| name | comfyui-skill-public |
| description | Portable ComfyUI workflow and API guidance for any install. Use when building, validating, or troubleshooting ComfyUI image/video workflows, discovering available nodes/models via /object_info, wiring loaders/encoders/VAEs/LoRAs correctly, submitting jobs through the REST or WebSocket APIs, training LoRAs with ComfyUI, adapting a workflow to an unknown user machine without assuming specific checkpoints, paths, hardware, or custom nodes. |
ComfyUI Portable Skill
Use this skill when the task is to work with ComfyUI in a reusable, installation-agnostic way.
Minimum trigger scope
Trigger on requests like:
- "Build a ComfyUI workflow"
- "Fix this ComfyUI workflow"
- "Use the ComfyUI API"
- "Why is this ComfyUI graph failing?"
- "Add a LoRA / model / node to this ComfyUI setup"
- "Train a LoRA on this dataset"
- "Run a batch on these images"
- "Generate a video with LTX"
- "Edit these photos with FLUX"
- "Run a video batch"
- "JoyCaption these images and generate prompts"
First move: assume zero install knowledge
Before writing or editing any workflow:
- Read setup.md if the install is unknown.
- Collect the missing setup fields from the user or discover them from the running ComfyUI instance via
/object_info.
- Prefer discovery over assumptions — confirm checkpoints, VAEs, encoders, and LoRAs from the target install.
- Stop and report missing requirements clearly.
Read path by task
- Cold-start instructions for an AI agent -> AGENTS.md
- Setup / first run / portability questions -> setup.md
- Prompting guide index (all families) -> prompting-guides/README.md
- LTX 2.3 video or image-to-video generation -> prompting-guides/ltx-2.3-prompting-guide.md
- FLUX 2 image generation/edit prompting -> prompting-guides/flux-2-prompting-guide.md
- WAN 2.2 video prompting -> prompting-guides/wan-2.2-prompting-guide.md
- Batch jobs with state tracking, monitoring, and cron job setup -> batch-operations.md
- Batch cron jobs, notification templates, and error recovery SOP -> cron-jobs.md
- FLUX image edit / FLUX T2I / LTX / WAN / Hunyuan node maps and code patterns -> reference-implementations.md
- Production-ready ComfyUI API helper (persistent WS, reconnect, env config) -> scripts/api_lib.py
- Sanitized batch script starter -> scripts/template_run_batch.py
- JoyCaption-driven image-to-prompt pipeline convention -> docs/joycaption-convention.md
- Demo workflow JSON files (FLUX image edit, FLUX T2I, LTX video) -> demo-workflows/ — for demonstration only; replace with your own workflows
- External dependencies (JoyCaption, ComfyUI, Python packages) -> dependencies.md
- API submission / queue / history / WebSocket -> api.md
- Programmatic graph building, compatibility checks, and debugging -> workflow-patterns.md
- Model-family requirements and LoRA loading/training guidance -> models.md
- Prompt construction (general) -> prompting-guides/general-prompting-guide.md
- Release history -> changelog.md
Which guide to use (decision flow)
- Is this first-time setup or unknown install? -> open setup.md first.
- Is the requested model family known?
- Is it a video task?
- Is it FLUX image generation/edit? -> prompting-guides/flux-2-prompting-guide.md
- If still unclear, use prompting-guides/general-prompting-guide.md, then refine with the family guide.
Global operating rules
- Treat node classes as discoverable, not guaranteed constants.
- Treat model filenames as examples until confirmed on the target install.
- Use filename-only model references in workflow JSON unless the install explicitly requires a subdirectory path.
- Use defensive parsing for
/history and /object_info; schema details can vary by ComfyUI version and custom nodes.
- Fail fast when a required node class, model, or custom node is missing.
- Keep setup-specific notes out of this file; put them in a per-user config or setup reference.
- Never edit a base workflow JSON file directly — always deep copy and patch.
- Demo workflow JSON files in demo-workflows/ are provided as working examples. Replace them with your own workflows adapted to your ComfyUI install.
- Never hardcode a ComfyUI host, input directory, Discord channel, model filename, or workflow path into a committed script. Use the scripts/api_lib.py env-var conventions (
COMFYUI_HOST, COMFYUI_WS, COMFYUI_INPUT_DIR) or pass values through function arguments.
- Before generating prompts for an image set, check for an existing
joycaption.md in the input folder (see docs/joycaption-convention.md). If found, ask the user: use it, update it, or start fresh.
Cold-read test
Before publishing or packaging changes, check that a fresh reader could answer:
- What does this skill do?
- When should it trigger?
- What must be discovered first?
- Where do install-specific values go?
- Which reference file should be opened for the current task?