content-machine

An autonomous, single-operator content studio: ideation → production (blog + video) → repurposing → scheduled auto-publish, across 3 niches and 7 platforms, at 200+ published pieces per month.

Personal system, not a packaged product. It runs my content operation end-to-end; published here to share the architecture. Expect hard-coded paths, opinionated conventions, and no install script.

What it does

flowchart LR
    A["<b>Ideate</b><br/>idea scorer<br/>32-file virality KB<br/>90-day no-repeat tracker"]
    B["<b>Generate</b><br/>Claude CLI subprocesses<br/>per-task model routing<br/>150 Python scripts"]
    C["<b>Render</b><br/>storyboard → HyperFrames<br/>→ FFmpeg composite<br/>word-synced captions"]
    D["<b>Schedule</b><br/>APScheduler daemon<br/>SQLite queue → launchd"]
    E["<b>Learn</b><br/>analytics ingest<br/>UTM star attribution"]
    A --> B --> C --> D --> E
    E -. "feeds next week's scoring" .-> A

Loading

One blog run fans out into every derivative — platform posts, carousel, deck, reel brief, thumbnail, worksheet — and the analytics loop closes back onto idea scoring, so what performed last week changes what gets written next week.

Architecture highlights

• 150 Python scripts, one orchestration model — all active code in v1/scripts/; entry points: run_blog_pipeline.py (one idempotent run produces every derivative of a blog: platform posts, social images, carousel, deck, reel brief, thumbnail, worksheet), run_video_pipeline.py (raw footage → trim → storyboard → HyperFrames per-beat compositions → FFmpeg composite), run_voiceover_week.py (audio-first lane: VO → B-roll montage long-form + auto portrait Shorts), scheduler.py (APScheduler daemon).
• Per-task LLM model routing — a single routing table (v1/scripts/lib/niche_config.py) assigns the right Claude model per task: top tier only for quality-critical hero content, workhorse tier for generation, small models strictly for mechanical classification. No model IDs hard-coded at call sites.
• Claude CLI as the LLM runtime — pipeline calls run claude -p subprocesses on subscription auth (no API key), with disk caching, timeout budgets per stage, and backoff only on transient failures.
• Virality as configuration — a 32-file knowledge base injects hooks, CTAs, and guardrails into every generator, routed per niche. Editing markdown changes generation behavior; prompts are versioned in code, never copy-pasted.
• Honesty guardrail — build-in-public posts about my open-source tools are generated with an explicit never-overclaim constraint per project.
• Closed loop — UTM-tagged repo links + analytics collection tie published content directly to GitHub star deltas, which feed the next week's idea scoring.

Stack

Python · Claude CLI (model-routed) · HyperFrames · FFmpeg · APScheduler · SQLite · launchd · Meta Graph API · YouTube (4 channels) · Medium · LinkedIn

Related tools

• autopilot-jobhunt — autonomous job-search agent · PyPI · MCP
• inbox-to-action — one-pass agentic inbox triage (PyPI, MCP)

Author

Tarun Gupta — tarunlnmiit.github.io · Case study of this system