SlimeFX is designed around a simple bet: artists should direct, not wire. We are building Claude integration so a VFX artist can describe an effect in plain language and get a procedural graph to refine - with tool calling and visual-reference analysis keeping the artist in control.
Architecture · design
From prompt to editable graph
Claude never draws pixels. It edits the same node graph an artist would, through a small set of typed tool calls, so everything it produces stays inspectable and editable.
- Prompt"Stylized shield hit, cyan ripple, short sparks"
- Claude APITool use against the SlimeFX node library and parameter ranges
- JSON graphNodes, parameters and connections, validated against a schema
- SlimeFXThe C++20 core loads the graph and simulates it in the viewport
{
"effect": "shield_hit",
"nodes": [
{ "id": "ring", "type": "spline_sweep", "params": { "radius": 1.2, "segments": 64 } },
{ "id": "ripple", "type": "noise_texture", "params": { "scale": 4.0, "speed": 0.8 } },
{ "id": "sparks", "type": "emitter", "params": { "rate": 40, "lifetime": 0.35 } }
],
"edges": [
{ "from": "ripple", "to": "ring.displace" },
{ "from": "ring", "to": "output.mesh" },
{ "from": "sparks", "to": "output.particles" }
]
}
- Schema first. Every graph is checked against the node library before it reaches the viewport, so bad output is rejected, not rendered.
- Tool calls, not free text. Claude adds nodes, sets parameters and connects edges through explicit tools, and each step is undoable.
- Visual references. A reference image can be sent alongside the prompt to guide palette, shape and timing.
- Same graph, same export. What the assistant builds goes through the same engine export as hand-built effects.