Khali — YouTube Shorts Factory
AI-powered YouTube Shorts production system. Catches trends before they peak, extracts transcripts, writes viral 60-second scripts, and generates videos — all from Telegram.
Built on OpenClaw so it's low-code and hands-off. Built to teach you how AI agent systems work while you use one.
---
What It Does
You send a YouTube link on Telegram
→ AI extracts the transcript
→ AI writes a 60-second short script
→ You approve (or edit)
→ AI generates a video with Google Veo 3.1 Lite
→ Video downloads to your machinePlus: A trend-hunting agent scans YouTube every 6 hours for rising videos in your niches, scores them, and alerts you before they peak.
---
Quick Start
# 1. Install OpenClaw (https://docs.openclaw.ai/start/getting-started)
curl -fsSL https://openclaw.ai/install.sh | bash
# 2. Clone this repo
git clone https://github.com/ciscoittech/khali-mvp.git
cd khali-mvp
# 3. Set up environment
cp .env.example .env
# Add your OPENROUTER_API_KEY and TELEGRAM_BOT_TOKEN
# 4. Build plugins
cd plugins/youtube-extractor && npm install && npm run build && cd ../..
cd plugins/video-generator && npm install && npm run build && cd ../..
cd plugins/trend-scout && npm install && npm run build && cd ../..
# 5. Start
openclaw startNeed an OpenRouter key? Sign up at openrouter.ai — it powers both the script writing (DeepSeek) and video generation (Veo).
Need a Telegram bot? Message @BotFather on Telegram, create a bot, copy the token.
---
How to Get the Most Out of This
Step 1: Start Using It (Day 1)
Open Telegram and send a YouTube link. That's it. The system will: 1. Pull the transcript 2. Write a 60-second script with a hook 3. Ask you to approve 4. Generate the video
The coach is always on — after every action, it explains what just happened and why. You don't need to understand the system to use it. You'll learn by doing.
Step 2: Read the Primer (When You're Curious)
Type primer in Telegram. You'll get a 5-minute guide covering:
- What OpenClaw is and how skills, plugins, and channels connect
- Systems Thinking (Donella Meadows) — see stocks, flows, and feedback loops in your production
- Theory of Constraints (Goldratt) — find the ONE bottleneck and fix it
- Hands-on exercises that teach each concept by doing it
- Documentation links to go deeper on anything
You can skip around — it's a reference, not a textbook. Come back to it anytime.
Step 3: Find Your Bottleneck (Week 1)
After producing a few shorts, type status. The daily digest shows:
- Throughput: shorts produced this week
- Inventory: scripts waiting for your approval
- Constraint: what's currently slowing the system down
If scripts are piling up, you are the constraint (Goldratt, Chapter 13). The system will tell you: "Your bottleneck is approval — I'm pausing scouting until you clear the queue."
Step 4: Catch Trends Early (Week 2)
Type find trends or let the auto-scanner run every 6 hours. It surfaces rising videos scored 1-10 on "shorts potential." When you see one you like, just say short 1 — the full pipeline runs automatically.
Configure your niches: set niches: tech, AI, cooking
Step 5: Assignment 001 — Close the Loop (Week 2-3)
Your first real assignment: connect Gmail to close the feedback loop.
Right now the system is an open loop — it makes shorts but never hears back. Did people comment? Did the algorithm pick it up? You're flying blind. That's the most dangerous thing in systems thinking (Meadows).
Type assignment 1 in Telegram for the full brief. Here's the summary:
1. Connect Gmail to OpenClaw as a channel 2. Route YouTube notifications (comments, analytics, subscribers) through a Gmail filter 3. Write your own skill (youtube-inbox-monitor.md) that reads comment emails and reports back 4. Answer 5 questions about what your inbox means as a system — what other emails are signals? Where does the constraint move? What's the leverage point?
You can use AI to help with the technical parts. The questions are what matter — they teach you to think in systems, not just use them. See skills/assignment-001-email-feedback-loop.md for the full guide.
Step 6: Change the Rules (Month 1+)
This is where the real power is. Open skills/shorts-producer.md in a text editor and change the hook formulas. That's leverage point 4 (Meadows) — you just changed a rule that affects every script the system writes.
Type explain mode to see exactly what prompts the AI receives and why. Turn it off with explain mode off.
Step 7: Go Deeper
| Want to... | Do this | |-----------|---------| | Understand the theory | Type primer — covers Meadows & Goldratt mapped to Khali | | See detailed breakdowns | Type explain mode — shows prompts, inputs/outputs, cinematography terms | | Learn OpenClaw concepts | Just use it — the coach explains after every action | | Build your own skills | Read Creating Skills, then edit any .md file in skills/ | | Add new tools | Read Building Plugins | | Browse community skills | Check ClawHub | | Check API costs | Visit OpenRouter Activity |
---
The Team (6 AI Skills)
| Skill | What It Does | When It Activates | |-------|-------------|-------------------| | shorts-producer | End-to-end: YouTube URL → transcript → script → video | When you send a YouTube link | | trend-hunter | Proactive trend discovery, scores videos 1-10 | Every 6 hours + "find trends" | | daily-digest | Morning brief + evening recap with systems health | 8am/9pm + "status" | | coach | Always-on mentor — explains with systems language | After every action | | explain-mode | Toggle deep-dive breakdowns of every step | "explain mode" on/off | | primer | Systems thinking foundation (Meadows + Goldratt) | "primer" anytime | | assignment-001 | First assignment: connect Gmail as a feedback loop | "assignment 1" |
The Tools (3 Plugins)
| Plugin | Tools | What It Calls | |--------|-------|---------------| | youtube-extractor | extract_youtube_transcript, write_shorts_script, scan_youtube_trends | YouTube captions + DeepSeek V4 Flash | | video-generator | generate_video | Google Veo 3.1 Lite via OpenRouter | | trend-scout | check_google_trends, get_channel_recent | Google Trends + YouTube RSS |
---
The Theory (Why It Works This Way)
This system is designed around two books. You don't need to read them — the primer and coach teach the concepts as you use the system — but they're worth reading if you want to go deep.
Thinking in Systems — Donella Meadows (2008)
> "A system is more than the sum of its parts."
Teaches you to see Khali as interconnected stocks, flows, and feedback loops — not separate tools. The trend scanner is an inflow. Your content library is a stock. Good scripts create a reinforcing loop with the YouTube algorithm. The primer maps every concept to what Khali actually does.
The Goal — Eliyahu Goldratt (1984)
> "The goal is not to improve one measurement in isolation. The goal is to improve the THROUGHPUT of the entire system."
Teaches you to find the ONE bottleneck and focus there. In Khali, the constraint is usually your approval speed — everything else is automated. The daily digest tracks throughput, inventory, and operating expense so you can see it.
---
Using Claude Code / Claude Desktop
This repo includes a built-in Claude agent and commands that run on Haiku (cheapest model) to help you learn and troubleshoot.
Setup
1. Open this repo in Claude Code or add it to Claude Desktop 2. The khali-guide agent is auto-discovered from .claude/agents/
Commands
| Command | What it does | |---------|-------------| | /help | Show all commands, prompts, and troubleshooting guides | | /primer | Walk through the systems thinking theory interactively | | /assignment | View your current assignment and track progress | | /status | Check system health — plugins, config, env, downloads | | /debug [error] | Paste an error message, get a diagnosis and fix |
Prompts You Can Use
Learning:
"explain [concept]"— plain-English explanation of any concept"why did that happen"— understand the last error or result"show me an example"— see working code for what you're building"what should I do next"— get guided next steps
Building:
"help me write a skill"— step-by-step skill creation"help me write a plugin"— step-by-step plugin creation"help me connect gmail"— Assignment 001 walkthrough"help me set up telegram"— bot setup walkthrough
Troubleshooting:
"debug [paste error]"— paste any error and get a fix"check my plugin"— validate plugin structure and build"check my skill"— validate skill markdown format"check my env"— verify .env has required keys"why won't it start"— diagnose OpenClaw startup issues
Common Issues
| Problem | Fix | |---------|-----| | openclaw: command not found | Install: curl -fsSL https://openclaw.ai/install.sh \| bash | | Cannot find module 'openclaw/plugin-sdk' | Build the plugin: cd plugins/[name] && npm install && npm run build | | 401 / unauthorized | Check .env has a valid OPENROUTER_API_KEY — get one at openrouter.ai | | address already in use :3126 | Another process on that port. Run lsof -i :3126 then kill -9 [PID] | | No captions available | Video doesn't have subtitles. Try a different video. | | content may have been filtered (Veo) | Prompt hit safety filter — remove celebrity names and retry | | Video generation timed out | Veo was slow. Try again, or check OpenRouter status | | Gmail won't connect | Need an App Password, not your regular password. 2FA required. | | Bot not responding on Telegram | Check TELEGRAM_BOT_TOKEN in .env. Make sure no other instance is using the same bot. |
---
Documentation
OpenClaw
- Getting Started — install and first chat in 5 minutes
- Skills — how agent behaviors work
- Creating Skills — write your own
- Plugins — how tools work
- Building Plugins — add new tools
- Telegram — bot setup
- Channels — all messaging platforms
- Cron / Scheduling — automated tasks
- Standing Orders — persistent rules
- Agent Loop — how the AI thinks
- Configuration — every setting
- ClawHub — community skills & plugins
- OpenClaw GitHub
OpenRouter
- Quickstart — get your API key
- API Reference — full API docs
- All Models — browse 400+ models
- Video Generation — how Veo works
- Usage & Costs — track spending
---
Project Structure
khali-mvp/
├── openclaw.json # Agent config (port 3126, Telegram + Email)
├── .env.example # Secrets template
│
├── skills/ # Agent behaviors (edit these to change how the AI works)
│ ├── primer.md # Systems thinking foundation + docs + exercises
│ ├── coach.md # Always-on mentor with 4 learning levels
│ ├── explain-mode.md # Deep-dive toggle for detailed breakdowns
│ ├── shorts-producer.md # YouTube → transcript → script → video pipeline
│ ├── trend-hunter.md # Trend discovery with Drum-Buffer-Rope pacing
│ └── daily-digest.md # Stats + systems health check
│
├── plugins/ # Tools (the code the AI calls)
│ ├── youtube-extractor/ # Transcript extraction + trend scanning
│ ├── video-generator/ # Veo 3.1 Lite video generation
│ └── trend-scout/ # Google Trends + YouTube RSS monitoring
│
├── downloads/ # Generated videos land here
└── services/agents/ # Legacy FastAPI prototype---
Built With the Framework From "From Prompts to Pipelines"
This system uses the same four concepts taught in the book:
| Concept | Where You See It in Khali | |---------|--------------------------| | Instruction | The skill files. shorts-producer.md tells the system exactly how to write scripts, what hooks to use, what format to follow. | | Memory | Your trend scores, production stats, and content library persist between sessions. The system remembers what worked. | | Control | The coach checks your output. The approval step prevents bad scripts from becoming bad videos. The trend scorer rejects low-potential topics. | | Flow | YouTube link in, video out. Five stages, each feeding the next. Remove one and the pipeline breaks. |
Want to go deeper? Read the free chapter: Why That Worked (And Why Most AI Doesn't). It teaches you the diagnostic framework you can use on Khali or any AI system.
Want to build your own? From Prompts to Pipelines teaches you to build three complete systems from scratch using OpenClaw, each more complex than the last. Khali uses the same patterns.
---
License
MIT
