Excalidraw Diagrams

Excalidraw diagrams (via the Excalidraw+ MCP)

Build scenes that read as a deliberate visual argument, not "boxes with text," and get them right the first time by knowing the MCP's non-obvious behaviours up front. The element JSON here is standard Excalidraw, so it also applies if you ever generate .excalidraw files directly — but this skill targets the hosted Excalidraw+ MCP (shareable scenes, collections), which is the usual case.

The workflow

1. Plan the structure first (see reference/layout.md). Pick the family (workflow / architecture / timeline / hierarchy / comparison / cycle / slides), the reading direction, and the regions — before placing anything. The test: if you deleted every label, would the layout alone still convey the idea?
2. Call read_excalidraw_format once before your first write this session. It's the MCP's own normative element guide (occasionally updated) and the server expects it. reference/elements.md is your fast working subset.
3. Create or target a scene. New: create_scene (needs a collectionId — get it from list_collections). Existing: find it via list_scenes / get_workspace; to edit an earlier one, get real element IDs with search_scene_content. The shareable link is https://app.excalidraw.com/s/<workspace_id>/<scene_id> — both IDs come back in the create_scene metadata (workspace and id); give this link to the user.
4. Add content with edit_scene_content — one coherent diagram per call. Build in placement order: regions/frames → containers → primary nodes → secondary nodes → labels → arrows (bound, once nodes are stable) → accents.
5. Add icons where they help (reference/library-icons.md + the bundled script) — real Postgres/React/Docker/AWS icons, inserted via the MCP.
6. Verify — this is mandatory, not optional (see "Verifying" below). Screenshot for layout, search_scene_content for text/geometry, fix, repeat.

Non-negotiables (each one silently ruins a diagram)

These are the traps that make a diagram look broken or fall apart on the next edit. Full explanations in reference/mcp-gotchas.md — internalize these five:

• Screenshots don't render text. A screenshot that looks "empty" usually has all its labels. Never judge text presence/content from a screenshot — use search_scene_content.
• New elements can land under existing fills (the server assigns indices non-monotonically, even on a fresh delete-free build). Icons placed on cards, or cards placed in a box, can render invisibly under the opaque fill — often inconsistently (some show, some don't). Simplest robust fix: give grouping containers AND cards backgroundColor:"transparent" (keep the coloured border) — a transparent fill can never occlude, so z-order stops mattering for content inside a box, and a white card on a white canvas looked identical anyway. Only fall back to an explicit high index (read the current max via search_scene_content, don't trust a literal like "b01") when you need one opaque element atop another. See reference/mcp-gotchas.md.
• textAlign:"center" makes the x you pass the CENTER of a standalone label (not the left edge). To center under a node at center X, pass x: X.
• Bound labels are always vertically centered in their shape. For an icon-above-label block, use a standalone label positioned yourself, not a bound label.
• Arrows pointing at shapes must be bound (startBinding/endBinding, mode "inside"), or they detach on the next edit. A bound arrow label (label:{}) is created only at add time — adding one via update silently no-ops (added:0), so re-add an arrow to (re)label it, and read the label as the arrow's labelText (it's invisible to a types:["text"] search). You can't delete a shape and its bound arrow in the same call — delete the arrow first.

Verifying (do not claim done without this)

A green edit_scene_content result is not verification. After building:

1. take_screenshot → check layout, spacing, overlap, arrow routing, colour, icon shapes. (Not text — it won't render.)
2. search_scene_content → confirm every label exists with the right text, and that positions/alignment are what you intended (e.g. equal top/bottom margins for a centered block; shared y across a row).
3. Fix and re-verify until both pass. Typical: 2–4 iterations.

Centering, alignment & icons — the fiddly parts

The bundled script does the error-prone math and the icon extraction. Run it from the skill directory:

# centering: where to put an icon + standalone label as one centered block in a box
python scripts/excalidraw_tools.py place --box X Y W H --icon-h H --font F

# find an icon pack, list its items, extract one positioned icon as add-JSON
python scripts/excalidraw_tools.py catalog <keyword>
python scripts/excalidraw_tools.py list <author/name.excalidrawlib>
python scripts/excalidraw_tools.py icon <lib> --item <idx|name> --at <cx> <top_y> --height <h> [--swap OLD=NEW]

place and icon both take --at as center-x, top-y (not the box's top-left x,y) — feed icon --at the exact center_x/top_y that place printed.

Row/column alignment is just shared y+height (row) or x+width (column) with a constant step — keeping those exact is what separates intentional from hand-nudged. Details and the reasoning in reference/layout.md.

Reference map

Read the file that matches the task; don't load them all up front.

• reference/mcp-gotchas.md — the MCP quirks above, in full. Read before any real scene work — these cause the most wasted time.
• reference/elements.md — element fields, fonts, semantic color palette, arrow bindings, frames/slides, charts. Your working reference while building.
• reference/layout.md — planning, spacing, centering/alignment math, and which structure to choose for a given idea (with a decision table).
• reference/library-icons.md — finding and inserting real community icons via the MCP; identifying unnamed icons; recommended packs.
• scripts/excalidraw_tools.py — catalog / list / icon / place. Run --help for any subcommand.

Professional vs hand-drawn

Excalidraw defaults to a hand-drawn look (roughness:1, Excalifont). For clean, technical diagrams set roughness:0, use fontFamily:8 (mono) or 7 (titles), and lean on gray/blue with one accent. For approachable/whiteboard diagrams, keep the defaults. Match the register to the audience.