Create an Excalidraw diagram or visual — architecture, flowchart, mind map — including casual 'draw this' / 'show me visually'. Also use when arc-maintaining-obsidian delegates Synthesis visuals beyond embedded Mermaid.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Create an Excalidraw diagram or visual — architecture, flowchart, mind map — including casual 'draw this' / 'show me visually'. Also use when arc-maintaining-obsidian delegates Synthesis visuals beyond embedded Mermaid.
category
knowledge
status
promoted
arc-diagramming-obsidian
Diagrams should ARGUE, not DISPLAY. A diagram is a visual argument — structure mirrors concept, shape carries meaning. If removing all text leaves a meaningless grid of boxes, the diagram has failed.
Pipeline
DESIGN → BUILD → VALIDATE → SAVE
You own the full workflow. For complex diagrams, delegate mechanical phases (Build, Validate, Save) to subagents that read from agents/ to keep context clean; for simple diagrams or when subagents aren't available, execute each phase yourself. The instructions below are self-contained either way.
Guidance splits into two layers, both applying across all four phases: HARD — physical/mechanical invariants the tools verify or that silently corrupt output; non-negotiable. SOFT — concept judgments that depend on what you're drawing; reasoning, not rules.
HARD: Invariants That Must Hold
Process Invariants
Detect theme before building. Choosing the wrong palette silently produces an unreadable diagram (dark palette on light background = pale washout; light palette on dark = near-invisible text). The Playwright renderer happily renders whatever colors you picked — no feedback loop to catch this.
If no response within 5 seconds, pick a palette from prompt signals ("dark mode" in the request, time of day) and state the assumption in your completion output. Never silently assume.
Every Validate iteration must render AND view PNG. You cannot judge composition, readability, or hierarchy from JSON alone. The overlap checker catches bounding-box overlaps but not visual problems — arrows crossing unrelated elements at acute angles, text too small to read, hierarchy failing because the hero doesn't dominate. View /tmp/diagram.png with the Read tool every iteration.
Save must be verified. Both ea.create() and the manual-fallback write path can silently produce format corruption — the canvas renders, markdown text bleeds through, and the Playwright renderer cannot detect this. Always run verify_saved_diagram.py after save.
Mechanical Invariants
ea.reset() at the start of every EA invocation — without it, elements accumulate from previous calls and appear in the output invisibly
Never change an element's id — this orphans connected arrows (they lose the shape they were bound to)
addText with box returns the BOX id, not the text id — use the returned value for connectObjects
viewBackgroundColor matches the detected theme — #1e1e1e for dark, #ffffff for light
Save format is byte-exact for the manual fallback path — any deviation (wrong frontmatter spacing, list-style tags instead of inline array, missing warning line) causes silent corruption. See references/save-format.md
Layout Trap Audit (Physical Collisions)
Four arrow-path collisions that recur across diagrams. Each is a rendered-pixel overlap, detectable by check_overlaps.pyafter build, but cheaper to prevent at design time by tracing arrow paths mentally.
Trap 1: Converging arrows (multiple arrows merging into one target) travel through the corridor below the source elements. Annotations placed directly below source boxes get crossed.
Trap 2: Back-edges (an arrow from a later zone returning to an earlier one) routed horizontally at the Y-midpoint of a zone they pass through will cross that zone's elements.
Trap 3: Decision-diamond "yes" / "no" labels placed at the bottom of the diamond sit directly in the fan-out exit path of any arrow leaving from the bottom.
Trap 4: Back-edge labels (free text describing the arrow) placed at the arrow's own vertical-run X-coordinate sit on top of the arrow line.
Mentally trace every planned arrow from source to target before writing EA code. If the path crosses an unrelated element, reposition before building. For fix strategies (arrow routing, anchor distribution, waypoint planning), read references/layout-heuristics.md Part 2.
SOFT: The Design Space
HARD keeps the diagram valid. SOFT is where you make it good — concept judgments that depend on what you're drawing.
Think first, draw second. For each major concept, answer before reaching for shapes: what does it DO (the verb)? What relationships and core transformation (input → output, state A → state B)? What would someone need to SEE to understand it?
Every element serves the concept. Before adding anything, ask: what does this communicate that labels alone don't? If "nothing", it's noise; if "this is where flow starts / a zone boundary / the concept is symmetric", it's doing work. Match two registers — language (monolingual prompt → monolingual labels, no bilingual subtitles unless asked) and conceptual (symmetric concept → equal-sized peers; don't fabricate a hero the concept doesn't claim).
Scale reflects real importance. Size the genuine hero (a convergence point, the concept the diagram is about) larger; size genuine peers equally. Size ranges in references/painters-toolkit.md are suggestions, not tiers.
Isomorphism self-check. Before building, ask: with all text removed, would the structure alone communicate the concept? A fan-out says "one source, many outputs"; a convergence says "many inputs, one result"; a cycle says "feedback". If the structure is just labeled boxes and arrows with no isomorphism, revisit. This is a check, not a gate — sometimes "A connects to B" is genuinely the concept.
Brushes — shape variety, subtitles, zone labels, accents, separators, containers, footers — are a vocabulary to pick from, not a menu to copy. See references/painters-toolkit.md. For pattern inspiration (fan-out, convergence, tree, timeline, cycle, assembly line, side-by-side, gap) see references/visual-patterns.md; if two adjacent sections look structurally identical, redesign one — visual monotony kills comprehension.
Depth and layout: for comprehensive/technical depth (research, multi-zoom, evidence artifacts) see references/depth-enhancements.md; for 20+ elements use references/plan_layout.py to compute coordinates automatically (systematic spacing + two-column separation that prevents common overlaps).
Phase 1: Build with ExcalidrawAutomate
Create elements using the EA API via obsidian eval. EA handles text sizing and arrow binding automatically. The build ends by exporting all elements as a .excalidraw JSON file to /tmp/diagram.excalidraw.
Before writing EA code, read references/layout-heuristics.md Part 1 for grid-based coordinate planning.
Beyond the Mechanical Invariants above: style before element (ea.style.* applies to the NEXT element created); stagger anchors when multiple arrows leave one shape (left/bottom/right, not all bottom); diamond text ≤ 12 chars (diamonds have ~50% less usable area than rectangles).
Read references/element-templates.md for the full EA API reference, raw JSON templates for Phase 2 fixes, and the binding checklist.
Mermaid shortcut: For simple flowcharts under 10 elements, use ea.addMermaid() instead. Only flowchart type produces native editable elements — other Mermaid types fall back to SVG images.
Phase 2: Validate (Mandatory)
After building, render to PNG, view it, fix defects. Up to 3 iterations — then save and report remaining issues.
ITERATION (repeat up to 3×):
1. CHECK — : "${ARCFORGE_ROOT:=$HOME/.agents/arcforge}"
cd ${ARCFORGE_ROOT}/skills/arc-diagramming-obsidian/references && \
uv run python check_overlaps.py /tmp/diagram.excalidraw
2. RENDER — : "${ARCFORGE_ROOT:=$HOME/.agents/arcforge}"
cd ${ARCFORGE_ROOT}/skills/arc-diagramming-obsidian/references && \
uv run python render_excalidraw.py /tmp/diagram.excalidraw \
--output /tmp/diagram.png --scale 2
View /tmp/diagram.png with the Read tool — HARD, every iteration.
3. JUDGE — Apply SOFT judgment. Design intent: correct patterns? Hero
dominant (if concept claims one)? Brushes from Painter's Toolkit
serving the concept? Defects: overlaps, crossings, uneven
spacing, text too small?
4. FIX — Edit .excalidraw JSON directly (Read → find element → Edit x/y).
Moving shapes does NOT break arrow binding (Excalidraw recalculates
from binding data, not coordinates).
Never change element IDs — this orphans connected arrows.
→ Next iteration, or proceed to Save if clean.
Read references/layout-heuristics.md Part 2 for fix strategies.
First-time setup (if renderer fails with missing deps):
Only use when obsidian eval code="typeof window.ExcalidrawAutomate" returns empty (EA plugin unavailable). Obsidian checks format heuristics — any deviation causes silent corruption. Read references/save-format.md for the byte-exact template.
Post-Save Verification (Mandatory)
: "${ARCFORGE_ROOT:=$HOME/.agents/arcforge}"cd${ARCFORGE_ROOT}/skills/arc-diagramming-obsidian/references && \
uv run python verify_saved_diagram.py <vault-path>/<name>.excalidraw.md
Exits non-zero on format corruption or render mismatch. For the manual-fallback path, also compares the re-rendered PNG against /tmp/diagram.png to catch JSON corruption. If verification fails, regenerate using the canonical template from references/save-format.md — not the file you just wrote.
Embed in Wiki Notes
Embed with ![[diagram-name]], placed outside bilingual callouts (diagrams are language-neutral).
Delegation (Optional)
For complex diagrams, spawn a subagent per mechanical phase to keep context clean. Pass the design spec and relevant context; each subagent reads its instructions and the reference files it needs:
Build — pass the design spec (theme, zones, elements, connections, brushes chosen); read agents/diagram-builder.md. Output: /tmp/diagram.excalidraw + element count.
Validate — pass the diagram path and a 1-2 sentence design intent; read agents/diagram-validator.md. Output: validated .excalidraw + PNG at /tmp/diagram.png + issues report.
Save — pass the diagram path, filename, folder, and embed target; read agents/diagram-saver.md. Output: vault path.
Quality gate: After validation returns, view the PNG yourself before Save. If it doesn't match your design intent, revise the spec and re-run Build, or give a new Validate pass specific fixes.
Do yourself: element generation, PNG render, overlap/layout checks (check_overlaps.py, plan_layout.py), direct .excalidraw.md vault write (or ea.create()). Delegate to obsidian:obsidian-cli: theme detection, vault-path/note search, embedding in a wiki note, reloading a changed diagram.
Reference Files (Read on Demand)
references/color-palette.md — Semantic colors for light + dark mode
references/visual-patterns.md — 9 patterns with layout guidance
