AtomicAuthor

Frederick Lowe, Oct 2023 (active)

Atomic Author dashboard showing article generation pipeline
Atomic Author dashboard showing article generation pipeline

AtomicAuthor is an AI-powered publishing platform I built for orchestrating multi-stage content generation at scale. It coordinates LLMs (Claude, GPT), image generation (Stability AI), and publishes directly to WordPress—all through a human-in-the-loop editorial workflow.

The Problem

Traditional content management treats AI as a bolt-on: generate text, copy-paste, manually edit, manually publish. This works for one article. It doesn't work for hundreds.

I wanted a system that could: - Queue and prioritize article generation across multiple topics - Break generation into discrete, retryable stages - Track costs and token usage across providers - Let editors rate and filter content before publication - Publish automatically when quality thresholds are met

The Architecture

Multi-Stage Pipeline

Articles flow through a configurable six-stage pipeline. Upon completion and validation of each stage, the MongoDB document containing the article is updated:

  1. Base Article — Core content generation via LLM
  2. Metadata — Categories, tags, dietary info (for recipe content)
  3. Image Prompt — Generate description for visual assets
  4. Image Generation — Create hero images via Stability AI
  5. External Links — Shopping integrations (Instacart for recipes)
  6. Finalization — Validation and WordPress preparation

Each stage is independently and retryable, up to the stage's configured maxProcessingAttempts.

Failed stages don't block the pipeline—the orchestrator automatically retries any failed generation from its point of failure.

Multi-Tenant Design

The platform supports multiple "accounts" (publishing projects): - Atomic — Base implementation with core logic - Chopped — Food/recipe publishing with Instacart integration

Each account has its own prompts, credentials, WordPress site, and custom stage logic. The wrapper pattern keeps core logic centralized while allowing per-tenant customization.

Consensus Rating System

A Bayesian rating system for quality control: - All content starts at 2.5 stars (neutral baseline) - Editors rate 1-5 stars, updating weighted averages - Auto-publish threshold: 3.2+ stars - Auto-reroll threshold: Below 1.8 stars

This lets a distributed editorial team manage thousands of articles through simple star ratings.

Technology Stack

Results

The platform has generated and published more than 10,000 articles, managed by a distributed editorial team.