Autonomous Content Pipeline

The problem

Producing regular long-form video content normally needs a writer, a voice, a video editor and someone managing publishing and performance — a real labor bottleneck for a solo operator or a small team wanting to run a content channel consistently.

What I built

A pipeline that takes a topic from research to a scheduled, published video with no manual editing. It researches content gaps, queues topics in a stateful job system, has Claude write the script, generates voiceover and AI imagery, assembles the video with burned captions, and publishes it to YouTube on a schedule through the official API — then feeds real performance data back into what gets made next.

Actual output

Real output from this pipeline. The channel it publishes to stays private — see the note below.

System flow

Technical depth

• Claude CLI is orchestrated as a scripted step in the pipeline, not used interactively — including typed error handling for auth and usage-limit failures.
• A stateful, YAML-driven job queue makes every stage resumable — a failed run picks back up rather than starting over.
• A statistical A/B testing framework compares different image-generation approaches against real view-through performance.
• Scheduled publishing goes through the official YouTube Data API with OAuth2, including metadata, scheduling and post-publish analytics tracking.
• A self-guarding scheduler (cron + lock files) prevents overlapping or duplicate runs.

Adapted to a second vertical

The same architecture was later adapted to a second content vertical, adding a retrieval-grounding layer that pulls cited excerpts from official government sources before script generation — so factual claims in a higher-accuracy-bar niche are traceable back to a primary source rather than generated from the model's own memory. That's the same pipeline, made more rigorous where the content demanded it.