ComfyUI-Expert

VideoAgent - ComfyUI Video Production Orchestrator

A session-scoped AI orchestrator for Claude Code that turns Claude into a senior video production technical director. It routes natural-language requests to 13 specialized skill modules covering the full pipeline: character image generation, video production, voice synthesis, LoRA training, and publishing – all driven by ComfyUI.

Why This Exists

Producing AI-generated video with ComfyUI involves juggling dozens of models, custom nodes, prompt styles, and hardware constraints. VideoAgent wraps all of that domain knowledge into a structured skill system that Claude reads on demand, so you can say things like:

• “Generate a photorealistic portrait of a dog using InstantID”
• “Make a talking head video where she says ‘Welcome to my channel’”
• “Train a LoRA from these 20 reference images”
• “Check for new ComfyUI video models released this month”

…and get validated, hardware-aware ComfyUI workflows without memorizing node names or VRAM budgets.

How It Works

video-agent.bat
  |
  |-- Writes state/session.json (active project, ComfyUI URL)
  |-- cd to this repo
  |-- Launches: claude
        |
        |-- Claude Code auto-loads CLAUDE.md (the orchestrator)
        |-- Local hooks fire (staleness check)
        |-- User's first message triggers foundation reads
        |-- Each request is routed to the right skill file

Key design decisions:

Decision	Rationale
Session-scoped, not global	Skills stay in this repo. Other Claude Code sessions are unaffected.
CLAUDE.md is the orchestrator	The only file Claude auto-loads from a project root. Contains the routing table and behavioral instructions.
Skills are read-on-demand markdown	No build step, no registration. Claude reads skills/{name}/SKILL.md when the routing table says to.
REST polling over WebSocket	Claude Code can’t hold persistent connections. Polling every 5s works fine for minute-long video generation.
Research is user-triggered	No cron jobs. A session-start hook reminds you when data is stale.

Prerequisites

• Claude Code installed and authenticated
• ComfyUI installed (local or remote)
• FFmpeg on PATH (for video assembly)
• PowerShell 7+ (for utility scripts)
• Windows (the launcher is a .bat file; WSL/Linux adaptation is straightforward)

Quick Start

1. Clone and launch

git clone https://github.com/MCKRUZ/ComfyUI-Expert.git
cd ComfyUI-Expert
video-agent.bat

With options:

video-agent.bat --project "my-video"          # Set active project
video-agent.bat --comfyui "http://<remote-ip>:8188"  # Remote ComfyUI
video-agent.bat --resume                       # Resume last session

2. Scan your ComfyUI installation

First time (or after installing new models/nodes), tell the agent:

Scan my ComfyUI installation at C:\ComfyUI

Or run the script directly:

pwsh -File scripts/scan-inventory.ps1 -ComfyUIPath "C:\ComfyUI"

This creates state/inventory.json – a cache of every model, custom node, and VRAM detail. The agent validates every workflow against this inventory before execution.

3. Start creating

Generate a photorealistic portrait using FLUX
Create a new project called "Character Showcase"
Add a character named Spot - German Shepard, shaggy hair, dark fur
Train a LoRA from these reference images
Research the latest ComfyUI video models

Architecture

3-Tier Context System

VideoAgent loads context incrementally to stay within Claude’s context window:

Tier	Files	Loaded When	Size
1: Foundation	foundation/*.md	Session start (first interaction)	~2K tokens
2: Working	projects/{name}/*	When working on a specific project	Varies
3: Reference	references/*.md	Only when a skill explicitly needs detail	Large

Skill Dependency Graph

CLAUDE.md (orchestrator - always loaded)
    |
    |-- Discovery (no dependencies)
    |   |-- comfyui-prompt-interview Vision clarification via conversation
    |
    |-- Foundation Skills (no dependencies)
    |   |-- comfyui-api              REST API connection
    |   |-- comfyui-inventory        Model/node discovery
    |   |-- project-manager          Project & character state
    |
    |-- Research (independent)
    |   |-- comfyui-research         Self-updating knowledge base
    |
    |-- Core Creation (depend on inventory)
    |   |-- comfyui-prompt-engineer  Model-specific prompt optimization
    |   |-- comfyui-workflow-builder Validated workflow JSON generation
    |
    |-- Production (depend on creation)
    |   |-- comfyui-video-pipeline   Wan 2.6 / LTX-2.3 / FramePack / AnimateDiff
    |   |-- comfyui-voice-pipeline   Qwen3-TTS / Chatterbox / F5-TTS / lip-sync
    |   |-- comfyui-lora-training    Dataset prep, training, evaluation
    |
    |-- Output (depend on production)
    |   |-- video-assembly           FFmpeg + Remotion composition
    |   |-- video-publisher          YouTube metadata & upload
    |
    |-- Support
        |-- comfyui-troubleshooter   Error diagnosis & fixes

Request Routing

When you make a request, CLAUDE.md routes it to the right skill:

You Say	Skill Loaded	What Happens
“I have an idea for…” / vague concept	comfyui-prompt-interview	Guided conversation to clarify vision before generation
“Generate a character portrait”	comfyui-workflow-builder	Checks inventory, builds workflow JSON, queues via API
“Craft a better prompt”	comfyui-prompt-engineer	Model-specific optimization (FLUX vs SDXL vs Wan)
“Create a video from this image”	comfyui-video-pipeline	Selects engine (Wan 2.6/FramePack/LTX-2.3), builds pipeline
“Clone this voice / make her talk”	comfyui-voice-pipeline	Voice synthesis + lip-sync pipeline
“Train a LoRA”	comfyui-lora-training	Dataset prep, training config, checkpoint evaluation
“Build a raw workflow”	comfyui-workflow-builder	Direct workflow construction with inventory validation
“Check for new models”	comfyui-research	Scans YouTube/GitHub/HuggingFace, updates references
“Something broke”	comfyui-troubleshooter	Error pattern matching, fix suggestions
“Assemble the final video”	video-assembly	FFmpeg or Remotion-based composition
“Upload to YouTube”	video-publisher	Metadata generation + upload delegation
“Create a new project”	project-manager	Project manifests, character profiles
“Connect to ComfyUI”	comfyui-api	Connection test, system info

The 13 Skills

Discovery

comfyui-prompt-interview – Guides you through a structured conversation before generating anything. Asks about subject, mood, style, setting, and intended use to produce a complete creative brief. Particularly useful for vague or high-level ideas (“I want something dramatic and cinematic”). Outputs a prompt-ready specification that feeds directly into comfyui-prompt-engineer or comfyui-workflow-builder.

Foundation

comfyui-api – Connects to ComfyUI’s REST API (default http://127.0.0.1:8188). Queues workflows, polls for results at 5-second intervals, handles image/model uploads, cancellations, and VRAM management. Supports online mode (live API) and offline mode (JSON export).

comfyui-inventory – Discovers every installed model, custom node, and VRAM configuration. Works online (API queries) or offline (directory scanning via scan-inventory.ps1). Caches results to state/inventory.json. Maps node classes to packages (e.g., ApplyInstantID -> ComfyUI_InstantID).

project-manager – Creates project structures with YAML manifests and character profiles. Tracks generation history (what settings worked), manages character identity (appearance, voice, LoRA, reference images), and updates defaults after successful runs.

Research

comfyui-research – Monitors 7 YouTube channels, 11 GitHub repos, and HuggingFace trending models. Extracts knowledge from tutorials (via transcript analysis), tracks releases, and generates staleness reports. Models older than 90 days and nodes older than 60 days get flagged.

Core Creation

comfyui-prompt-engineer – Model-specific prompt optimization for FLUX, SDXL, SD1.5, and Wan. Adjusts prompts for identity methods (InstantID, PuLID, IP-Adapter, LoRA), recommends CFG scales per model, and provides negative prompt templates. Integrates with character profiles for context.

comfyui-workflow-builder – Generates ComfyUI workflow JSON from natural language. Validates every model and node against inventory before output. Supports text-to-image, identity-preserved generation, video (Wan/AnimateDiff), upscaling, and inpainting patterns. Includes VRAM estimation per component.

Production

comfyui-video-pipeline – Orchestrates video engines based on requirements:

• LTX-2.3: 4K production quality, audio+video, GGUF-quantized options
• Wan 2.6: 1080p reference-to-video, native audio generation, built-in lip-sync
• Wan 2.2 MoE 14B: Film-level quality, first+last frame control, 24GB+ VRAM
• FramePack: Long videos (60+ sec), VRAM-invariant (works on 6GB)
• HunyuanVideo 1.5: Lightweight quality alternative at 8.3B params
• AnimateDiff V3: Fast iteration, motion LoRAs, 4-8 step Lightning

Includes post-processing (RIFE frame interpolation, face enhancement, deflicker, color correction) and a dedicated talking-head pipeline.

comfyui-voice-pipeline – Seven voice synthesis tools (Qwen3-TTS, Chatterbox Turbo, F5-TTS, TTS Audio Suite, IndexTTS-2, RVC, ElevenLabs) and four lip-sync methods (Wav2Lip, SadTalker, LivePortrait, LatentSync 1.6). Three complete pipelines: Quick (image-to-talk), Quality (image-to-video-to-lip-sync), and Premium (expression transfer). Wan 2.6 now also supports native lip-sync as a fifth path.

comfyui-lora-training – Training tools (AI-Toolkit for FLUX, Kohya_ss for SDXL, FluxGym/SimpleTuner for low VRAM). Covers dataset preparation (15-30 images, captioning strategy), hyperparameter guidance, checkpoint evaluation, and LoRA + zero-shot method combination.

Output

video-assembly – Two modes: FFmpeg (concatenation, audio mixing, subtitles, transitions) and Remotion (animated captions, motion graphics, React-based templates). Audio normalization to -16 LUFS for YouTube. Quality presets (CRF 15-28).

video-publisher – Thin orchestrator that delegates to global YouTube skills for research, title/thumbnail optimization, upload, and analytics. Generates platform-specific metadata (YouTube, Shorts, Instagram Reels, TikTok).

Support

comfyui-troubleshooter – Diagnoses four error categories (server, workflow, quality, performance). Covers the top 10 common errors (OOM, missing nodes, precision mismatch, burned faces, etc.) with quick fixes. Includes a quality decision tree and missing-dependency resolution.

Model Landscape

The agent tracks the top models across five categories:

Image Generation

Model	Best For	VRAM
FLUX.2 [dev]	Photorealism, 4MP, multi-reference (up to 10 images)	24GB+
FLUX.2 [klein]	Fast generation, low VRAM (4B/9B distilled)	12-20GB+
FLUX Kontext	Iterative character editing	12-32GB
Qwen-Image 2.0	Typography, 2K resolution, layered editing	24GB+
Z-Image	Non-distilled quality (Base) / fast (Turbo, 8 steps)	12-16GB+

Identity Preservation

Method	Best For	VRAM
InfiniteYou	Highest identity fidelity	24GB
FLUX Kontext	Edit without retraining	12-32GB
PuLID Flux 2	FLUX.2 family (Klein + Dev)	24-40GB
PuLID Flux II	FLUX.1 dual characters, no pollution	24-40GB

Video Generation

Model	Best For	VRAM
LTX-2.3	4K audio+video, portrait mode, production	24GB+
Wan 2.6	Reference-to-video, lip-sync, native audio gen, 1080p	24GB+
Wan 2.2 MoE	Film-level quality, first+last frame control	24GB+
HunyuanVideo 1.5	Lightweight flagship quality (8.3B params)	24GB
FramePack	Long videos (60s+), VRAM-invariant	6GB+
SkyReels V1	Human-centric, 33 expressions, cinematic	24GB+
AnimateDiff V3	Fast iteration, motion LoRAs	8GB+

Voice / TTS

Tool	Best For	License
TTS Audio Suite	Unified 11-engine platform, 23 languages	Multi
Qwen3-TTS	10 languages, zero-shot clone, voice design	Open
Chatterbox Turbo	Emotion tags, sub-200ms latency	MIT
IndexTTS-2	8-emotion vector control via sliders	Open
F5-TTS	Zero-shot cloning, works on 6GB VRAM	MIT

Lip-Sync

Tool	Best For
LatentSync 1.6	Highest accuracy (ByteDance)
Wan 2.6 native	Reference-to-video with built-in lip-sync
Wav2Lip	Proven, works with any face
SadTalker	Head movement + expressions

Full specs and download links are in references/models.md.

Hardware Profile

VideoAgent is configured for an RTX 5090 (32GB VRAM) but works with any GPU. The agent adjusts recommendations based on available VRAM.

Workload	32GB Status	Notes
FLUX.1-dev FP16	Native	No quantization needed
Wan 2.2 14B	Native	Full quality
FramePack	Overkill	Designed for 6GB
PuLID Flux II	Native	Dual-character generation
InfiniteYou	Native	Both SIM and AES variants
LoRA Training (FLUX)	Native	No quantization needed

Recommended ComfyUI launch flags: --highvram --fp8_e4m3fn-unet

Project Structure

ComfyUI-Expert/
|-- video-agent.bat              Launcher (writes session config, opens Claude)
|-- CLAUDE.md                    Orchestrator (routing table, behavior, rules)
|-- .claude/
|   +-- settings.local.json     Project-local hooks & permissions
|
|-- foundation/                  Tier 1: Quick reference (~2K tokens)
|   |-- agent-persona.md         Communication style & principles
|   |-- api-quick-ref.md         ComfyUI REST API cheat sheet
|   |-- hardware-profile.md      GPU specs, VRAM capabilities
|   |-- model-landscape.md       Top 3 models per category
|   +-- skill-registry.md        Skill list & dependency map
|
|-- skills/                      13 skill modules (read on demand)
|   |-- comfyui-api/
|   |-- comfyui-inventory/
|   |-- comfyui-lora-training/
|   |-- comfyui-prompt-engineer/
|   |-- comfyui-prompt-interview/
|   |-- comfyui-research/
|   |-- comfyui-troubleshooter/
|   |-- comfyui-video-pipeline/
|   |-- comfyui-voice-pipeline/
|   |-- comfyui-workflow-builder/
|   |-- project-manager/
|   |-- video-assembly/
|   +-- video-publisher/
|
|-- references/                  Tier 3: Deep reference (loaded on demand)
|   |-- models.md                Full model catalog & download links
|   |-- workflows.md             Complete workflow node configurations
|   |-- lora-training.md         Training parameters & best practices
|   |-- voice-synthesis.md       Voice tools in depth
|   |-- prompt-templates.md      Model-specific prompt strategies
|   |-- troubleshooting.md       Error database with solutions
|   |-- research-log.md          Full technique survey (ongoing, replaces research-2025.md)
|   |-- staleness-report.md      Freshness tracking for all entries
|   +-- evolution.md             Update protocol & changelog
|
|-- projects/                    Per-project state (gitignored)
|-- state/                       Runtime state (gitignored)
|   |-- session.json             Active project & ComfyUI URL
|   +-- inventory.json           Cached models/nodes from scan
|
|-- scripts/                     Utility scripts
|   |-- scan-inventory.ps1       Offline ComfyUI directory scanner
|   |-- connect-comfyui.ps1      Connection test & diagnostics
|   |-- staleness-check.ps1      Session-start hook (checks research age)
|   |-- deploy.ps1               Sync references to global skill
|   +-- sync-skills.ps1          Sync all skills to global Claude skills directory
|
|-- agent/
|   +-- AGENT.md                 Extended orchestration spec
|
|-- openclaw/                    OpenClaw compatibility layer
|   |-- AGENTS.md                Orchestration rules (CLAUDE.md equivalent)
|   |-- SOUL.md                  Agent persona
|   |-- TOOLS.md                 Available tools & API reference
|   |-- setup.ps1                Install skills into OpenClaw workspace
|   +-- openclaw.example.json    Config template
|
+-- docs/
    |-- architecture.md          System design decisions
    +-- getting-started.md       Quick start guide

Example Workflows

Character Image Generation

1. "Create a new project called Character Showcase"
2. "Add a character named Sage - auburn hair, green eyes, freckles"
3. "Generate a photorealistic portrait of Sage using InstantID"

The agent: reads inventory -> loads workflow-builder skill -> loads prompt-engineer skill -> generates optimized prompt -> builds validated workflow JSON -> queues via ComfyUI API -> polls for result.

Talking Head Video

1. "Make Sage say 'Hello everyone, welcome to my channel'"

The agent orchestrates a multi-step pipeline: voice synthesis (Chatterbox/F5-TTS) -> video generation (Wan 2.2) -> lip-sync (LatentSync) -> face enhancement -> assembly.

Research Update

1. "Check for new ComfyUI models and techniques"

The agent: loads research skill -> checks YouTube channels, GitHub repos, HuggingFace trending -> extracts knowledge from tutorials -> updates reference files -> generates staleness report.

Hooks & Automation

VideoAgent uses Claude Code’s hooks system for lightweight automation:

Hook	Event	What It Does
Staleness check	SessionStart	Warns if research data is older than 2 weeks

Configured in .claude/settings.local.json (project-local, doesn’t affect other sessions).

Scripts Reference

Script	Purpose	Usage
scan-inventory.ps1	Scan ComfyUI models & nodes offline	pwsh -File scripts/scan-inventory.ps1 -ComfyUIPath "C:\ComfyUI"
connect-comfyui.ps1	Test ComfyUI connection & show diagnostics	pwsh -File scripts/connect-comfyui.ps1
staleness-check.ps1	Check research freshness (session hook)	Runs automatically at session start
deploy.ps1	Sync references to global comfyui-character-gen skill	pwsh -File scripts/deploy.ps1
sync-skills.ps1	Sync all skill files to global Claude skills directory	pwsh -File scripts/sync-skills.ps1

Customization

Different GPU

Edit foundation/hardware-profile.md with your GPU specs. The agent reads this at session start and adjusts VRAM recommendations accordingly.

Different ComfyUI location

Pass it at launch:

video-agent.bat --comfyui "http://<remote-ip>:8188"

Or edit the default in video-agent.bat (line 18).

Adding models to the landscape

Edit foundation/model-landscape.md (top 3 quick reference) and references/models.md (full catalog). Or just ask the agent to run a research update.

Adapting for Linux/macOS

Replace video-agent.bat with a shell script that:

1. Writes state/session.json
2. cds to the repo
3. Runs claude

Update the PowerShell script paths in .claude/settings.local.json to use pwsh (which is cross-platform).

Troubleshooting

Issue	Solution
Claude doesn’t act as VideoAgent	Launch via video-agent.bat, not plain claude
“Model not found” in workflow	Run inventory scan, then ask to regenerate
ComfyUI won’t connect	Run pwsh -File scripts/connect-comfyui.ps1
Staleness hook not firing	Check .claude/settings.local.json is valid JSON
Skills leaking to other sessions	They shouldn’t – skills are local files, not globally installed

OpenClaw Compatibility

VideoAgent skills also work with OpenClaw. Both platforms follow the AgentSkills specification, so the skill files are cross-compatible. The openclaw/ directory contains everything needed.

How it maps

Claude Code	OpenClaw Equivalent	Notes
CLAUDE.md (orchestrator)	AGENTS.md + SOUL.md + TOOLS.md	Split across three workspace files
video-agent.bat (launcher)	OpenClaw daemon	OpenClaw runs as a persistent service
.claude/settings.local.json (hooks)	openclaw.json (config)	Different config format
Auto-loads from project root	Skills in ~/.openclaw/workspace/skills/	Must be installed to workspace
{skill}/SKILL.md frontmatter	Same + metadata.openclaw block	Already added to all 12 skills

Setup

# 1. Run the setup script (copies skills + workspace files into OpenClaw)
pwsh -File openclaw/setup.ps1

# Or use symlinks to keep them in sync with the repo
pwsh -File openclaw/setup.ps1 -Symlink

# Or specify a custom OpenClaw workspace path
pwsh -File openclaw/setup.ps1 -OpenClawDir "~/.openclaw/workspace"

# 2. Add skill config to your ~/.openclaw/openclaw.json
# See openclaw/openclaw.example.json for the full template -- at minimum:

{
  "skills": {
    "entries": {
      "comfyui-api": {
        "enabled": true,
        "env": { "COMFYUI_URL": "http://127.0.0.1:8188" }
      },
      "comfyui-inventory": {
        "enabled": true,
        "env": { "COMFYUI_PATH": "C:\\ComfyUI" }
      }
    }
  }
}

# 3. Restart OpenClaw to pick up the new skills

What the setup script does

1. Copies (or symlinks) all 12 skill folders into ~/.openclaw/workspace/skills/
2. Copies AGENTS.md, SOUL.md, TOOLS.md into the workspace root
3. Copies foundation/ and references/ alongside skills for reference access

What’s different in OpenClaw

• Skill routing: Claude Code uses a routing table in CLAUDE.md. OpenClaw uses keyword matching on the description field in each skill’s frontmatter – the descriptions are already written to support this.
• Requirements gating: OpenClaw validates metadata.openclaw.requires at load time (checks that binaries/env vars exist). Skills that fail requirements are excluded from the session.
• No session hook: The staleness-check hook is Claude Code specific. In OpenClaw, ask the agent to check for stale research manually, or set up a cron job.
• File references: OpenClaw skills can use {baseDir} to reference files relative to the skill folder. The current skills use relative paths that work in both environments.

Files

openclaw/
|-- AGENTS.md                Orchestration rules (CLAUDE.md equivalent)
|-- SOUL.md                  Agent persona
|-- TOOLS.md                 Available tools and API reference
|-- setup.ps1                Installation script
+-- openclaw.example.json    Config template for ~/.openclaw/openclaw.json

License

MIT