youtube-shorts-pipeline

Verticals v3

The open source AI content engine with built-in niche intelligence.

Topic in. Published Short out. Any niche. ~$0.11 per video.

Quickstart · Hosted Version

Repo note: the product is called Verticals v3. The GitHub repository is youtube-shorts-pipeline.

python -m verticals run --topic "Sam Altman just mass-fired 200 safety researchers" --niche tech

That one command researches the topic, writes a hook driven script tuned to tech YouTube, generates cinematic b roll, records a natural voiceover, burns in animated captions, adds mood matched background music, generates a thumbnail, and uploads it to YouTube. ~90 seconds of video, ~3 minutes of wall time, ~$0.11 in API costs.

What Changed in v3

v2 was an esports news pipeline. v3 is a general purpose content engine that works for any niche, any topic, any creator.

The biggest change: Niche Intelligence. Every stage of the pipeline now reads from a niche profile that shapes script tone, visual style, caption aesthetics, music mood, and thumbnail strategy. Ship a cooking Short and it writes like a cooking creator, generates food photography b roll, and picks warm upbeat background music. Ship a true crime Short and the tone shifts to suspenseful, the visuals go dark and cinematic, and the music drops to ambient tension.

15 niches ship out of the box. Build your own in 5 minutes.

Other highlights: multi provider LLM support (Claude, Gemini, GPT, Ollama local), free TTS via Edge TTS, YouTube upload, topic discovery, resumable stages, and a local-first config model.

Current Release: v3.1.0

v3.1.0 brings community-contributed providers and reliability fixes: MiniMax (LLM + TTS) and 60db (TTS) as optional providers, niche-aware caption fonts so CJK and other non-Latin scripts render correctly, a working edge-tts pin (6.x is rejected by Microsoft with 403s), and clearer errors for the most-reported setup problems.

Implemented today:

• research with DuckDuckGo plus optional source scraping
• script and metadata generation through Claude, Gemini, GPT, Ollama, MiniMax, LiteLLM, or Claude CLI
• b roll and thumbnail image generation through Gemini Imagen, with fallback frames
• voiceover through Edge TTS, ElevenLabs, MiniMax, 60db, or macOS say
• Whisper captions with ASS burn-in plus SRT export, niche-configurable fonts
• ffmpeg assembly with Ken Burns motion, background music, and voice ducking
• private-by-default YouTube upload

Not shipped yet: Gradio UI, Docker, Colab, TikTok/Reels/X upload, Pexels, Replicate, ComfyUI, and Kokoro TTS. Those are roadmap items, not current features.

How It Works

┌─────────────────────────────────────────────────────────────────┐
│                        NICHE PROFILE                            │
│  Loaded once. Shapes every stage. 15 built in or bring your own │
└─────────────┬───────────────────────────────────────────────────┘
              │
              ▼
┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐
│ RESEARCH │→ │  SCRIPT  │→ │ VISUALS  │→ │  VOICE   │→ │ CAPTIONS │→ │ ASSEMBLE │→ UPLOAD
│          │  │          │  │          │  │          │  │          │  │          │
│ DuckDuck │  │ LLM with │  │ Gemini   │  │ ElevenLabs│  │ Whisper  │  │ ffmpeg   │
│ Go + web │  │ niche    │  │ fallback │  │ Edge TTS │  │ word     │  │ Ken Burns│
│ scraping │  │ persona  │  │ frames   │  │ say      │  │ level    │  │ + music  │
│          │  │ + hooks  │  │          │  │          │  │ ASS+SRT  │  │ ducking  │
└──────────┘  └──────────┘  └──────────┘  └──────────┘  └──────────┘  └──────────┘

Stage by stage:

Research — Searches DuckDuckGo (and optionally scrapes source URLs) for live facts. Every name, number, and claim in the final script traces back to this research. This is the anti hallucination gate: the LLM is instructed to use only facts from research data, never its training knowledge.

Script — An LLM (your choice of provider) writes a 60 to 90 second voiceover script using the niche profile's tone, pacing rules, and hook patterns. The profile tells the LLM things like "open with a question, not a statement" for tech niches or "open with a shocking statistic" for finance niches. Output includes the script, b roll image prompts, thumbnail prompt, and platform metadata for YouTube/TikTok/Instagram/X.

Visuals — Generates 3 b roll frames via Gemini Imagen, then auto crops them to 9:16 portrait. If image generation fails, the pipeline uses simple fallback frames so assembly can still complete. The niche profile shapes the visual vocabulary: a fitness niche generates gym and movement imagery, a science niche generates diagrams and lab visuals.

Voice — Text to speech via your configured provider: Edge TTS (free, cross platform, 300+ voices, recommended default), ElevenLabs (premium, most natural), or macOS say (fallback). The niche profile suggests voice characteristics (pace, energy, tone) but the final voice selection is yours.

Captions — Whisper generates word level timestamps. The pipeline produces both ASS (burned in with word by word yellow highlight) and SRT (uploaded to YouTube for closed captions). Caption styling follows the niche profile: bold energetic fonts for gaming, clean minimal for tech, warm handwritten feel for lifestyle.

Assemble — ffmpeg combines animated b roll (Ken Burns zoom/pan effects), voiceover, burned in captions, and background music with automatic voice ducking. Music selection is mood matched to the niche profile.

Upload — Publishes to YouTube (private by default) with title, description, tags, SRT captions, and AI generated thumbnail. TikTok and Instagram export coming in v3.1.

Niche Intelligence

This is what makes Verticals different from every other AI video tool.

A niche profile is a YAML file that tells the pipeline how to think about content for a specific audience. It shapes every stage without requiring any prompt engineering from you.

# niches/tech.yaml
name: tech
display_name: "Tech & AI News"

script:
  tone: "informed, slightly opinionated, conversational"
  pacing: "fast, dense with facts, no filler"
  hooks:
    - pattern: "contrarian_take"
      template: "Everyone is celebrating {topic}. Here's why that's a problem."
    - pattern: "breaking_news"
      template: "This just happened and nobody is talking about it."
    - pattern: "prediction"
      template: "{topic} changes everything. Here's what happens next."
    - pattern: "explainer"
      template: "Let me explain {topic} in 60 seconds because most people are getting this wrong."
    - pattern: "comparison"
      template: "{thing_a} vs {thing_b}. One of these wins and it's not even close."
  cta_variants:
    - "Follow for daily tech breakdowns."
    - "Subscribe. I cover AI news nobody else is talking about."
    - "Drop a comment: do you agree?"
  word_count: "150 to 170"
  forbidden: ["like and subscribe", "smash that bell", "what's up guys"]

visuals:
  style: "clean, minimal, dark backgrounds, neon accents"
  mood: "futuristic, sleek, professional"
  subjects: ["circuit boards", "code on screens", "server rooms", "product shots", "data visualizations"]
  avoid: ["stock photo people smiling at laptops", "generic office", "clipart"]

voice:
  pace: "slightly fast, ~160 wpm"
  energy: "confident, authoritative but not robotic"
  suggested_voices:
    edge_tts: "en-US-GuyNeural"
    elevenlabs: "JBFqnCBsd6RMkjVDRZzb"

captions:
  highlight_color: "#00FF88"
  font_weight: "bold"
  position: "lower_third"

music:
  mood: "ambient electronic, subtle energy, no lyrics"
  energy: "medium"

thumbnail:
  style: "dark background, bold white/green text, product or face focus"
  text_position: "left_aligned"

15 built-in niches plus a general fallback: tech, gaming, finance, fitness, cooking, travel, true_crime, science, politics, entertainment, sports, fashion, education, motivation, comedy, and general.

Build your own by copying any profile and editing it. Drop the YAML in niches/ and reference it with --niche your_niche_name.

Quickstart

CLI

git clone https://github.com/rushindrasinha/youtube-shorts-pipeline.git
cd youtube-shorts-pipeline
pip install -r requirements.txt

python -m verticals run --topic "your topic" --niche tech

The old --news flag still works for backwards compatibility, but --topic is the recommended public API.

Hosted Version

Use verticals.gg if you want the workflow without local setup.

Examples

Two worked end-to-end examples with real commands and the exact output to expect at each stage live in examples/:

• Tech news Short, full pipeline — draft, produce, upload with the default providers
• Zero-cost draft with topic discovery — trending topics plus a local Ollama draft, no paid keys

CLI Commands

Full pipeline (topic to published Short)

python -m verticals run --topic "headline" --niche tech
python -m verticals run --topic "headline" --niche cooking --provider ollama
python -m verticals run --discover --niche gaming --auto-pick

Individual stages

python -m verticals draft --topic "headline" --niche tech
python -m verticals produce --draft <path> --lang en
python -m verticals upload --draft <path> --lang en
python -m verticals topics --niche tech --limit 20

Useful flags

--niche NAME         Niche profile (default: general)
--provider NAME      LLM provider: claude, gemini, openai, ollama, minimax (default: claude)
--voice NAME         TTS provider: edge, elevenlabs, minimax, 60db, say (default: edge)
--platform NAME      Draft target: shorts, reels, tiktok, all (default: shorts)
--lang CODE          Language: en, hi, es, pt, de, fr, ja, ko (default: en)
--dry-run            Draft only, skip produce and upload
--force              Redo all stages even if completed
--verbose            Debug logging

Provider Support

LLM (script generation)

Provider	Cost	Setup	Notes
Claude (Anthropic)	~$0.02/script	ANTHROPIC_API_KEY	Best quality. Default.
Gemini (Google)	Free tier available	GEMINI_API_KEY	Good quality, generous free tier.
GPT (OpenAI)	~$0.01/script	OPENAI_API_KEY	Solid alternative.
Ollama (local)	Free	Install Ollama + pull model	No API key needed. Quality varies by model.
Claude CLI	Free w/ Max sub	Install Claude Code	Uses Claude Max subscription, no API key. See Claude authentication.
MiniMax	Pay-as-you-go	MINIMAX_API_KEY	OpenAI-compatible API.

TTS (voiceover)

Provider	Cost	Setup	Notes
Edge TTS	Free	None	Recommended default. 300+ voices, cross platform.
ElevenLabs	~$0.05/video	ELEVENLABS_API_KEY	Most natural. Premium.
MiniMax	Pay-as-you-go	MINIMAX_API_KEY	Streaming TTS, English voices.
60db	Pay-as-you-go	SIXTYDB_API_KEY	Native Indic-language voices, low per-character cost. List voices: python -m verticals voices --provider 60db.
macOS say	Free	macOS only	Basic fallback.

Claude authentication

The claude provider works with any one of these, checked in order:

1. ANTHROPIC_API_KEY — a standard Anthropic API key from console.anthropic.com. This is not the same thing as a Claude Code OAuth token; the two are not interchangeable.
2. A logged-in Claude Code CLI (claude login with a Claude Max subscription). The pipeline shells out to claude -p and no API key is needed.
3. CLAUDE_CODE_OAUTH_TOKEN — a long-lived token from claude setup-token, with the claude CLI installed. Useful for CI and headless machines. The token is consumed by the CLI, not sent to the Anthropic API directly.

Visuals (b roll)

Provider	Cost	Setup	Notes
Gemini Imagen	Free tier available	GEMINI_API_KEY	Default image provider.
Fallback frames	Free	None	Solid-color fallback frames if image generation fails.

Upload

Platform	Status	Auth
YouTube	Stable	OAuth (setup wizard)
TikTok	v3.1	Coming soon
Instagram Reels	v3.1	Coming soon
X (Twitter)	v3.1	Coming soon

$0.00 Mode (completely free)

Yes, you can run this with zero API spend:

python -m verticals draft \
  --topic "your topic" \
  --niche tech \
  --provider ollama

This creates the script and metadata with a local LLM. Full video production still needs visuals, TTS, captions, and ffmpeg; Edge TTS is free, while Gemini visuals require GEMINI_API_KEY.

Configuration

All keys stored in ~/.verticals/config.json with 0600 permissions:

Variable	Required	Used By
ANTHROPIC_API_KEY	If using Claude API	Script generation
CLAUDE_CODE_OAUTH_TOKEN	If using Claude CLI headless	Script generation (via claude CLI)
GEMINI_API_KEY	If using Gemini visuals/LLM	B roll + thumbnails
OPENAI_API_KEY	If using GPT	Script generation
ELEVENLABS_API_KEY	If using ElevenLabs	Premium voiceover
MINIMAX_API_KEY	If using MiniMax	Script generation + voiceover
SIXTYDB_API_KEY	If using 60db	Voiceover

Environment variables override config file values.

Note on GEMINI_API_KEY: it must be an AI Studio key from aistudio.google.com/apikey. Vertex AI or service-account credentials fail with 403: Method doesn't allow unregistered callers. The same 403 appears when the key is simply not set in the environment the pipeline runs in (a common cause inside Docker — pass it with docker run -e GEMINI_API_KEY=...).

Topic Discovery

Discover trending topics from multiple sources, filtered by niche relevance:

python -m verticals topics --niche tech --limit 20

Source	Method	Auth	Niche Filtering
Reddit	.json API	None	Subreddit mapping per niche
RSS	feedparser	None	Configurable feeds per niche
Google Trends	pytrends	None	Geo + category filtering
Twitter/X	Public API	Optional	Keyword filtering
TikTok	Apify	Optional	Hashtag mapping
YouTube Trending	RSS/API	None	Category mapping
Hacker News	API	None	Tech/startup default

Configure per niche in your profile:

# In niches/tech.yaml
discovery:
  reddit: ["technology", "artificial", "MachineLearning", "singularity"]
  rss: ["https://hnrss.org/frontpage", "https://techcrunch.com/feed"]
  google_trends_category: "t"
  youtube_trending_category: "28"

Cost Per Video

Configuration	Cost
Premium (Claude + Gemini + ElevenLabs)	~$0.11
Budget (Gemini + Gemini + Edge TTS)	~$0.04
Draft-only local (Ollama)	$0.00
Voice-only free path (Edge TTS)	$0.00 for voice generation

Quality Limits

Verticals is built for repeatable short-form production, not for replacing an editor on story-led creative work.

It works best for:

• news explainers
• niche trend breakdowns
• list-style educational Shorts
• fast social experiments where volume matters

It is not ideal for:

• cinematic storytelling
• creator personality pieces
• videos that require taste, live footage, or strong art direction
• factual topics where the research source quality is weak

Use draft --dry-run or run --dry-run when testing a new niche. The most important human checkpoint is the draft: hook, factual claims, and whether the video has a real reason to exist.

Distribution Loop

The pipeline is only the production layer. The useful operating loop is:

• Pick a niche with existing short-form demand.
• Generate and review a small batch of Shorts.
• Publish consistently on the platforms where that niche already moves.
• Use one clear call to action: free resource, newsletter, waitlist, or product.
• Read performance weekly, then tighten the niche profile and hooks.

This repo lowers the cost of testing. It does not remove the need for niche selection, distribution taste, or a real conversion path.

Project Structure

verticals/
├── verticals/
│   ├── __main__.py            # CLI entry point
│   ├── config.py              # Keys, paths, setup wizard
│   ├── niche.py               # Niche profile loader
│   ├── llm.py                 # Claude / Gemini / GPT / Ollama
│   ├── research.py            # DuckDuckGo research gate
│   ├── draft.py               # Script generation with niche intelligence
│   ├── broll.py               # Gemini image generation + Ken Burns
│   ├── tts.py                 # ElevenLabs / Edge / say
│   ├── captions.py            # Whisper + ASS/SRT
│   ├── music.py               # Track selection + ducking
│   ├── assemble.py            # ffmpeg final assembly
│   ├── thumbnail.py           # Thumbnail generation + text overlay
│   ├── upload.py              # YouTube upload
│   ├── topics/                # Multi source topic engine
│   ├── state.py               # Resume capability
│   ├── retry.py               # Exponential backoff
│   └── log.py                 # Structured logging
├── niches/                    # 15 built in niche profiles
│   ├── tech.yaml
│   ├── gaming.yaml
│   ├── finance.yaml
│   ├── fitness.yaml
│   ├── cooking.yaml
│   ├── travel.yaml
│   ├── true_crime.yaml
│   ├── science.yaml
│   ├── politics.yaml
│   ├── entertainment.yaml
│   ├── sports.yaml
│   ├── fashion.yaml
│   ├── education.yaml
│   ├── motivation.yaml
│   ├── comedy.yaml
│   └── general.yaml           # Default fallback
├── tests/
├── scripts/
│   └── setup_youtube_oauth.py
├── references/
│   ├── setup.md
│   └── troubleshooting.md
├── pyproject.toml
└── requirements.txt

Testing

pip install -e ".[dev]"
python -m pytest tests/ -q

Security

All security measures from v2 carry forward, plus:

Credential storage: Config and tokens use 0600 permissions via atomic os.open(). API key handling: All providers send keys via headers, never URL parameters. Upload privacy: YouTube uploads default to private. Prompt injection: Research snippets truncated to 300 chars with boundary markers. LLM output fields are type checked before use. OAuth scopes: Minimum required scopes per platform. Niche profiles: YAML parsed with safe_load (no code execution). Dependency pinning: Compatible release bounds on all packages.

Roadmap

v3.0 (this release) Niche intelligence, multi provider LLM support, Edge TTS default, topic discovery, resumable stages, YouTube upload

v3.1 (planned) TikTok/Instagram/X upload, multi language niche profiles, A/B script variants (generate 2, pick better), scheduled batch production

v3.2 (planned) Analytics integration (which Shorts performed best), niche profile auto tuning based on performance data, series support (multi episode narrative arcs)

Later Web UI, Docker, Google Colab, additional visual providers, stock footage fallback

Built By

Dr Rushindra Sinha — MD, Stanford GSB, Full Stack Developer.

Built the first game server at 17 (went #1 globally, acquired before finishing med school). Co-founded Global Esports — South Asia's only Valorant Champions Tour Pacific franchise. Now building AI tools for creators and operators at aarees.com.

Follow: @irushi on X · @rushindrasinha on Instagram

---

More From This Stack

Product	What it does
verticals.gg	Hosted version of this pipeline — no setup, no terminal, just results
thumbnail.gg	AI thumbnail generation with deep niche intelligence and CTR optimization
aarees.com	The AI agent platform powering both products
Global Esports	South Asia's VCT Pacific franchise — where the esports niche profile was battle-tested

---

License

MIT