I am currently building an ai story app called Inky.
Most people think building an ai story app is about writing a clever system prompt. It isn't. It's about architecting a system that can handle state, memory, and narrative logic across multiple turns without hallucinating the plot into a corner. Most 'AI apps' are just thin wrappers around a single API call. They are brittle, expensive to run, and offer no real moat.
Inky is different. It is a product of my studio’s operating model: AI as the team, not just a feature. I am building this in public to show the difference between a prompt-engineered toy and an architected system.
The Artifact: Why Inky Exists
Inky is a digital product designed to help users co-create long-form narratives. The goal isn't to have the AI write a book for you—that’s boring. The goal is to build a system that acts as a creative partner, maintaining the 'world state' while the user drives the intent.
When you are building an ai story app, the core challenge is consistency. If a character has blue eyes in chapter one, they cannot have green eyes in chapter four. If the story takes place in a desert, it shouldn't start raining because the LLM forgot the setting. Solving this requires more than a large context window; it requires a structured data layer that sits between the user and the model.
Moving Beyond the Wrapper
I learned the hard way that relying on a single 'God Prompt' is a recipe for failure. As the story grows, the prompt becomes bloated, the model loses focus, and the cost per token skyrockets.
Instead, I use agentic engineering. Inky is powered by VERA, the custom orchestration layer I built for my studio. VERA treats different LLMs and specialized prompts as workers in a factory line.
Agentic Engineering in Practice
In the Inky architecture, the work is split across several specialized agents:
- The Archivist: This agent doesn't write prose. Its only job is to extract facts from the current turn and update the 'World Bible' (a structured JSON object in the database).
- The Director: This agent analyzes the user's input against the current narrative arc and decides the tone and direction of the next beat.
- The Weaver: This is the prose engine. It takes the instructions from the Director and the facts from the Archivist to generate the actual text.
By decoupling these concerns, I can use a cheaper, faster model for the Archivist and a high-reasoning model for the Weaver. This makes the system more durable and significantly more profitable.
