AI Content Machine

An autonomous content studio publishing 200+ pieces a month across 7 platforms for 3 distinct niches — 150 Python scripts on the Claude CLI with per-task model routing, a HyperFrames video pipeline, and a scheduler daemon that fires posts while the operator sleeps.

01Why it exists

Consistent multi-platform publishing is a full-time job: research, drafting, design, video, scheduling, repurposing, analytics. Doing it for three niches (Data Science · Life · Poetry) manually is impossible. The machine treats content as an engineering problem: every stage scripted, model-routed, scheduled, and measured — with hard rules like a 90-day no-repeat tracker and an honesty guardrail baked into generation.

02How it works

1. Ideate & scoreAn idea scorer ranks candidates weekly; a 32-file virality knowledge base injects hooks, CTAs, and guardrails into every generator, routed per niche. An annual tracker enforces no repeated angles within 90 days.
2. Generate with model routingAll pipeline calls run through the Claude CLI on subscription OAuth. A single routing table assigns the right model per task: Fable 5 for hero blogs, Sonnet 5 as the workhorse, Opus 4.8 for storyboards, Haiku for mechanical classification only.
3. One-run blog pipelineOne idempotent command turns a blog into every derivative: platform posts, social images, carousel, deck, reel brief, thumbnail, and a gated worksheet — then stages them for publishing.
4. V2 video pipelineRaw footage → trim → storyboard → HyperFrames per-beat compositions → FFmpeg composite, with word-synced captions, B-roll montage, and screen-zoom effects. A voiceover-first lane builds long-form + portrait Shorts from audio alone.
5. Auto-publish & learnAn APScheduler daemon fires pending posts from SQLite across YouTube (4 channels), Medium, LinkedIn, and Instagram/Threads/Facebook via the Meta Graph API. Analytics collect star deltas via UTM attribution and feed the next week's ideas.

03Results

• 200+ published pieces/month sustained across 7 platforms and 3 niches
• 150 Python scripts orchestrated end-to-end — human touchpoints reduced to recording, a ~10-minute approval pass, and community replies
• Per-task LLM model routing from one config: quality where it pays, cheap models for mechanical work, zero API key for pipeline calls
• Closed loop: UTM star attribution ties content directly to GitHub stars on the open-source tools it promotes

04Stack