ai-shorts-pipeline

For anyone whose calendar is full but whose audience still expects to hear from them.

Daily content has become a tax on every operator, founder, and builder trying to grow a name. Most days you don't have an hour to record, edit, caption, and post. ai-shorts-pipeline automates the production so you can keep showing up — even on the days you can't. The point isn't to make videos faster for fun. The point is consistency for the people whose real job isn't creating content but who still need to.

It's a Claude Code skill. You give it a topic and a couple of source links. It hands back a finished 60-second vertical video, a voiceover in your cloned voice, and per-platform captions for YouTube Shorts, Instagram Reels, LinkedIn, and X. ~5–7 minutes per video on a normal laptop.

Under the hood, the skill orchestrates these moving parts in a single render pass:

• ElevenLabs for voiceover (your cloned voice), with word-level timestamp alignment
• Vertex Veo for cinematic B-roll (text-to-video and Imagen→Veo image-to-video)
• Three.js for custom 3D scenes — particle systems, brand strikethroughs, data-viz
• Remotion for everything else: composition, timing, encoding, frame extraction
• HeyGen (optional) for presenter-led avatar segments at the hook and CTA

Plus a local-only marketing HQ dashboard (multi-page Express app) for queue management, posting cadence, and per-platform publishing kits with copy-buttons and char-limit enforcement.

Battle-tested on @AIinBusiness — daily videos rendered by this exact pipeline.

This repo is honest about what it does and doesn't do.

What you get from cloning this repo:

• A skill protocol Claude follows for every video — editing rules, scene grammar, pacing curves, anti-patterns, mandatory QA workflow. The decisions that took us months to settle, condensed into one file.
• A working marketing HQ dashboard you can run on localhost. Queue, calendar, roadmap, performance, ideas, hooks, playbook.
• The render pipeline: Remotion compositions, ~40 scene types, the Three.js primitives library, gen-audio + Veo + HeyGen helpers, the orchestrator that turns a topic into a spec, and 12 synthesized SFX. Clone, set keys, render.

What you bring yourself:

• Your voice clone settings (ElevenLabs stability, similarity, style — these are tuned to a specific voice and tone; yours will be different)
• Your topic-research process (which AI news matters this week, who you're writing for, the contrarian angle — that's editorial judgment, not code)
• Your hooks library (you build it as you ship and learn)
• Your performance benchmarks (your channel's 3-second retention floor, your save-to-share ratio — yours, not mine)
• The take. Always the take.

The plumbing is here. The framework gives you ~70% of the craftsmanship of framing. The 30% — the voice you tune, the hooks you build, the takes you bring — is what makes the channel yours instead of mine. That's a feature, not a gap.

For the full skill protocol — editing rules, scene grammar, anti-patterns, QA workflow — see SKILL.md.

flowchart LR
    A[Topic + sources] --> B[Claude Code]
    B -->|drafts narration| C[Script]
    C -->|word-level alignment| D[ElevenLabs voiceover]
    C -->|scene-by-scene spec| E[Remotion composition]
    E -->|backdrop + 3D| F[Three.js scenes]
    E -->|cinematic B-roll| G[Vertex Veo]
    E -->|optional avatar| H[HeyGen]
    D --> I[Render pass]
    E --> I
    F --> I
    G --> I
    H --> I
    I --> J[1080×1920 MP4]
    I --> K[Per-platform publishing kit]
    J --> L[Marketing HQ dashboard]
    K --> L
    L -->|YouTube · Instagram · LinkedIn · X| M[Posted]

    style B fill:#f59e0b,stroke:#92400e,color:#1c1917
    style I fill:#f59e0b,stroke:#92400e,color:#1c1917
    style L fill:#1f2937,stroke:#f59e0b,color:#f5f5f4
    style M fill:#10b981,stroke:#064e3b,color:#1c1917

The skill is the orchestration layer. Claude reads SKILL.md, picks the right scenes for each beat, generates the spec, and runs the render. The dashboard is where you actually live — queueing, scheduling, copying captions, tracking what shipped.

git clone https://github.com/Khizergenfox/ai-shorts-pipeline.git
cd ai-shorts-pipeline
npm install
cp .env.example .env
# Dashboard runs without any keys. Render pipeline (v0.2) needs ELEVENLABS_API_KEY etc.

npm run dashboard      # opens http://localhost:5173

The dashboard is the marketing HQ — queue, calendar, roadmap, performance, ideas, hooks, playbook. All local, all yours, no internet exposure.

Multi-page Express app, no build step. Each section is its own URL so it's easy to extend (drop a route file + a view file).

The Today page knows what day it is and tells you what mode you're in (post / engage / ideate / rest), the one CTA, and the posting playbook for today. Posting windows per platform, engagement velocity rules, the whole thing scoped to right now.

Drag-to-reorder list of rendered videos waiting to ship. Top of list = next post day. Click any video to see per-platform captions with copy buttons, live char counters on the X tweet, and a mark-posted form that promotes it to the Uploaded section.

Personal-brand milestones with Day-N or specific-date triggers, anchored to a launch date you set. Add milestones from the form, see them surface on the Today page when their trigger fires.

Reusable opening-line templates with tags (contrarian, honest-reversal, from-the-trenches, etc.). One click to copy. You build your own library as you ship and learn what hooks land.

Marketing strategy notes and an append-only "Lessons learned" log. Every shipped video gets a one-line entry: what worked, what didn't, what to do differently. Over time this becomes the most valuable file in the repo.

A single example frame from a recent rendered short, showing the news-article scene type doing its job (real publication style, real headline, scene-driven typography):

The skill enforces "literal before metaphorical" — real news article mockups before any abstract 3D explanation. Editing rules and scene grammar are in SKILL.md.

• SKILL.md — the skill protocol Claude reads to render videos: editing rules, scene grammar, pacing curves, the question-to-visual decision rubric, the QA workflow, the HeyGen avatar pipeline
• dashboard/ — local marketing operations (multi-page Express, no build step, drag-to-reorder queue, copy-to-clipboard captions, append-only "Lessons learned" log)
• Render pipeline — Remotion entry + master composition + ~40 scene type implementations
• 3D toolkit — 18 Three.js primitives in src/effects/three/ (BarRace, MathField, NodeNetwork, Stat3DExtrude, LogoOrbit, BloomHalo, CameraRig, RevenueTimeChart, and more) that you compose into custom story_3d variants for your topics
• 9 production scripts — render-full, gen-audio, gen-veo-broll-cinematic, gen-veo-from-imagen, gen-heygen-clip, poll-heygen-clips, render-3d, test-vertex-auth, plus the shared Vertex client
• All 12 synthesized SFX in public/sfx/ so renders have sound on first run (thump, riser, click, whoosh, fire-whoosh, glitch, alert-beep, success-ding, terminal-enter, notification, chart-rise, typing)
• Example spec + example narration — specs/example-news.json and scripts/example.txt so you can see the JSON shape and render end-to-end
• Voice tuning is env-var driven — ELEVENLABS_STABILITY / _SIMILARITY_BOOST / _STYLE / _USE_SPEAKER_BOOST. Defaults are ElevenLabs' neutral starting point (0.5 / 0.75 / 0.0). Tune to your cloned voice.

Clone, set keys, drop a script in scripts/example.txt, then:

node scripts/gen-audio.mjs scripts/example.txt example
node scripts/render-full.mjs example

Your first video lands in out/example-v1.mp4.

• v0.1 — Skill spec, marketing HQ dashboard, example data, architecture diagram
• v0.2 — Render pipeline (Remotion compositions, render scripts, gen-audio, Veo b-roll, HeyGen avatar helpers, all SFX)
• v0.3 — YouTube + LinkedIn + X uploaders integrated into the dashboard
• v0.4 — Auto-fetch performance metrics from each platform's API
• v0.5 — Voice cloning setup walkthrough

Khizer Hussain · @AIinBusiness on YouTube

Also building GenFox.AI. This is the side project I run in public, for people who want to ship daily AI content without a video team.

MIT. Use it, fork it, ship your own channel.

If you ship something built on this, tag me — happy to boost it.