Authoring from memory of how generic diagram tools look — skipping these reads — is the single biggest cause of broken graphs: drifted field names, all-to-all edge meshes, and wires routed straight through nodes. Skipping the reads is the bug, not a shortcut.

Decide Audit up front. Ask the user whether they want an Audit defect review of the finished flow (§ Audit). Settle it before you draw — never auto-annotate defects with icons.

1. Understand the topology

Group nodes by category, decide reading direction, mark the critical connections. Draw only the edges that exist in the real structure — never wire every node in one group to every node in the next (a mesh). If a band-to-band connection comes out all-to-all, a structural element is missing (usually a convergence node — many inputs feed one decision); see layout-tips.md §10b. Logic-check the draft (step 7, pre-draw) before you place anything.

2. Read the references once

schema.md for every field, layout-tips.md for positioning — read them instead of inlining a schema from memory. schema.md is gated against the engine (schema-xref), so any drift is caught.

3. Pin layout + sizing

Pin the shape before drawing (§ Choose the layout first). The engine sizes nodes from their labels — you don't set w/h except on diamonds. Plan spacing per schema.md § Sizing rules of thumb + layout-tips.md §3.

4. Fill the DATA SECTION

Scaffold a copy into output/ — node new-flow.mjs my-flow → output/my-flow.html — (or copy template-v2.html by hand). Then in the DATA SECTION set <title>, the title lines, W/H, and fill nodes[], edges[], and optionally BANDS[], COLUMNS{} / HCOLUMNS[], CANVASES[], STATUS_LEGEND{}. Minimal shape (full fields: schema.md § Node / § Edge):

// node — required: id, label, cat, type, desc, x, y
{ id:'FEED', label:'Feeding', cat:'CORE LOOP', type:'core',
  role:'action', desc:'Tap to feed.\nRestores hunger.', x:220, y:120 }
// edge — required: f, t, fromPt, toPt
{ f:'FEED', t:'FOOD', fromPt:'bot', toPt:'top', route:'vhv', style:'solid' }

Declare node ownership on both sides — a node's band/free (plus column/colFree in H) and the band/column members[] must agree (gated A27/A28). The engine does the rest: bands auto-fit, lanes auto-grow, loose entry nodes wrap in a LIFECYCLE band. (schema.md § Node ownership · § Column ownership · § Auto LIFECYCLE band · § Canvas filter.)

5. Features & styling — all optional, all in the references

Add only what the flow needs; every field is defined in schema.md, every heuristic in layout-tips.md:

• Node types — built-ins + the custom-type recipe: § Adding a custom type.
• Edge routing — vhv/hvh/3leg per edge; pick fromPt/toPt so lines exit/enter clean (arrowhead points at toPt): § Routers + layout-tips.md §6, §9.
• Edge styling — solid/dashed/bold + crit: § Edge styling decision matrix.
• Critical paths — crit:true + lbl:'🔑N +verb', 3–7 per 30 edges: layout-tips.md §8.
• Decisions — kind:'diamond' + branch:'Y'/'N': § Decision diamond.
• Status dots + author flags — status (+ STATUS_LEGEND) and flag (a lightweight per-node "look here" note): § Status dot · § Audit marker. For a real defect review — bugs, logic holes, geometry defects like a wire through a node — use the Audit extension, not flag icons (§ Audit).

6. Render — three modes (fall back in order; don't skip ahead)

• Mode 1 — inline interactive widget. Render template-v2.html via the host's HTML tool (e.g. mcp__visualize__show_widget). Pan/zoom/hover native — best for iteration.
• Mode 2 — inline static SVG (your own output, not a tool call). If Mode 1 is unavailable, write an <svg>…</svg> block directly in your reply, using arcgram conventions (bands as rects, nodes border+dark bg, vhv/hvh/3leg edges, pills, crit red). Don't switch to a different diagram tool — the chat-native SVG routes around the failed tool.
• Mode 3 — standalone HTML export. Write template-v2.html with your data to a file the user opens. Always available with filesystem access; use for handoff.

7. Self-check — run the validators (headless Node, under skills/<tool>/)

The diagram is a spec — check it, don't eyeball the render. Pre-draw (as soon as nodes/edges exist) — Checkpoint catches dangling edges, orphans, a diamond missing a branch, duplicate ids:

node skills/checkpoint/checkpoint.mjs your.html

Post-draw — Validate runs Checkpoint always, and Reconcile too when you pass matching code (a no-code run reports INCOMPLETE, never a silent pass):

node skills/validate/validate.mjs your.html [code-file ...]

The engine also self-checks on render (stamps self-check ran … | N findings into the subtitle); the CLI is the enforce-by-construction version.

8. Save + syntax-check (Mode 3)

Extract the inline JS and syntax-check before opening — catches typos that silently blank the canvas:

awk '/<script>/{flag=1;next}/<\/script>/{flag=0}flag' your.html > /tmp/x.js && node --check /tmp/x.js

Open in browser. Confirm.

Audit — defect review (opt-in; use the extension, not flag icons)

Audit is a defect-review overlay for a drawn flow: a pulsing ring + a short reason tag on each marked node or edge, plus an "Audit" findings list. It self-hides when there are no marks, so a clean flow looks identical with or without it.

Ask first, mark only when asked. Offer the user an Audit pass and settle it up front — do not auto-annotate defects.

Use the extension — not the flag field. The Audit UI is extensions/arcgram-bugmarks.js. The per-node flag field draws a small icon but is not the Audit overlay — it must never stand in for a defect review. Load the extension after the engine (inline the <script> for a self-contained file), then call ArcgramBugs.set([...]):

<script src="extensions/arcgram-bugmarks.js"></script>
<script>
  ArcgramBugs.set([
    { level:'high',   anchor:{ node:'NODE_ID' },       reason:'wire crosses node', desc:'route around — a wire through a node is a level-1 defect' },
    { level:'medium', anchor:{ edge:{ f:'A', t:'B' } }, reason:'authored concern',  desc:'…' }
  ]);
</script>

level marks the source: high = a gate / geometry defect (e.g. a wire through a node), medium = an authored concern, low = an agent advisory. Anchor each mark to a node:'ID' or an edge:{ f, t }. Full API + fields: extensions/README.md.

Hard rules

Non-negotiables. Adjust everything else freely; these hold on every diagram.

• Write permission (two halves). The agent edits only the DATA SECTION — everything below END OF DATA SECTION is the engine, off-limits. AND the agent proposes; the human holds final write-authority — draft and adjust the diagram data freely, but do not commit it as final shared ground truth without the human's confirmation.
• Author from schema.md, never from memory of what generic diagram tools look like. Read this body + schema.md before generating; name fields, don't invent them.
• Single source. schema.md owns every field; README.md owns the positioning. Point to them — don't restate a field definition or re-argue the job inside this file.
• Keep attribution. The "Made with Arcgram" mark + NOTICE stay intact when you redistribute (Apache §4(d)); an unbranded / no-attribution build needs the commercial license (WATERMARK-AND-COMMERCIAL-TERMS.md).

License

Apache-2.0. Attribution rules → § Hard rules (the canonical copy); full terms, trademark, and the unbranded commercial build → LICENSE · NOTICE · WATERMARK-AND-COMMERCIAL-TERMS.md.