>> the editorial engine
The assistant knows the journal now. Not because it was told what to write, but because it has read what has already been written.
The core is a vector store: nineteen markdown files, each high-signal. Identity documents, design language, tone, selected posts, an anti-style guide, fifty short concrete observations. The model retrieves from these on each query. The quality of the retrieval determines the quality of the response.
A vector store is not a database. It stores meaning, not records. The same query asked two different ways returns the same material because the system matches by semantic proximity, not exact string. This matters for something as inherently imprecise as "write in the voice of this journal."
The hard part was not the technical setup. It was deciding what belonged in the store and what did not. Code patterns and file structure were excluded — they can be derived by reading the repo. Git history was excluded — that belongs in commit messages. What remained were the things that would not survive in any other form: the philosophy, the tone, the recurring preoccupations, the lines already written that define what a new line should feel like.
An observations file was added last: fifty short concrete notes from experience. Not abstractions. Not principles. Observations. The model needed friction against actuality — something to pull it away from manifesto language toward the grounded specific. More realia improved the output faster than any additional instruction.
The result behaves more like an editorial engine than an assistant. It does not generate text from scratch. It operates inside a defined aesthetic and checks each response against it. When it works, the output is indistinguishable from something written in this register. When it fails, it sounds like an AI.
The failure modes are instructive. The model hallucinates files that do not exist, reaches for noir atmosphere when restraint is called for, abstracts when a concrete observation would do more. These are not model failures. They are knowledge base failures. Better content improves the system faster than a larger model.
The store is not finished. It is a working document. What gets added to the journal goes into the knowledge base. The assistant gets incrementally more accurate as the writing accumulates. That is the right direction.