I Turned Generative AI Into My Writing Team
In Episode 1 of this series I wrote about the concept and the outline. This time, I want to talk about how I actually built this novel — using generative AI as a team while I moved into a directing role.
My role wasn’t “writer.” It was “director.”
My main job is running a company and managing development projects. So this time too, it never occurred to me to do all the writing myself, by hand. I treated AI as a team of subordinates, and took on the structure, character design, worldbuilding, giving instructions, and directing the revisions — almost exactly the approach I use in my regular work.
Rather than creating from zero, I took the position of designing “what to have it write, under what constraints, and how.” In fact, I spent a considerable amount of time building out the plot structure, the characters, and the worldbuilding.
It really was close to “boss and subordinate”
The feeling that clicked into place most, once I actually tried this, was that it worked exactly like the relationship between a boss and a subordinate.
I gave a great many instructions, but unless those instructions were systematized and rule-based, the AI couldn’t handle them well. “Just make it feel good somehow” doesn’t move it anywhere. So in the end, I read through the entire text myself and built each instruction as a set together with its reasoning — “this is like this, so do it this way.” It was almost identical to how I manage development projects normally.
And it was a long-haul task. Just like a development project with no end in sight. The period ran roughly 25 days, with nearly 200 git commits, and the main text ran past 300,000 characters. The revision requests (issues) I filed, split across GitHub and a separately managed ledger, easily exceeded 500. Partway through, I gave up on aligning direction through minor fixes and went ahead with a full-scale rewrite. Even I have to admit — it was a proper “development project.” (laughs)
Splitting roles: claude.ai and Claude Code
The tools broke down into two main ones.
- claude.ai … creative judgment calls, judging whether the writing was good or bad, an advisor role.
- Claude Code … file-level editing. Batch replacements across chapters, cross-cutting fixes, version control with git.
The manuscript — the prologue plus about 30 chapters — was managed entirely as Markdown files. In the past I used to hand-name versions like v1/v2/v3…, but switching to git made history management dramatically easier.
The instruction document as a “blueprint”
What made it possible to have the AI write a long novel with consistent quality was preparing an instruction document (a handoff document) in advance.
HANDOFF_GUIDE.md… established rules, each character’s manner of speech, formation charts, notes on historical accuracy, the order in which tactics are revealed, and a chapter-by-chapter progress tracker, all on one sheet.historical_elements_to_weave_in.md… historical elements woven into the story (memories of 1966, faith in the long ball, Hillsborough and the Taylor Report, each club’s golden era, etc.) organized with priority levels.
In short, I put every premise into files so the AI could work from the same premises even when starting a new session. Without this, the AI happily strays outside the setting.
What’s interesting is that this instruction document grew by a line every time I worked on it. For instance, the procedure for the “patrol check” — the process of reviewing the text — gained a “check this too, next time” every time I caught an oversight, and before I knew it, it had become a whole section of the procedure manual. This wasn’t “leaving it to the AI”; it felt more like I was jointly growing the checking process together with the AI. Every time something failed, one more rule got added — no different from a human editing environment.
Historical fact as “constraint,” not “seasoning”
Another thing I valued was treating the ages, club histories, and historical events of real players not as flavor but as hard structural constraints. The ban on English clubs from European competition after Hillsborough, the league’s balance of power, each club’s golden era — these “facts that cannot be moved” formed the foundation on which I laid the fictional story of Wandle. Build it on top of a lie and the world instantly turns cheap, so I didn’t compromise here.
To that end, I kept running plain, merciless checks over and over. Calculating characters’ ages (how old someone would be in a given year), eliminating anachronistic expressions (not letting characters speak concepts or words that didn’t yet exist in that era), and even verifying details of daily life and culture — like not having characters toast with beer mugs. Checking, one by one: “as 1980s Britain, could this really have happened?” It’s tedious work, but neglect it and the world instantly rings false.
Incidentally, quite a few of these historical facts and small details came not just from knowledge I already had, but from things that surfaced in conversation with the AI, which I then seasoned and adopted myself. Thanks to that, the book is dense with the full-on obsessiveness of a tactics nerd and a football history nerd. Honestly, there were moments I felt I’d exceeded, by quite a margin, the limits of what I could write as a single author. I hope that comes through as something enjoyable, too.
Assigning “symbols” to tactics, like squad numbers
For Wandle’s original tactics, I assigned names and sequential numbers using Greek letters (α through θ / First through Eighth) and managed them as a single system. As the story progresses, James’s “bag of tricks” keeps growing, and that bag of tricks was what I was cataloguing as a setting.
The numbers I’d settled on once got reshuffled many times over as I kept writing. On one occasion I carried out a “major surgery” of renumbering and reclassification (reorganizing things like: δ = the long throw demoted to a set piece, ε = the false nine, ζ = isolation, and so on). It’s backstage work invisible to readers as mere numbers, but whether this kind of setting backbone exists or not completely changes how consistent the tactical depictions feel.
A concrete example: how I searched for a “great player who never broke through”
I think this is the clearest example, so I’ll lay out the actual exchange.
James, the protagonist, discovers still-unknown talent, and I worked together with the AI to search history for “players to be discovered.” The first names the AI raised were real Englishmen who, at the time, were languishing in non-league football. Iain Dowie, Eddie McGoldrick… Historically accurate, but honestly, neither I nor Japanese readers could quite connect with them.
So I gave a new order. “I want bigger names. Players at the level Japanese football fans would recognize from watching the World Cup, the Premier League, the Champions League.”
And then they came pouring out: late bloomers who were essentially unknown until their early twenties and then blossomed spectacularly in later years, one after another. Each came back with details attached — “where they were and what they were doing in the early 1990s,” “when they had their breakthrough.” Almost none of this was knowledge I already had, and I selected from it and worked it into the story with my own seasoning. I won’t name names here, but every single one is a face any Japanese football fan would surely have gasped at, at least once, watching the World Cup, the Premier League, or the Champions League. Who ends up wearing the Wandle shirt, I’ll leave as a treat for the main story.
And this is where I ran a single thread that gives the work its character. James isn’t gathering talent just because he already knows the “results” — he sees still-unrecognized talent with his own eyes and wins them over himself. And then he “sets them free to the world, even a single season earlier than history did” — that’s the story of passion I distilled it down to.
(As an aside — I agonized quite a bit with the AI over the question of “once the time comes when they’d historically have shone, should James send them off to their real clubs instead of hoarding them at Wandle, or keep them at Wandle to the very end of the story?” Even I have to admit — that was “a fairly complicated” thing to be consulting an AI about.)
…Reading this far, you might think “so AI made everything easy, huh.” But next time, I’ll write about how it wasn’t actually that simple. The theme: “Editing the Japanese text turned out to be only slightly easier than just writing it myself.”
Originally published in Japanese at https://clazytech.com/2026/06/1642/. Translated with LLM assistance and reviewed before publication.