I didn't build WriteAssist because I wanted AI to write a book for me.
I built it because of a book I'd already given up on.
The thing I shelved
Sixteen hours of my mother. I recorded her in pieces over one winter, her voice and my questions, the stories coming out in no particular order, and I had every minute of it transcribed. Then I went looking for someone to turn the pile into a book. This was a few years ago. AI existed, but it was early, and I wasn't using it for anything that mattered yet. So I did the normal thing. I found a writer on Fiverr who would listen to the whole thing, take notes, and draft the book. Seven hundred dollars. I put down the deposit.
Two weeks later she came back and said seven hundred wasn't going to cover it. It was closer to three thousand. I understood, it was a mountain of work, but three thousand wasn't a number I had for this. So I thanked her, she returned my deposit, and we parted on good terms. The transcripts went on a shelf. And that was that.
It wasn't a pause. I had quietly decided the book was never going to happen, and I'd made my peace with it. That decision has a name, and at the time I didn't have the name for it.
Why it mattered enough to hurt
It wasn't that the book didn't matter. It mattered more than almost anything else I'd thought about making, and that was exactly the problem. Sixteen hours is a lot of hours. And people don't tell their lives in order. She'd start a story about her own mother, jump to last year, circle back forty minutes later with the part that actually mattered. Beautiful to listen to. A nightmare to turn into chapters. I had the raw material and no path from the pile to the book, and once the Fiverr plan fell through I stopped looking for one.
Years later, the door reopened
A couple of years went by and AI got good. Really good. One day I thought about the transcripts again, still sitting where I'd left them, and for the first time the wall I'd hit looked like it might have a door in it. So I went back to the thing I'd shelved and did what a lot of people did the first time these tools got good. I dropped a chunk of transcript in and asked for a chapter.
And it was good. Genuinely good. It found a thread I hadn't seen, opened on a scene with her at the kitchen table, carried her phrasing into the prose. I read it twice. I've built real software with AI and I lean on it constantly, and even I got starry-eyed for a second. I went to bed that night planning the cover.
The crack
I asked for the next chapters. And the cracks showed up the way they always do with long AI work. Quietly, in the details only you would catch.
Chapter five had my grandmother dying in 1991. She didn't. The model stated it like a fact it had always known, no hedging, no flag, just a confident sentence about a death that never happened. Other things slipped too. People walked into scenes as if you'd already been introduced, relatives named like you were supposed to know them, no setup, no moment where the reader actually meets them. The model could write. What it could not do was keep my family straight.
That's the part that stuck with me. Not the magic chapter. The wrong year on my grandmother. Because it told me exactly what I was dealing with. The model could write a better sentence than I could on a tired night. It could not hold the whole living thing in its head at once. Twenty chapters, a real timeline, real people who have to stay consistent across all of it. It would write any single piece beautifully and lose the thread between them.
What I actually needed
The easy conclusion was "AI isn't good enough for this yet." Wait for a bigger context window, try again next year. I sat with that and decided it was wrong.
The model was not the problem. The context window was. And a bigger window wouldn't have saved me, because I'd still be handing it the whole sixteen-hour pile and hoping it held everything in mind at once. What was missing wasn't capacity. It was structure. I didn't need a better prompt. I needed a system around the prompt. Keep the window fixed and let the scaffolding do the remembering.
I've seen what happens without the system
I know what that structure looks like, because it's how enterprise infrastructure has worked for as long as I've been in it. You don't let one person do everything. The engineer who configures the firewalls is not the one walking Sally through changing the batteries in her mouse. The database administrator does not reboot the firewall. The help desk tech does not reinstall the domain controller. Each role has a lane, a defined scope, and the access that lane needs and nothing more.
In small shops, one person does do all of it. The same hands touch the firewall, the database, and Sally's mouse. Those people are not bad at their jobs. Usually they're heroic. But there is only so much one head can hold, and we have all seen how those environments go. Things get missed. Nobody is watching the part that quietly broke. Capacity is finite, and when you pretend it isn't, it finds you.
That is the same failure as one giant prompt. One prompt doing everything is the overloaded admin who touches all of it and slowly loses the thread. The fix in infrastructure is separation of duties: bounded roles, least privilege, a clean handoff between them. The fix for AI is the same shape. Many small agents, each with one job, each holding only what that job needs.
The build
WriteAssist is what came out of that. It's built for Claude Code, and the shape of it is the shape I'd give anything I had to actually trust.
Instead of one giant prompt, there's a plan for each chapter before a word of prose gets written. Instead of hoping the model remembers the facts, there's a compendium, one source of truth for the timeline and the people, and every task reads it first. That's where my grandmother's real history lives now, so a chapter can't quietly invent a new version of it.
Instead of asking the model nicely not to do the thing I can't stand, there's a hook at the harness level that physically blocks it. Mine bans the em dash. Policy says don't. A hook says can't. The point isn't the em dash. The point is that a rule I care about gets enforced by the tool, not requested in a prompt the model can talk itself out of on the next pass.
The reviewers can't touch what they review. There's a panel of them, each looking at one thing, continuity, voice, pacing, the rules, and every one is locked to read-only. A reviewer that can quietly "fix" the problem it's supposed to flag isn't a reviewer. Same reason the auditor doesn't sign the checks. And each revision runs in its own isolated branch, so a bad pass never corrupts the good draft and I can see exactly what changed.
None of it is exotic. The writer never has to hold the whole book in its head, because neither does a human writer. It's the boring discipline of running things that aren't allowed to fall over, pointed at a memoir instead of a data center.
The part worth keeping
The wrong year on my grandmother taught me something bigger than a writing tool.
The model is the easy part to upgrade. Every few months it gets better on its own and I don't lift a finger. What doesn't upgrade for free is the system around it. The plan. The source of truth. The rules you refuse to leave to chance. The decision about what each piece is allowed to touch. That part is judgment, and it's the part that decides whether sixteen hours of your mother's voice becomes a book or stays a folder. The system around the model matters more than the model. A weaker model inside a stronger system beats a stronger model with no system at all, and a bigger model with no system would just have gotten my grandmother wrong faster.
What AI actually replaced
AI didn't replace a writer in my life. I'm not a writer. What it replaced was my resignation, the quiet decision I'd already made that this book was never going to happen. The transcripts are off the shelf. The book is moving, slowly, the way the real ones do. But it's moving, and it's moving because I stopped hunting for the perfect prompt and built the thing that should have been there the whole time.
So if you're sure you don't have a project for AI, look again. You might not be missing a project. You might be sitting on a resignation, some good thing you quietly decided was never going to happen because it was too big for one head. Find it. Then write the plan, not the prompt. Add one hook, not one more rule you'll keep having to repeat. Give the work a source of truth and make everything read it first.
WriteAssist is on GitHub, MIT licensed, free to clone and open in Claude Code.