Skip to content
Back to search

comfyui-expert

SkillWorks with Claude CodeGitHub

Session-scoped Claude Code agent for ComfyUI video production — 12 specialized skills for image gen, video, voice cloning, LoRA training, and publishing

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/mckruz-comfyui-expert/badge.svg)](/listings/mckruz-comfyui-expert)

HTML

<a href="/listings/mckruz-comfyui-expert"><img src="/api/listings/mckruz-comfyui-expert/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

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

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 509 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # VideoAgent - ComfyUI Video Production Orchestrator
  2. A session-scoped AI orchestrator for [Claude Code](https://docs.anthropic.com/en/docs/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](https://github.com/comfyanonymous/ComfyUI).
  3. ## Why This Exists
  4. 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:
  5. - _"Generate a photorealistic portrait of a dog using InstantID"_
  6. - _"Make a talking head video where she says 'Welcome to my channel'"_
  7. - _"Train a LoRA from these 20 reference images"_
  8. - _"Check for new ComfyUI video models released this month"_
  9. ...and get validated, hardware-aware ComfyUI workflows without memorizing node names or VRAM budgets.
  10. ## How It Works
  11. ```
  12. video-agent.bat
  13. |
  14. |-- Writes state/session.json (active project, ComfyUI URL)
  15. |-- cd to this repo
  16. |-- Launches: claude
  17. |
  18. |-- Claude Code auto-loads CLAUDE.md (the orchestrator)
  19. |-- Local hooks fire (staleness check)
  20. |-- User's first message triggers foundation reads
  21. |-- Each request is routed to the right skill file
  22. ```
  23. **Key design decisions:**
  24. | Decision | Rationale |
  25. |----------|-----------|
  26. | Session-scoped, not global | Skills stay in this repo. Other Claude Code sessions are unaffected. |
  27. | `CLAUDE.md` is the orchestrator | The only file Claude auto-loads from a project root. Contains the routing table and behavioral instructions. |
  28. | Skills are read-on-demand markdown | No build step, no registration. Claude reads `skills/{name}/SKILL.md` when the routing table says to. |
  29. | REST polling over WebSocket | Claude Code can't hold persistent connections. Polling every 5s works fine for minute-long video generation. |
  30. | Research is user-triggered | No cron jobs. A session-start hook reminds you when data is stale. |
  31. ## Prerequisites
  32. - **[Claude Code](https://docs.anthropic.com/en/docs/claude-code)** installed and authenticated
  33. - **[ComfyUI](https://github.com/comfyanonymous/ComfyUI)** installed (local or remote)
  34. - **[FFmpeg](https://ffmpeg.org/)** on PATH (for video assembly)
  35. - **[PowerShell 7+](https://github.com/PowerShell/PowerShell)** (for utility scripts)
  36. - Windows (the launcher is a `.bat` file; WSL/Linux adaptation is straightforward)
  37. ## Quick Start
  38. ### 1. Clone and launch
  39. ```cmd
  40. git clone https://github.com/MCKRUZ/ComfyUI-Expert.git
  41. cd ComfyUI-Expert
  42. video-agent.bat
  43. ```
  44. With options:
  45. ```cmd
  46. video-agent.bat --project "my-video" # Set active project
  47. video-agent.bat --comfyui "http://<remote-ip>:8188" # Remote ComfyUI
  48. video-agent.bat --resume # Resume last session
  49. ```
  50. ### 2. Scan your ComfyUI installation
  51. First time (or after installing new models/nodes), tell the agent:
  52. ```
  53. Scan my ComfyUI installation at C:\ComfyUI
  54. ```
  55. Or run the script directly:
  56. ```powershell
  57. pwsh -File scripts/scan-inventory.ps1 -ComfyUIPath "C:\ComfyUI"
  58. ```
  59. 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.
  60. ### 3. Start creating
  61. ```
  62. Generate a photorealistic portrait using FLUX
  63. Create a new project called "Character Showcase"
  64. Add a character named Spot - German Shepard, shaggy hair, dark fur
  65. Train a LoRA from these reference images
  66. Research the latest ComfyUI video models
  67. ```
  68. ## Architecture
  69. ### 3-Tier Context System
  70. VideoAgent loads context incrementally to stay within Claude's context window:
  71. | Tier | Files | Loaded When | Size |
  72. |------|-------|-------------|------|
  73. | **1: Foundation** | `foundation/*.md` | Session start (first interaction) | ~2K tokens |
  74. | **2: Working** | `projects/{name}/*` | When working on a specific project | Varies |
  75. | **3: Reference** | `references/*.md` | Only when a skill explicitly needs detail | Large |
  76. ### Skill Dependency Graph
  77. ```
  78. CLAUDE.md (orchestrator - always loaded)
  79. |
  80. |-- Discovery (no dependencies)
  81. | |-- comfyui-prompt-interview Vision clarification via conversation
  82. |
  83. |-- Foundation Skills (no dependencies)
  84. | |-- comfyui-api REST API connection
  85. | |-- comfyui-inventory Model/node discovery
  86. | |-- project-manager Project & character state
  87. |
  88. |-- Research (independent)
  89. | |-- comfyui-research Self-updating knowledge base
  90. |
  91. |-- Core Creation (depend on inventory)
  92. | |-- comfyui-prompt-engineer Model-specific prompt optimization
  93. | |-- comfyui-workflow-builder Validated workflow JSON generation
  94. |
  95. |-- Production (depend on creation)
  96. | |-- comfyui-video-pipeline Wan 2.6 / LTX-2.3 / FramePack / AnimateDiff
  97. | |-- comfyui-voice-pipeline Qwen3-TTS / Chatterbox / F5-TTS / lip-sync
  98. | |-- comfyui-lora-training Dataset prep, training, evaluation
  99. |
  100. |-- Output (depend on production)
  101. | |-- video-assembly FFmpeg + Remotion composition
  102. | |-- video-publisher YouTube metadata & upload
  103. |
  104. |-- Support
  105. |-- comfyui-troubleshooter Error diagnosis & fixes
  106. ```
  107. ### Request Routing
  108. When you make a request, `CLAUDE.md` routes it to the right skill:
  109. | You Say | Skill Loaded | What Happens |
  110. |---------|-------------|--------------|
  111. | "I have an idea for..." / vague concept | `comfyui-prompt-interview` | Guided conversation to clarify vision before generation |
  112. | "Generate a character portrait" | `comfyui-workflow-builder` | Checks inventory, builds workflow JSON, queues via API |
  113. | "Craft a better prompt" | `comfyui-prompt-engineer` | Model-specific optimization (FLUX vs SDXL vs Wan) |
  114. | "Create a video from this image" | `comfyui-video-pipeline` | Selects engine (Wan 2.6/FramePack/LTX-2.3), builds pipeline |
  115. | "Clone this voice / make her talk" | `comfyui-voice-pipeline` | Voice synthesis + lip-sync pipeline |
  116. | "Train a LoRA" | `comfyui-lora-training` | Dataset prep, training config, checkpoint evaluation |
  117. | "Build a raw workflow" | `comfyui-workflow-builder` | Direct workflow construction with inventory validation |
  118. | "Check for new models" | `comfyui-research` | Scans YouTube/GitHub/HuggingFace, updates references |
  119. | "Something broke" | `comfyui-troubleshooter` | Error pattern matching, fix suggestions |
  120. | "Assemble the final video" | `video-assembly` | FFmpeg or Remotion-based composition |
  121. | "Upload to YouTube" | `video-publisher` | Metadata generation + upload delegation |
  122. | "Create a new project" | `project-manager` | Project manifests, character profiles |
  123. | "Connect to ComfyUI" | `comfyui-api` | Connection test, system info |
  124. ## The 13 Skills
  125. ### Discovery
  126. **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`.
  127. ### Foundation
  128. **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).
  129. **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`).
  130. **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.
  131. ### Research
  132. **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.
  133. ### Core Creation
  134. **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.
  135. **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.
  136. ### Production
  137. **comfyui-video-pipeline** -- Orchestrates video engines based on requirements:
  138. - **LTX-2.3**: 4K production quality, audio+video, GGUF-quantized options
  139. - **Wan 2.6**: 1080p reference-to-video, native audio generation, built-in lip-sync
  140. - **Wan 2.2 MoE 14B**: Film-level quality, first+last frame control, 24GB+ VRAM
  141. - **FramePack**: Long videos (60+ sec), VRAM-invariant (works on 6GB)
  142. - **HunyuanVideo 1.5**: Lightweight quality alternative at 8.3B params
  143. - **AnimateDiff V3**: Fast iteration, motion LoRAs, 4-8 step Lightning
  144. Includes post-processing (RIFE frame interpolation, face enhancement, deflicker, color correction) and a dedicated talking-head pipeline.
  145. **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.
  146. **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.
  147. ### Output
  148. **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).
  149. **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).
  150. ### Support
  151. **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.
  152. ## Model Landscape
  153. The agent tracks the top models across five categories:
  154. ### Image Generation
  155. | Model | Best For | VRAM |
  156. |-------|----------|------|
  157. | FLUX.2 [dev] | Photorealism, 4MP, multi-reference (up to 10 images) | 24GB+ |
  158. | FLUX.2 [klein] | Fast generation, low VRAM (4B/9B distilled) | 12-20GB+ |
  159. | FLUX Kontext | Iterative character editing | 12-32GB |
  160. | Qwen-Image 2.0 | Typography, 2K resolution, layered editing | 24GB+ |
  161. | Z-Image | Non-distilled quality (Base) / fast (Turbo, 8 steps) | 12-16GB+ |
  162. ### Identity Preservation
  163. | Method | Best For | VRAM |
  164. |--------|----------|------|
  165. | InfiniteYou | Highest identity fidelity | 24GB |
  166. | FLUX Kontext | Edit without retraining | 12-32GB |
  167. | PuLID Flux 2 | FLUX.2 family (Klein + Dev) | 24-40GB |
  168. | PuLID Flux II | FLUX.1 dual characters, no pollution | 24-40GB |
  169. ### Video Generation
  170. | Model | Best For | VRAM |
  171. |-------|----------|------|
  172. | LTX-2.3 | 4K audio+video, portrait mode, production | 24GB+ |
  173. | Wan 2.6 | Reference-to-video, lip-sync, native audio gen, 1080p | 24GB+ |
  174. | Wan 2.2 MoE | Film-level quality, first+last frame control | 24GB+ |
  175. | HunyuanVideo 1.5 | Lightweight flagship quality (8.3B params) | 24GB |
  176. | FramePack | Long videos (60s+), VRAM-invariant | 6GB+ |
  177. | SkyReels V1 | Human-centric, 33 expressions, cinematic | 24GB+ |
  178. | AnimateDiff V3 | Fast iteration, motion LoRAs | 8GB+ |
  179. ### Voice / TTS
  180. | Tool | Best For | License |
  181. |------|----------|---------|
  182. | TTS Audio Suite | Unified 11-engine platform, 23 languages | Multi |
  183. | Qwen3-TTS | 10 languages, zero-shot clone, voice design | Open |
  184. | Chatterbox Turbo | Emotion tags, sub-200ms latency | MIT |
  185. | IndexTTS-2 | 8-emotion vector control via sliders | Open |
  186. | F5-TTS | Zero-shot cloning, works on 6GB VRAM | MIT |
  187. ### Lip-Sync
  188. | Tool | Best For |
  189. |------|----------|
  190. | LatentSync 1.6 | Highest accuracy (ByteDance) |
  191. | Wan 2.6 native | Reference-to-video with built-in lip-sync |
  192. | Wav2Lip | Proven, works with any face |
  193. | SadTalker | Head movement + expressions |
  194. Full specs and download links are in `references/models.md`.
  195. ## Hardware Profile
  196. VideoAgent is configured for an **RTX 5090 (32GB VRAM)** but works with any GPU. The agent adjusts recommendations based on available VRAM.
  197. | Workload | 32GB Status | Notes |
  198. |----------|:-----------:|-------|
  199. | FLUX.1-dev FP16 | Native | No quantization needed |
  200. | Wan 2.2 14B | Native | Full quality |
  201. | FramePack | Overkill | Designed for 6GB |
  202. | PuLID Flux II | Native | Dual-character generation |
  203. | InfiniteYou | Native | Both SIM and AES variants |
  204. | LoRA Training (FLUX) | Native | No quantization needed |
  205. Recommended ComfyUI launch flags: `--highvram --fp8_e4m3fn-unet`
  206. ## Project Structure
  207. ```
  208. ComfyUI-Expert/
  209. |-- video-agent.bat Launcher (writes session config, opens Claude)
  210. |-- CLAUDE.md Orchestrator (routing table, behavior, rules)
  211. |-- .claude/
  212. | +-- settings.local.json Project-local hooks & permissions
  213. |
  214. |-- foundation/ Tier 1: Quick reference (~2K tokens)
  215. | |-- agent-persona.md Communication style & principles
  216. | |-- api-quick-ref.md ComfyUI REST API cheat sheet
  217. | |-- hardware-profile.md GPU specs, VRAM capabilities
  218. | |-- model-landscape.md Top 3 models per category
  219. | +-- skill-registry.md Skill list & dependency map
  220. |
  221. |-- skills/ 13 skill modules (read on demand)
  222. | |-- comfyui-api/
  223. | |-- comfyui-inventory/
  224. | |-- comfyui-lora-training/
  225. | |-- comfyui-prompt-engineer/
  226. | |-- comfyui-prompt-interview/
  227. | |-- comfyui-research/
  228. | |-- comfyui-troubleshooter/
  229. | |-- comfyui-video-pipeline/
  230. | |-- comfyui-voice-pipeline/
  231. | |-- comfyui-workflow-builder/
  232. | |-- project-manager/
  233. | |-- video-assembly/
  234. | +-- video-publisher/
  235. |
  236. |-- references/ Tier 3: Deep reference (loaded on demand)
  237. | |-- models.md Full model catalog & download links
  238. | |-- workflows.md Complete workflow node configurations
  239. | |-- lora-training.md Training parameters & best practices
  240. | |-- voice-synthesis.md Voice tools in depth
  241. | |-- prompt-templates.md Model-specific prompt strategies
  242. | |-- troubleshooting.md Error database with solutions
  243. | |-- research-log.md Full technique survey (ongoing, replaces research-2025.md)
  244. | |-- staleness-report.md Freshness tracking for all entries
  245. | +-- evolution.md Update protocol & changelog
  246. |
  247. |-- projects/ Per-project state (gitignored)
  248. |-- state/ Runtime state (gitignored)
  249. | |-- session.json Active project & ComfyUI URL
  250. | +-- inventory.json Cached models/nodes from scan
  251. |
  252. |-- scripts/ Utility scripts
  253. | |-- scan-inventory.ps1 Offline ComfyUI directory scanner
  254. | |-- connect-comfyui.ps1 Connection test & diagnostics
  255. | |-- staleness-check.ps1 Session-start hook (checks research age)
  256. | |-- deploy.ps1 Sync references to global skill
  257. | +-- sync-skills.ps1 Sync all skills to global Claude skills directory
  258. |
  259. |-- agent/
  260. | +-- AGENT.md Extended orchestration spec
  261. |
  262. |-- openclaw/ OpenClaw compatibility layer
  263. | |-- AGENTS.md Orchestration rules (CLAUDE.md equivalent)
  264. | |-- SOUL.md Agent persona
  265. | |-- TOOLS.md Available tools & API reference
  266. | |-- setup.ps1 Install skills into OpenClaw workspace
  267. | +-- openclaw.example.json Config template
  268. |
  269. +-- docs/
  270. |-- architecture.md System design decisions
  271. +-- getting-started.md Quick start guide
  272. ```
  273. ## Example Workflows
  274. ### Character Image Generation
  275. ```
  276. 1. "Create a new project called Character Showcase"
  277. 2. "Add a character named Sage - auburn hair, green eyes, freckles"
  278. 3. "Generate a photorealistic portrait of Sage using InstantID"
  279. ```
  280. 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.
  281. ### Talking Head Video
  282. ```
  283. 1. "Make Sage say 'Hello everyone, welcome to my channel'"
  284. ```
  285. The agent orchestrates a multi-step pipeline: voice synthesis (Chatterbox/F5-TTS) -> video generation (Wan 2.2) -> lip-sync (LatentSync) -> face enhancement -> assembly.
  286. ### Research Update
  287. ```
  288. 1. "Check for new ComfyUI models and techniques"
  289. ```
  290. The agent: loads research skill -> checks YouTube channels, GitHub repos, HuggingFace trending -> extracts knowledge from tutorials -> updates reference files -> generates staleness report.
  291. ## Hooks & Automation
  292. VideoAgent uses Claude Code's [hooks system](https://docs.anthropic.com/en/docs/claude-code/hooks) for lightweight automation:
  293. | Hook | Event | What It Does |
  294. |------|-------|-------------|
  295. | Staleness check | `SessionStart` | Warns if research data is older than 2 weeks |
  296. Configured in `.claude/settings.local.json` (project-local, doesn't affect other sessions).
  297. ## Scripts Reference
  298. | Script | Purpose | Usage |
  299. |--------|---------|-------|
  300. | `scan-inventory.ps1` | Scan ComfyUI models & nodes offline | `pwsh -File scripts/scan-inventory.ps1 -ComfyUIPath "C:\ComfyUI"` |
  301. | `connect-comfyui.ps1` | Test ComfyUI connection & show diagnostics | `pwsh -File scripts/connect-comfyui.ps1` |
  302. | `staleness-check.ps1` | Check research freshness (session hook) | Runs automatically at session start |
  303. | `deploy.ps1` | Sync references to global `comfyui-character-gen` skill | `pwsh -File scripts/deploy.ps1` |
  304. | `sync-skills.ps1` | Sync all skill files to global Claude skills directory | `pwsh -File scripts/sync-skills.ps1` |
  305. ## Customization
  306. ### Different GPU
  307. Edit `foundation/hardware-profile.md` with your GPU specs. The agent reads this at session start and adjusts VRAM recommendations accordingly.
  308. ### Different ComfyUI location
  309. Pass it at launch:
  310. ```cmd
  311. video-agent.bat --comfyui "http://<remote-ip>:8188"
  312. ```
  313. Or edit the default in `video-agent.bat` (line 18).
  314. ### Adding models to the landscape
  315. 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.
  316. ### Adapting for Linux/macOS
  317. Replace `video-agent.bat` with a shell script that:
  318. 1. Writes `state/session.json`
  319. 2. `cd`s to the repo
  320. 3. Runs `claude`
  321. Update the PowerShell script paths in `.claude/settings.local.json` to use `pwsh` (which is cross-platform).
  322. ## Troubleshooting
  323. | Issue | Solution |
  324. |-------|---------|
  325. | Claude doesn't act as VideoAgent | Launch via `video-agent.bat`, not plain `claude` |
  326. | "Model not found" in workflow | Run inventory scan, then ask to regenerate |
  327. | ComfyUI won't connect | Run `pwsh -File scripts/connect-comfyui.ps1` |
  328. | Staleness hook not firing | Check `.claude/settings.local.json` is valid JSON |
  329. | Skills leaking to other sessions | They shouldn't -- skills are local files, not globally installed |
  330. ## OpenClaw Compatibility
  331. VideoAgent skills also work with [OpenClaw](https://github.com/openclaw/openclaw). Both platforms follow the [AgentSkills specification](https://agentskills.io), so the skill files are cross-compatible. The `openclaw/` directory contains everything needed.
  332. ### How it maps
  333. | Claude Code | OpenClaw Equivalent | Notes |
  334. |-------------|--------------------|----|
  335. | `CLAUDE.md` (orchestrator) | `AGENTS.md` + `SOUL.md` + `TOOLS.md` | Split across three workspace files |
  336. | `video-agent.bat` (launcher) | OpenClaw daemon | OpenClaw runs as a persistent service |
  337. | `.claude/settings.local.json` (hooks) | `openclaw.json` (config) | Different config format |
  338. | Auto-loads from project root | Skills in `~/.openclaw/workspace/skills/` | Must be installed to workspace |
  339. | `{skill}/SKILL.md` frontmatter | Same + `metadata.openclaw` block | Already added to all 12 skills |
  340. ### Setup
  341. ```powershell
  342. # 1. Run the setup script (copies skills + workspace files into OpenClaw)
  343. pwsh -File openclaw/setup.ps1
  344. # Or use symlinks to keep them in sync with the repo
  345. pwsh -File openclaw/setup.ps1 -Symlink
  346. # Or specify a custom OpenClaw workspace path
  347. pwsh -File openclaw/setup.ps1 -OpenClawDir "~/.openclaw/workspace"
  348. ```
  349. ```powershell
  350. # 2. Add skill config to your ~/.openclaw/openclaw.json
  351. # See openclaw/openclaw.example.json for the full template -- at minimum:
  352. ```
  353. ```json
  354. {
  355. "skills": {
  356. "entries": {
  357. "comfyui-api": {
  358. "enabled": true,
  359. "env": { "COMFYUI_URL": "http://127.0.0.1:8188" }
  360. },
  361. "comfyui-inventory": {
  362. "enabled": true,
  363. "env": { "COMFYUI_PATH": "C:\\ComfyUI" }
  364. }
  365. }
  366. }
  367. }
  368. ```
  369. ```powershell
  370. # 3. Restart OpenClaw to pick up the new skills
  371. ```
  372. ### What the setup script does
  373. 1. Copies (or symlinks) all 12 skill folders into `~/.openclaw/workspace/skills/`
  374. 2. Copies `AGENTS.md`, `SOUL.md`, `TOOLS.md` into the workspace root
  375. 3. Copies `foundation/` and `references/` alongside skills for reference access
  376. ### What's different in OpenClaw
  377. - **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.
  378. - **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.
  379. - **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.
  380. - **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.
  381. ### Files
  382. ```
  383. openclaw/
  384. |-- AGENTS.md Orchestration rules (CLAUDE.md equivalent)
  385. |-- SOUL.md Agent persona
  386. |-- TOOLS.md Available tools and API reference
  387. |-- setup.ps1 Installation script
  388. +-- openclaw.example.json Config template for ~/.openclaw/openclaw.json
  389. ```
  390. ## License
  391. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.