diff --git a/optional-skills/creative/comfyui-skill-openclaw/SKILL.md b/optional-skills/creative/comfyui-skill-openclaw/SKILL.md
new file mode 100644
index 000000000000..0811f234bc88
--- /dev/null
+++ b/optional-skills/creative/comfyui-skill-openclaw/SKILL.md
@@ -0,0 +1,342 @@
+---
+name: comfyui-skill-openclaw
+description: "Generate images, video, and audio through ComfyUI workflows using the comfyui-skill CLI. Import workflows, manage dependencies, execute across multiple servers, track history, and optionally manage via Web UI. Use when the user asks to generate images, run ComfyUI workflows, or manage ComfyUI resources."
+version: 1.0.0
+requires: "ComfyUI server running locally or via Comfy Cloud; CLI auto-installed via uvx or pip"
+author: HuangYuChuh
+license: Apache-2.0
+platforms: [macos, linux, windows]
+metadata:
+  hermes:
+    tags: [comfyui, image-generation, stable-diffusion, flux, creative, generative-ai, ai-art, workflow]
+    related_skills: [stable-diffusion-image-generation, image_gen]
+    category: creative
+---
+
+# ComfyUI Agent Skill (OpenClaw)
+
+Generate images, video, and audio through ComfyUI using the `comfyui-skill` CLI.
+Workflows become callable "skills" with named parameters — the agent never touches raw node graphs.
+
+**Reference files:**
+- `references/cli-reference.md` — complete command reference (27 commands)
+- `references/api-notes.md` — underlying REST API (for debugging)
+- `scripts/comfyui_setup.sh` — workspace initialization
+
+## When to Use
+
+- User asks to generate images with Stable Diffusion, SDXL, Flux, or other diffusion models
+- User wants to run a specific ComfyUI workflow
+- User wants to chain generative steps (txt2img → upscale → face restore)
+- User needs ControlNet, inpainting, img2img, or other advanced pipelines
+- User asks to manage ComfyUI queue, check models, or install custom nodes
+- User wants video/audio generation via AnimateDiff, Hunyuan, AudioCraft, etc.
+- User wants to import, organize, or configure workflows visually (Web UI)
+
+## How It Works
+
+1. **Import** a workflow JSON (editor or API format) → CLI extracts a parameter schema
+2. **Run** with friendly args (`--args '{"prompt": "a cat"}'`) → CLI injects values into nodes
+3. **Retrieve** outputs → CLI downloads generated files locally
+
+The agent never sees node IDs or graph wiring. The CLI handles:
+- Editor → API format conversion (resolves reroutes, widget ordering via `/object_info`)
+- Auto-upload of local images referenced in args
+- Dependency checking (missing custom nodes, models)
+- WebSocket streaming with polling fallback
+- Multi-server routing (`server_id/workflow_id`)
+- Idempotent execution via `--job-id`
+
+## CLI Invocation
+
+**Zero-install (recommended):**
+
+```bash
+uvx --from comfyui-skill-cli comfyui-skill [OPTIONS] COMMAND [ARGS]
+```
+
+**If installed via pip/pipx:**
+
+```bash
+comfyui-skill [OPTIONS] COMMAND [ARGS]
+```
+
+For brevity, examples below use an alias:
+
+```bash
+COMFY="uvx --from comfyui-skill-cli comfyui-skill"
+# Or if pip-installed: COMFY="comfyui-skill"
+```
+
+**Always pass `--json` for structured output** the agent can parse:
+
+```bash
+$COMFY --json list
+$COMFY --json run my-workflow --args '{"prompt": "a cat"}'
+```
+
+## Setup & Onboarding
+
+### 1. ComfyUI Must Be Running
+
+The CLI talks to a running ComfyUI server. If the user doesn't have one:
+
+- Point them to https://docs.comfy.org/installation
+- Supports: NVIDIA (CUDA), AMD (ROCm), Intel Arc, Apple Silicon (MPS), CPU-only
+- Desktop app available for Windows/macOS; manual install for Linux
+- Comfy Cloud available for users without a GPU (https://platform.comfy.org)
+
+### 2. Initialize a Workspace
+
+The CLI reads `config.json` and `data/` from its working directory:
+
+```bash
+bash scripts/comfyui_setup.sh
+```
+
+Or manually:
+
+```bash
+mkdir -p ~/.hermes/comfyui && cd ~/.hermes/comfyui
+$COMFY --json server add --id local --url http://127.0.0.1:8188 --name "Local ComfyUI"
+```
+
+For Comfy Cloud:
+
+```bash
+$COMFY --json server add --id cloud --url https://cloud.comfy.org \
+  --name "Comfy Cloud" --api-key "comfyui-xxxxxxxxxxxx"
+```
+
+### 3. Verify Connection
+
+```bash
+$COMFY --json server status
+```
+
+Should return `"status": "online"`. If offline, user needs to start ComfyUI.
+
+### 4. Import a Workflow
+
+```bash
+$COMFY --json workflow import /path/to/workflow.json --name my-workflow
+```
+
+Auto-detects format (editor or API), converts if needed, extracts parameter schema.
+
+To import from the ComfyUI server's saved workflows:
+
+```bash
+$COMFY --json workflow import --from-server
+```
+
+## Core Workflow
+
+### Step 1: List Available Skills
+
+```bash
+$COMFY --json list
+```
+
+Returns all imported workflows with parameter schemas and `param_count`.
+- `required: true` → ask the user if not provided
+- `required: false` → infer from context or omit
+- Never expose node IDs; only use business parameter names
+
+### Step 2: Check Dependencies (First Run)
+
+```bash
+$COMFY --json deps check my-workflow
+```
+
+If `is_ready` is false:
+
+```bash
+$COMFY --json deps install my-workflow --all
+```
+
+Missing models must be downloaded manually — CLI reports which folder to place them in.
+
+### Step 3: Execute
+
+**Blocking (recommended for most use):**
+
+```bash
+$COMFY --json run my-workflow --args '{"prompt": "a beautiful sunset", "seed": 42}'
+```
+
+Blocks until done, streams progress, downloads outputs.
+
+**Non-blocking (for long jobs):**
+
+```bash
+# Submit
+$COMFY --json submit my-workflow --args '{"prompt": "..."}'
+# Returns: {"prompt_id": "abc-123"}
+
+# Poll — each call is a SEPARATE tool invocation, do NOT loop in shell
+$COMFY --json status abc-123
+# Returns: {"status": "running"} or {"status": "success", "outputs": [...]}
+```
+
+**Polling pattern (critical):** Each `status` call must be a separate bash command.
+Do NOT write a shell loop. Read the JSON, report progress to the user, then call again.
+
+### Step 4: Present Results
+
+On success, `outputs` contains file paths. Show them to the user via image preview or file reference.
+
+## Quick Decision Tree
+
+| User says | Command |
+|-----------|---------|
+| "generate an image" / "draw" | `run <skill> --args '{"prompt": "..."}'` |
+| "import this workflow" | `workflow import <path>` |
+| "use this image" (img2img) | `upload <image>` then `run` with the reference |
+| "inpaint this" | `upload <mask> --mask` then `run` |
+| "what workflows do I have" | `list` |
+| "what models are available" | `models list checkpoints` |
+| "check if everything's installed" | `deps check <skill>` |
+| "what failed" / "show history" | `history list <skill>` |
+| "cancel that" | `cancel <prompt_id>` |
+| "free up GPU memory" | `free` |
+| "which nodes exist for X" | `nodes search <query>` |
+| "manage workflows visually" | `python3 ./ui/open_ui.py` (Web UI) |
+
+## Web UI (Optional)
+
+The project ships a dedicated Web UI for visual workflow management:
+
+```bash
+python3 ./ui/open_ui.py
+```
+
+The Web UI provides:
+- Visual workflow import and parameter schema editing
+- Drag-and-drop workflow ordering
+- Multi-server configuration and health monitoring
+- One-click dependency checking and installation
+- Execution history browser
+- i18n support (English, 简体中文, 繁體中文, 日本語, 한국어, Español)
+
+The Web UI is a companion to the CLI, not a replacement. Agents use the CLI; humans use the Web UI for setup and configuration.
+
+**Source:** https://github.com/HuangYuChuh/ComfyUI_Skills_OpenClaw
+
+## Multi-Server
+
+Skills are addressed as `server_id/workflow_id`:
+
+```bash
+$COMFY --json list                              # all servers
+$COMFY --json run local/txt2img --args '{...}'  # specific server
+$COMFY --json run cloud/flux --args '{...}'     # different server
+$COMFY --json server stats --all                # VRAM/RAM across all servers
+```
+
+If `server_id` is omitted, the default server is used.
+
+## Image Upload (img2img / Inpainting)
+
+```bash
+# Upload input image
+$COMFY --json upload /path/to/photo.png
+
+# Upload mask for inpainting
+$COMFY --json upload /path/to/mask.png --mask --original photo.png
+
+# Auto-upload: if a param has type "image" and value starts with /, ./, ../, ~,
+# the CLI uploads it automatically
+$COMFY --json run inpaint --args '{"image": "./photo.png", "mask": "./mask.png", "prompt": "fill with flowers"}'
+```
+
+## Model Discovery
+
+```bash
+$COMFY --json models list                  # all folder types
+$COMFY --json models list checkpoints      # checkpoint files
+$COMFY --json models list loras            # LoRA files
+$COMFY --json models list controlnet       # ControlNet models
+```
+
+Folders: `checkpoints`, `loras`, `vae`, `controlnet`, `clip`, `clip_vision`,
+`upscale_models`, `embeddings`, `unet`, `diffusion_models`.
+
+## Node Discovery
+
+```bash
+$COMFY --json nodes list                   # all nodes, grouped by category
+$COMFY --json nodes list -c sampling       # filter by category
+$COMFY --json nodes info KSampler          # full details of one node
+$COMFY --json nodes search "upscale"       # fuzzy search
+```
+
+## Queue & System
+
+```bash
+$COMFY --json queue list                   # running + pending jobs
+$COMFY --json queue clear                  # clear pending
+$COMFY --json cancel <prompt_id>           # cancel specific job
+$COMFY --json free                         # unload models + free VRAM
+$COMFY --json server stats                 # system info (VRAM, RAM, GPU)
+```
+
+## Workflow Management
+
+```bash
+$COMFY --json workflow import <path> --name <id>    # import from file
+$COMFY --json workflow import --from-server          # import from ComfyUI server
+$COMFY --json workflow enable <skill_id>             # enable
+$COMFY --json workflow disable <skill_id>            # disable
+$COMFY --json workflow delete <skill_id>             # delete
+$COMFY --json info <skill_id>                        # show schema + details
+```
+
+## Idempotent Execution
+
+For retries that shouldn't burn extra GPU:
+
+```bash
+$COMFY --json run my-workflow --args '{"prompt": "..."}' --job-id "unique-key-123"
+```
+
+If `unique-key-123` was already executed, returns the cached result instantly (O(1) file check).
+
+## Pitfalls
+
+1. **Working directory matters** — The CLI reads `config.json` and `data/` from CWD.
+   Always `cd` to the workspace. If `list` returns empty, you're in the wrong directory.
+
+2. **Editor format needs a live server** — Importing editor-format workflows calls
+   `/object_info` to resolve widget ordering. API-format imports work offline.
+
+3. **Missing custom nodes** — Always `deps check` before first run. "class_type not found"
+   means missing nodes.
+
+4. **JSON args quoting** — Wrap `--args` in single quotes: `--args '{"prompt": "a cat"}'`.
+
+5. **Comfy Cloud differences** — Cloud uses `/api/` prefix and `X-API-Key` auth.
+   The CLI handles this transparently when configured with `--api-key`.
+
+6. **Model names are exact** — Case-sensitive, includes extension. Use
+   `models list checkpoints` to discover installed models.
+
+7. **Long generations** — Video and high-step workflows can take minutes. Use `run`
+   for blocking or `submit` + `status` for non-blocking.
+
+8. **Concurrent limits (Cloud)** — Free/Standard: 1 job. Creator: 3. Pro: 5.
+
+9. **Config portability** — Use `config export` / `config import` to transfer setups.
+
+10. **Cloud API nodes unauthorized** — Workflows using Kling, Sora, or other paid API nodes
+    need a Comfy Cloud API Key. Generate one at https://platform.comfy.org and configure
+    via `server add --api-key` or the Web UI's server settings.
+
+## Verification Checklist
+
+- [ ] `uvx --from comfyui-skill-cli comfyui-skill --version` runs successfully
+- [ ] `server status` returns online
+- [ ] Workspace dir has `config.json` and `data/`
+- [ ] At least one workflow imported (`list` returns non-empty)
+- [ ] `deps check` passes for imported workflows
+- [ ] Test run completes and outputs are saved
diff --git a/optional-skills/creative/comfyui-skill-openclaw/references/api-notes.md b/optional-skills/creative/comfyui-skill-openclaw/references/api-notes.md
new file mode 100644
index 000000000000..5bcd80668e55
--- /dev/null
+++ b/optional-skills/creative/comfyui-skill-openclaw/references/api-notes.md
@@ -0,0 +1,104 @@
+# ComfyUI REST API Notes
+
+The `comfyui-skill` CLI wraps these endpoints. This reference is for debugging,
+understanding errors, or advanced use when the CLI doesn't cover a specific need.
+
+## Endpoints the CLI Uses
+
+| Endpoint | Method | CLI Command |
+|----------|--------|-------------|
+| `/system_stats` | GET | `server status`, `server stats` |
+| `/prompt` | POST | `run`, `submit` |
+| `/history/{prompt_id}` | GET | `status`, `run` (polling) |
+| `/history` | GET | `history list --server` |
+| `/queue` | GET | `queue list` |
+| `/queue` | POST | `queue clear`, `queue delete` |
+| `/interrupt` | POST | `cancel` |
+| `/free` | POST | `free` |
+| `/object_info` | GET | `nodes list`, `workflow import` (schema extraction) |
+| `/object_info/{class}` | GET | `nodes info` |
+| `/models` | GET | `models list` |
+| `/models/{folder}` | GET | `models list <folder>`, `deps check` |
+| `/view` | GET | `run` (output download) |
+| `/upload/image` | POST | `upload` |
+| `/upload/mask` | POST | `upload --mask` |
+| `/node_replacements` | GET | `workflow import` (deprecated node detection) |
+| `/internal/logs/raw` | GET | `logs show` |
+| `/workflow_templates` | GET | `templates list` |
+| `/global_subgraphs` | GET | `templates subgraphs` |
+| `/v2/userdata` | GET | `workflow import --from-server` |
+| `/ws` | WebSocket | `run` (real-time progress) |
+
+### Cloud-specific
+
+| Endpoint | Method | Purpose |
+|----------|--------|---------|
+| `/api/jobs` | GET | Job listing with filtering |
+| `/api/jobs/{id}` | GET | Job details |
+
+### ComfyUI Manager (optional plugin)
+
+| Endpoint | Method | CLI Command |
+|----------|--------|-------------|
+| `/manager/queue/start` | GET | `deps install` |
+| `/manager/queue/install` | POST | `deps install` (custom nodes) |
+| `/manager/queue/install_model` | POST | `deps install --models` |
+| `/manager/queue/status` | GET | `deps install` (progress) |
+
+## Local vs Cloud Differences
+
+| | Local | Cloud |
+|---|---|---|
+| Base URL | `http://127.0.0.1:8188` | `https://cloud.comfy.org` |
+| Route prefix | none | `/api` |
+| Auth | none or bearer token | `X-API-Key` header |
+| Job status | Poll `/history/{id}` | `/api/jobs/{id}` |
+| Output download | Direct bytes from `/view` | 302 redirect → signed URL |
+| WebSocket | `ws://host:port/ws?clientId={uuid}` | `wss://host/ws?clientId={uuid}&token={key}` |
+| Concurrent jobs | Sequential | Tier-limited (Free: 1, Creator: 3, Pro: 5) |
+
+The CLI handles all of these differences transparently based on the server config.
+
+## Workflow JSON Format (API Format)
+
+```json
+{
+  "node_id_string": {
+    "class_type": "NodeClassName",
+    "inputs": {
+      "param_name": "value",
+      "linked_input": ["source_node_id", output_index]
+    }
+  }
+}
+```
+
+- Node IDs are strings (`"3"`, not `3`)
+- Links: `["node_id", output_index]` — 0-based int
+- `class_type` must match exactly (case-sensitive)
+
+## POST /prompt Payload
+
+```json
+{
+  "prompt": { "<workflow>" },
+  "client_id": "uuid",
+  "extra_data": {
+    "api_key_comfy_org": "key-for-paid-api-nodes"
+  }
+}
+```
+
+The CLI constructs this from the imported workflow + injected parameters.
+
+## WebSocket Message Types
+
+| Type | When | Key Fields |
+|------|------|------------|
+| `execution_start` | Prompt begins | `prompt_id` |
+| `executing` | Node running (`null` = done) | `node`, `prompt_id` |
+| `progress` | Sampling steps | `node`, `value`, `max` |
+| `executed` | Node output ready | `node`, `output` |
+| `execution_success` | All nodes done | `prompt_id` |
+| `execution_error` | Failure | `exception_type`, `exception_message` |
+| `execution_interrupted` | User cancelled | `prompt_id` |
diff --git a/optional-skills/creative/comfyui-skill-openclaw/references/cli-reference.md b/optional-skills/creative/comfyui-skill-openclaw/references/cli-reference.md
new file mode 100644
index 000000000000..016da6e3cb23
--- /dev/null
+++ b/optional-skills/creative/comfyui-skill-openclaw/references/cli-reference.md
@@ -0,0 +1,172 @@
+# comfyui-skill CLI Reference
+
+Complete command map for `comfyui-skill` v0.2.x.
+
+**Invocation:** `uvx --from comfyui-skill-cli comfyui-skill [OPTIONS] COMMAND [ARGS]`
+
+Or if installed as a tool: `comfyui-skill [OPTIONS] COMMAND [ARGS]`
+
+## Global Options
+
+| Option | Short | Description |
+|--------|-------|-------------|
+| `--version` | `-V` | Show version |
+| `--json` | `-j` | JSON output (always use for agent parsing) |
+| `--output-format` | | `text`, `json`, or `stream-json` (NDJSON events) |
+| `--server` | `-s` | Server ID override |
+| `--dir` | `-d` | Data directory (default: CWD) |
+| `--verbose` | `-v` | Verbose output |
+| `--no-update-check` | | Skip CLI update check |
+
+## Standalone Commands
+
+### `list`
+List all available skills across all enabled servers. Returns `param_count` per workflow.
+
+### `info <SKILL_ID>`
+Show skill details and parameter schema. Skill ID: `server_id/workflow_id` or `workflow_id`.
+
+### `run <SKILL_ID> [OPTIONS]`
+Execute a skill (blocking — waits for completion, streams progress).
+
+| Option | Short | Description |
+|--------|-------|-------------|
+| `--args` | `-a` | JSON parameters (default: `{}`) |
+| `--only` | | Comma-separated node IDs for partial execution |
+| `--priority` | `-p` | Queue priority (lower = first, negative = jump queue) |
+| `--validate` | | Validate workflow without executing (dry run) |
+| `--job-id` | | Idempotency key — reuse cached result if already executed |
+
+### `submit <SKILL_ID> [OPTIONS]`
+Submit a skill (non-blocking — returns `prompt_id` immediately). Same options as `run` except no streaming.
+
+### `status <PROMPT_ID>`
+Check execution status. Returns: `queued` (with `position`), `running`, `success` (with `outputs`), or `error`.
+
+### `upload [FILE_PATH] [OPTIONS]`
+Upload a file to ComfyUI for use in workflows.
+
+| Option | Description |
+|--------|-------------|
+| `--from-output` | Reuse output from a previous prompt_id as input |
+| `--mask` | Upload as mask (for inpainting) |
+| `--original` | Original image filename (for mask upload) |
+
+### `cancel <PROMPT_ID>`
+Cancel a running or queued job. Interrupts if running, removes from queue if pending.
+
+### `free [OPTIONS]`
+Release GPU memory.
+
+| Option | Short | Description |
+|--------|-------|-------------|
+| `--models` | `-m` | Unload all models from VRAM |
+| `--memory` | | Free all cached memory |
+
+## Command Groups
+
+### `server` — Manage ComfyUI Servers
+
+| Subcommand | Description |
+|------------|-------------|
+| `server list` | List all configured servers |
+| `server status [SERVER_ID]` | Check if server is online |
+| `server stats [SERVER_ID]` | System stats: VRAM, RAM, GPU, versions (`--all` for all servers) |
+| `server add` | Add server (`--id`, `--url` required; `--name`, `--output-dir`, `--auth`, `--api-key` optional) |
+| `server enable <SERVER_ID>` | Enable a server |
+| `server disable <SERVER_ID>` | Disable a server |
+| `server remove <SERVER_ID>` | Remove a server from config (does not delete workflow data) |
+
+### `workflow` — Manage Workflows
+
+| Subcommand | Description |
+|------------|-------------|
+| `workflow import [JSON_PATH]` | Import workflow (`--name`, `--from-server`, `--preview`, `--check-deps`) |
+| `workflow enable <SKILL_ID>` | Enable a workflow |
+| `workflow disable <SKILL_ID>` | Disable a workflow |
+| `workflow delete <SKILL_ID>` | Delete a workflow |
+
+### `models` — Discover Models
+
+| Subcommand | Description |
+|------------|-------------|
+| `models list [FOLDER]` | List models in a folder (checkpoints, loras, vae, controlnet, etc.) |
+
+### `nodes` — Discover Nodes
+
+| Subcommand | Description |
+|------------|-------------|
+| `nodes list` | List all node classes (`-c` to filter by category) |
+| `nodes info <NODE_CLASS>` | Full details of a node type |
+| `nodes search <QUERY>` | Fuzzy search across names/categories |
+
+### `deps` — Dependency Management
+
+| Subcommand | Description |
+|------------|-------------|
+| `deps check <SKILL_ID>` | Check if dependencies are installed (returns `is_ready`) |
+| `deps install <SKILL_ID>` | Install missing deps (`--repos` git URLs, `--models`, `--all`) |
+
+### `history` — Execution History
+
+| Subcommand | Description |
+|------------|-------------|
+| `history list [SKILL_ID]` | List history (`--server`, `--status`, `--limit`, `--sort`) |
+| `history show <SKILL_ID> <RUN_ID>` | Show specific run details |
+
+### `queue` — Queue Management
+
+| Subcommand | Description |
+|------------|-------------|
+| `queue list` | Show running and pending jobs |
+| `queue clear` | Clear all pending jobs |
+| `queue delete <PROMPT_IDS...>` | Remove specific jobs from queue |
+
+### `logs` — Server Logs
+
+| Subcommand | Description |
+|------------|-------------|
+| `logs show` | Show recent server logs (`--lines` / `-n`, default: 50) |
+
+### `templates` — Discover Templates
+
+| Subcommand | Description |
+|------------|-------------|
+| `templates list` | Workflow templates from custom nodes |
+| `templates subgraphs` | Reusable subgraph components |
+
+### `config` — Configuration
+
+| Subcommand | Description |
+|------------|-------------|
+| `config export` | Export config + workflows as bundle (`--output`, `--portable-only`) |
+| `config import <INPUT_PATH>` | Import bundle (`--dry-run`, `--apply-environment`, `--no-overwrite`) |
+
+## Config File Format
+
+Located at `<workspace>/config.json`:
+
+```json
+{
+  "default_server": "local",
+  "servers": [
+    {
+      "id": "local",
+      "name": "Local ComfyUI",
+      "url": "http://127.0.0.1:8188",
+      "enabled": true,
+      "output_dir": "./outputs",
+      "auth": "",
+      "comfy_api_key": ""
+    }
+  ]
+}
+```
+
+**Server fields:**
+- `id` — unique identifier (no spaces/slashes/dots)
+- `url` — ComfyUI base URL
+- `enabled` — whether server is active
+- `output_dir` — where outputs are saved (relative to workspace)
+- `auth` — bearer token for authenticated servers
+- `comfy_api_key` — Comfy Cloud API key
diff --git a/optional-skills/creative/comfyui-skill-openclaw/scripts/comfyui_setup.sh b/optional-skills/creative/comfyui-skill-openclaw/scripts/comfyui_setup.sh
new file mode 100644
index 000000000000..8e3017d11262
--- /dev/null
+++ b/optional-skills/creative/comfyui-skill-openclaw/scripts/comfyui_setup.sh
@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+# Initialize a comfyui-skill workspace directory.
+# Usage: bash scripts/comfyui_setup.sh [WORKSPACE_DIR] [--url COMFYUI_URL]
+#
+# Creates the workspace, adds a default local server config,
+# and verifies the connection.
+
+set -euo pipefail
+
+WORKSPACE="${1:-$HOME/.hermes/comfyui}"
+COMFYUI_URL="http://127.0.0.1:8188"
+
+# Parse optional --url flag
+for arg in "$@"; do
+    case "$prev" in
+        --url) COMFYUI_URL="$arg"; prev="";;
+        *) prev="$arg";;
+    esac
+done
+
+# Detect CLI: prefer uvx, fall back to direct command
+if command -v uvx &>/dev/null; then
+    COMFY="uvx --from comfyui-skill-cli comfyui-skill"
+elif command -v comfyui-skill &>/dev/null; then
+    COMFY="comfyui-skill"
+else
+    echo "ERROR: Neither uvx nor comfyui-skill found."
+    echo "Install one of:"
+    echo "  pip install uv          # then uvx handles everything"
+    echo "  pip install comfyui-skill-cli"
+    exit 1
+fi
+
+echo "==> Initializing ComfyUI skill workspace at: $WORKSPACE"
+mkdir -p "$WORKSPACE"
+cd "$WORKSPACE"
+
+# Create config if missing
+if [ ! -f config.json ]; then
+    echo "==> Creating default config (server at $COMFYUI_URL)"
+    $COMFY --json server add --id local --url "$COMFYUI_URL" --name "Local ComfyUI"
+    echo "==> Config created: $WORKSPACE/config.json"
+else
+    echo "==> config.json already exists, skipping"
+fi
+
+# Verify connection
+echo "==> Checking server connection..."
+if $COMFY --json server status 2>/dev/null | grep -q '"online"'; then
+    echo "==> ComfyUI is reachable!"
+    $COMFY --json server stats 2>/dev/null || true
+else
+    echo "==> ComfyUI is not reachable at $COMFYUI_URL"
+    echo "    Start ComfyUI first, or re-run with a different URL:"
+    echo "    bash scripts/comfyui_setup.sh $WORKSPACE --url http://YOUR_HOST:PORT"
+    echo ""
+    echo "    Install ComfyUI: https://docs.comfy.org/installation"
+fi
+
+echo ""
+echo "==> Workspace ready: $WORKSPACE"
+echo "    Always cd here before running comfyui-skill commands."