Autopilot Shorts

Fully automated YouTube Shorts pipeline that generates and uploads viral short-form content -- and learns from its own performance to get better over time.

Topic -> Script -> AI Images -> TTS -> Video -> Upload -> Learn -> Repeat

Zero manual steps. Set it on a cron and walk away.

---

Why This Exists

Most "YouTube automation" tools generate a video and call it done. This pipeline closes the loop: it scrapes your channel's actual view counts, identifies what worked and what didn't, and feeds that data back into the next generation prompt. Over time, the content evolves toward what your audience actually watches.

It also fights content repetition with 5 independent diversity systems -- topic rotation, hook sampling, voice rotation, history exclusion, and performance-informed generation -- so your channel never feels like a bot even though it is one.

How It Works

                    +------------------+
                    |   cron (3x/day)  |
                    +--------+---------+
                             |
                    +--------v---------+
                    | Scrape YouTube    |  yt-dlp grabs view counts
                    | metrics           |  for all your videos
                    +--------+---------+
                             |
                    +--------v---------+
                    | Generate topic    |  LLM picks topic informed by
                    | (self-improving)  |  top/bottom performers
                    +--------+---------+
                             |
                    +--------v---------+
                    | Write script      |  3-4 sentences, scroll-stopping
                    | + hook sampling   |  hook, diverse patterns
                    +--------+---------+
                             |
               +-------------+-------------+
               |                           |
     +---------v----------+     +----------v---------+
     | Generate 7 images  |     | Text-to-speech     |
     | (Together.ai FLUX) |     | (edge-tts, rotated |
     |                    |     |  across 11 voices)  |
     +--------+-----------+     +----------+----------+
              |                            |
              +-------------+--------------+
                            |
                   +--------v---------+
                   | Render video      |  MoviePy: images + audio
                   | 1080x1920         |  + subtitles + background music
                   +--------+---------+
                            |
                   +--------v---------+
                   | Upload to YouTube |  Selenium drives Firefox
                   | via Studio        |  headless on server
                   +--------+---------+
                            |
                   +--------v---------+
                   | Save to history   |  Track topic/hook/voice
                   | + clean up files  |  for future diversity
                   +-------------------+

Stack

Component	Tool
LLM	Ollama (any model -- qwen2.5:3b, llama3, etc.)
Image Gen	Together.ai FLUX
TTS	edge-tts (11 voices, auto-rotated)
Video Render	MoviePy + ImageMagick
Subtitles	faster-whisper
Upload	Selenium + Firefox (headless)
Metrics	yt-dlp
Scheduling	cron (or any scheduler)

Runs on any Linux box. Tested on AWS EC2.

Setup

1. Prerequisites

• Python 3.12
• Ollama installed with a model pulled:

  curl -fsSL https://ollama.ai/install.sh | sh
  ollama pull qwen2.5:3b

• ImageMagick:

  sudo apt install imagemagick    # Ubuntu/Debian
  brew install imagemagick        # macOS

• Firefox (for headless upload):

  sudo apt install firefox

• Together.ai API key -- sign up at together.ai for image generation

2. Install

git clone https://github.com/CodeSarthak/autopilot-shorts.git
cd autopilot-shorts

python3.12 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

cp config.example.json config.json

3. Configure

Edit config.json with your settings. The critical ones:

{
  "firefox_profile": "/path/to/your/firefox/profile",
  "headless": true,
  "nanobanana2_api_key": "your-together-ai-api-key",
  "imagemagick_path": "/usr/bin/convert"
}

See docs/Configuration.md for all options.

4. Set Up Firefox Profile

The upload uses Selenium to drive Firefox, which needs to be logged into YouTube Studio:

1. Create a dedicated Firefox profile: firefox -P (create a new profile)
2. Open YouTube Studio in that profile and log in
3. Copy the profile path (find it at about:profiles in Firefox)
4. Set firefox_profile in config.json to that path

5. Configure YouTube Account

Create .mp/youtube.json:

[
  {
    "id": "any-unique-uuid",
    "nickname": "MyChannel",
    "firefox_profile": "/path/to/firefox/profile",
    "niche": "your channel niche (e.g. tech, fitness, finance, history)",
    "language": "English"
  }
]

6. Add Background Music

Drop .mp3 files into the Songs/ directory. One is randomly selected per video. Use royalty-free tracks.

7. Run

# Single run
venv/bin/python3.12 src/cron.py youtube YOUR_ACCOUNT_UUID qwen2.5:3b

# Automate with cron (3x/day example)
crontab -e

0 6 * * * cd /path/to/autopilot-shorts && venv/bin/python3.12 src/cron.py youtube YOUR_UUID qwen2.5:3b >> pipeline.log 2>&1
0 14 * * * cd /path/to/autopilot-shorts && venv/bin/python3.12 src/cron.py youtube YOUR_UUID qwen2.5:3b >> pipeline.log 2>&1
0 22 * * * cd /path/to/autopilot-shorts && venv/bin/python3.12 src/cron.py youtube YOUR_UUID qwen2.5:3b >> pipeline.log 2>&1

The Self-Improving Loop

This is what makes this project different. Before every run:

1. yt-dlp scrapes view counts for all videos on your channel
2. The top 3 and bottom 3 performers are identified
3. This data is injected into the LLM topic generation prompt:

LEARN FROM PAST PERFORMANCE -- these titles got the MOST views on our channel:
- How One Decision Changed Everything (842 views)
- The Hidden Cost Nobody Talks About (717 views)
These titles got the FEWEST views:
- Topic That Was Too Broad (2 views)
- Generic Title Without a Hook (0 views)
Generate topics that lean toward what performed well.

The LLM learns to generate content similar to what actually gets views. The more videos you publish, the better the signal, the better the content.

Content Diversity System

Five independent systems prevent your channel from feeling robotic:

System	How It Works
Topic categories	8 default categories (customize in youtube.json) -- one randomly picked per run
Hook patterns	Pool of 15 templates -- 3-4 randomly sampled per prompt
Voice rotation	11 edge-tts voices (male/female, US/UK/AU) -- last 3 excluded
History exclusion	Last 10 topics and hooks are excluded from generation
Performance feedback	Top/bottom performers steer future topics

License

AGPL-3.0. See LICENSE.

Credits

Originally forked from MoneyPrinterV2.