Excalidraw render skill

Requires: npx (Node) and network access on the first run, which fetches @swiftlysingh/excalidraw-cli. Also needs the ability to spawn a sub-agent because the cost model below depends on it. If sub-agents are unavailable in this harness, say so and offer to author the JSON directly at higher token cost rather than silently doing it.

Make a polished Excalidraw diagram from a rough input, without baby-sitting. The main agent thinks in pixels (by looking at rendered PNGs); a Haiku sub-agent does the JSON work.

Scope: draw what the user described, at the level of detail they gave. Don't add components they didn't mention because the architecture "should" have a cache or a load balancer, don't expand a three-box sketch into a full system map, and don't redesign their architecture while diagramming it. If something in the input is genuinely ambiguous, draw your best reading and say what you assumed.

Core principle: cost discipline

A .excalidraw file is huge: a 10-box diagram is 3K to 5K tokens. Reading or editing it with the main model is wasteful and slow. Vision on a rendered PNG is far cheaper and far more accurate for layout judgment.

The main agent MUST:

• ✅ Read the rendered PNG to judge correctness (this is what vision is for)
• ✅ Plan layouts up-front (positions, sizes, colors) as an architectural judgment
• ✅ Critique specific visual issues from the PNG (overlaps, clipping, misalignment)
• ✅ Dispatch surgical fix instructions to a Haiku sub-agent