Writing with AI

1

Writing it for you, and editing what you wrote

"I used AI" covers two completely different things. At the Hoshi Shinichi Award, a Japanese science-fiction prize, three of the four winning entries turned out to have used generative AI. Around the same time, Rie Kudan won the Akutagawa Prize, Japan's most prominent literary award, and described ChatGPT as "a different kind of editor." The first is handing over the act of writing. The second is asking someone to read what you already wrote.

Software development settled this distinction a while ago. The AI developers reach for every day is less a machine that writes code in their place than one that reads the code already there and helps fix it. For people who write prose, that second kind of tool barely exists yet.

Even with an AI that reads every page, you are the one writing. Where the story goes, which sentence survives — those stay human decisions. What a reading AI changes is not that, but the quality of who you get to talk to about it.

Core is built to read what you wrote, not to write in your place.

2

The AI you paste into, and the AI that goes and reads

If you have ever taken a long manuscript to ChatGPT, the routine is familiar. Copy the text in, paste it, act on what comes back, paste the revision in again. Come back another day and you start over, pasting and re-explaining what the work is about.

This is not a matter of using the tool badly. It is structural: you have been cast as the one who does the pasting. A chat AI can only see what you hand it, and it answers without knowing about the chapters you left out.

An AI that goes and reads removes that whole routine. You write down what you want to ask, and it opens the chapters it needs. That sounds like a story about saving copy and paste, but what actually changes is the answer. A general observation from something that has never seen your work, and a note from something that has read it, are different objects.

There is also the worry that it will simply assert things about your work that aren't so. This is hallucination: an AI returning a confident, plausible error. Reading the manuscript directly makes that far less likely, because when you ask about chapter three, what gets consulted is the chapter three you actually wrote. Facts from outside the work, though — history, technical detail, anything you didn't write down — are a separate matter, and still need checking the ordinary way.

3

What it costs to read the whole work

Reading an entire manuscript ought to come at a price. There is a ceiling on how much you can hand a model at once, and the more you hand it, the slower and more expensive the reply gets. Built naively, this stops being usable the moment the manuscript runs to a hundred thousand words.

Core manages it by not sending the full text. What the model normally receives is a summary of the work, and the full text of a chapter is fetched only when that chapter is actually needed. Being able to ask one question about consistency across a long book, and having the wait and the cost stay reasonable, do not normally coexist. Making them coexist is what this part of the design is for.

The examples here are drawn from novels, but nothing changes for a dissertation or a work of nonfiction.

4

Vertical Japanese as a condition

Core was built first for Japanese, which brings a requirement most writing tools never face: the text runs top to bottom, right to left, and the editor has to stay that way while you write in it.

Almost no editor handles this well. Tools designed around English don't attempt it, and the ones that do often break down at typesetting or export. Core keeps the composition vertical from drafting through to a finished book or ebook, and holds a long work together in chapters while doing it.

If you write in English, none of this is a feature you will use. It is worth knowing anyway, because it says something about what the editor underneath is built to survive.

5

Proving who wrote it

"I wrote this myself and it got flagged as AI." You see this constantly now, on coursework, on applications, from people with nothing to hide.

The conversation tends to drift toward making detection more accurate. But there is a ceiling on what you can learn from a finished manuscript alone: a piece of polished prose does not carry a record of how it came to be.

So Core doesn't try to judge the finished text. It records the process of writing as it happens. Rather than inferring the truth afterward, you hold the evidence that you were writing from the beginning.

Ending the "did AI write this?" debate with provenance / About the writing certificate

6

Why this tool exists

All of the above started as my own frustration.

In programming, tools like Claude Code read the files themselves, take in the whole codebase, and act on it. Once you have worked that way, using a chat AI on a novel starts to feel thin. It only follows as far as you have explained, and every session begins again with something that knows nothing about the work.

I wanted an AI that would go and read the manuscript, take in the shape of the whole thing, and think alongside me — an editor. I didn't build it because it looked like a market. I built it because I wanted to use it. Core is developed by one person.

7

How it differs from other writing tools

There is nothing unusual anymore about a writing app with AI in it. Word has one. Notion has one. What differs is how the AI relates to your work, and it sorts into three kinds: a general-purpose chat bolted to the side, a generator that produces prose from a prompt, and something that reads the whole work and talks with you about it.

Core is the third. Sections 1 through 3 are the reasons for choosing it.

Comparisons with individual tools have their own pages.

ChatGPT Scrivener Ulysses Notion Word + Copilot