comfyui-workflow-builder

This skill translates natural-language image and video generation requests into valid, executable ComfyUI workflow JSON. It addresses a common failure mode when models try to hand-author ComfyUI graphs: hallucinated node names, wrong output indices, dangling connections, and model settings that do not match the chosen checkpoint. By encoding the correct node-graph format, connection conventions, and model-specific rules, it produces graphs that ComfyUI can actually run.

The documented workflow is a six-step process: parse the request into output type, source material, identity method, and quality level; check a cached inventory of available checkpoints, identity models, ControlNet models, custom nodes, and VRAM; select a pipeline pattern from a table covering text-to-image, identity-preserved image, LoRA character, image-to-video (Wan and AnimateDiff), talking head, upscale, and inpainting; generate the JSON; validate every class_type and model filename against the inventory plus connection and resolution checks; and output by either queuing through the companion comfyui-api skill (online) or saving to disk (offline). It ships concrete templates (for example a full FLUX text-to-image graph), a VRAM estimation table by component, and a list of common mistakes to avoid such as wrong output indices and CFG values that are too high for InstantID.

It targets generative-media practitioners, ComfyUI power users, and agent builders who want to go from a prompt to a runnable pipeline. It explicitly excludes ComfyUI installation, custom node development, Python scripting, model training, and hardware advice, keeping its scope to workflow construction. An eval harness (test cases, benchmark, trigger tests) accompanies the skill.