Stop Prompting. Start Designing. Building a Creative AI Pipeline
How I built an automated generative AI pipeline that creates a fictional museum, and what it taught me about prompt engineering, creative constraints, and AI product design.
Everyone is building generative AI apps. That’s not a criticism. It’s the current reality of the tech field. But there’s a category of problem that gets less attention: creative coherence in AI generated content. This is an exciting new frontier, and the one I wanted to explore.
A Museum That Builds Itself
The Ravensfield Collection is a fully automated content pipeline that generates a fictional museum, one artifact at a time. Each run produces a piece of flash fiction about a mysterious artifact, complete with quotes from fictional experts and an AI-generated image of the piece. The pipeline runs on a daily cron job.
The fiction lives in a specific, carefully defined world: the Ravensfield Collection of the Uncanny, narrated by an unnamed Chief Curator with an erudite, detached voice that masks a deep existential dread. Every story follows a strict five-paragraph structure. Every artifact has provenance, a medium, a year, and an artist. The imagery is generated from the same description the text uses. The whole thing coheres, not because AI is magical, but because it was designed to.
Design the Rules, Not the Output
The most common mistake I see in AI-generated creative work is treating AI as the author. Give it a vague instruction, get a vague output back, call it a day. The results tend to be technically acceptable and creatively inert, the prose equivalent of a stock photo.
The Ravensfield Collection is built on a different premise: AI needs constraints to produce coherent creative output, not freedom. The more precisely you define its rules, its voice, its structure, and its limits, the more surprising and alive the output feels within that world. You’re no longer the one with hands on the clay. You’re the one who designed the mold.
That’s the real work of prompt engineering: not crafting the perfect instruction, but architecting the system around it.
That shift is worth understanding before you start. Frustration sets in quickly when you treat an AI creative project like painting a canvas or putting words on a page. It’s neither of those things. The creative flow has moved upstream, into the design of rules and systems, and it can be a lot of fun and deeply satisfying on its own terms.
Under the Hood
The pipeline runs in five steps:
- A Madlibs-style generator assembles a randomized creative brief from curated word lists: an art movement, a visual technique, a story subgenre, a protagonist, a theme, and a fate.
- Anthropic’s Claude receives the brief and produces a detailed conservation-style description of a fictional artifact.
- That description fans out in parallel: Claude writes the full story and metadata while Leonardo.ai generates the image, both working from the same source material.
- Everything gets stored atomically in a database — either all of it or none of it. No partial entries, ever.
- A final vision-consistency check has Claude compare the generated image against the written story to catch any contradictions before the piece goes live.
It runs across multiple AI models and produces one coherent output, every day, without intervention.
A Tool, Not a Mind
No tool is without limits.
The system can generate thousands of unique combinations, but only from a vocabulary a human assembled. Quality works the same way: vague prompts produce vague stories. And Claude doesn’t “understand” the Chief Curator’s dread any more than a piano understands Chopin. It executes within a shaped space.
That’s also what makes it worth doing well: creative AI systems amplify your design decisions, which means the quality is entirely in your hands.
The Bigger Picture
The next five articles each tackle one of the core decisions this project demanded: pipeline architecture, controlled randomness, system prompt design, image generation, and multimodal feedback loops. I decided to have some fun and create a weird fiction pocket universe, but the pattern holds for any team building an AI product.
The full repository is public if you want to dig in directly. Otherwise, let’s start from the beginning.