I Ran Generative AI Like a Writing Team to Build My Novel
In part 1 of this series, I wrote about the inspiration and the plot outline. This time I want to talk about how I actually built this novel — using generative AI as a team while I handled the direction.
The role wasn’t “author.” It was “director.”
My actual job is running a company and managing development projects. So this time too, it never occurred to me to write the whole thing myself by hand. I treated the AI as a team of subordinates and took on the direction of structure, character design, worldbuilding, instructions, and revision — almost the same approach as my day job.
Rather than creating from zero, I positioned myself on the side that designs “what to write, under what constraints, and how to make it write.” In practice, I spent a considerable amount of time building the plot structure, the characters, and the world.
It really was close to “boss and subordinate”
Once I actually tried this, the feeling that fit best was that it closely resembled the working relationship between a boss and a subordinate.
I gave instruction after instruction, but unless those instructions were systematized into rules, the AI couldn’t handle them properly. “Just make it feel good somehow” doesn’t move it. So in the end, I’d read through the entire text and build instructions paired with reasoning — “this part is like this, so do it like that.” It was almost identical to how I manage development projects day to day.
And it was a long haul of work. Just like a development project with no end in sight. The period ran roughly 25 days, git commits numbered close to 200, and the main text exceeded 300,000 characters. The revision requests (issues) I filed while managing things split across GitHub, combined with a separately managed ledger, comfortably exceeded 500. At one point I gave up trying to align direction through minor fixes and went for a full rewrite. Quite the “development project,” if I say so myself. (laughs)
Division of labor: claude.ai and Claude Code
The tools fell into two main categories.
- claude.ai … creative judgment calls, assessing the quality of the prose, acting as a sounding board.
- Claude Code … file-level editing. Batch replacements across chapters, cross-cutting fixes, version control with git.
The manuscript — a prologue plus about 30 chapters — was managed entirely as Markdown files. In the past I’d manually name files v1/v2/v3 and so on, but switching to git made history management far easier all at once.
The handoff document as a “blueprint”
What proved effective in getting the AI to write a long work with consistent quality was creating a handoff document first.
HANDOFF_GUIDE.md… a single sheet consolidating established rules, each character’s speech patterns, formation tables, historical-accuracy notes, the order in which tactics get revealed, and a chapter-by-chapter progress tracker.historical_elements_to_weave_in.md… historical elements woven into the story (memories of 1966, the cult of 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 that a new session could start from the same footing. Without this, the AI would casually stray from the established setup.
What’s interesting is that this handoff document grew by one line with every work session. Take the steps for the “sweep check” that reviews the prose: every time an oversight slipped through, a line got added saying “check this too, next time,” and before I knew it, it had become a whole section of the procedure manual. It wasn’t so much that I delegated to the AI — it felt closer to co-developing the checking procedure itself together with the AI. Every failure adding one more rule — that’s no different from a human editorial workflow.
Historical fact as “constraint,” not “seasoning”
Another thing I prioritized was treating the ages and club histories of real players, and historical events, as hard structural constraints rather than flavor. European competition bans after Hillsborough, the balance of power among leagues, each club’s golden era — on top of these “immovable facts,” I layered the fictional story of Wandle. Build it on top of falsehoods and the whole world turns cheap instantly, so I wouldn’t budge on this.
To that end, I kept running relentless, unglamorous checks. Calculating characters’ ages (how old someone would be in a given year), eliminating anachronisms (never letting a character say a concept or phrase that didn’t exist yet in that era), and even verification down to the level of daily life and culture — like not letting anyone toast with a beer mug when it wasn’t period-appropriate. Pinning down, one by one, “would this really be plausible as 1980s Britain?” It’s unglamorous work, but neglect it and the whole world instantly rings false.
Incidentally, quite a few of these historical facts and bits of trivia weren’t just things I already knew — a fair amount came out of dialogue with the AI, which I then seasoned myself and adopted. Thanks to that, the book is dense with tactics-nerd and football-history-nerd material, fully cranked up. Honestly, I even had the sense that I’d gone well beyond the limits of what I could have written as a single author. I hope readers enjoy that aspect too.
Tactics got “symbols,” not squad numbers
For Wandle’s original tactics, I assigned names and sequential numbers using Greek letters (α through θ, the first through eighth), managing them as a single system. As the story progresses, James’s “bag of tricks” keeps growing, and I had that bag of tricks laid out in full as part of the setup.
The numbering I’d settled on got reshuffled repeatedly as I kept writing. On one occasion I carried out what amounted to major surgery — renumbering and reclassifying everything (reorganizing things so that δ, the long throw-in, gets downgraded to a set piece, ε becomes the false nine, ζ becomes isolation, and so on). It’s backstage work readers never see as numbers, but whether or not this kind of structural backbone exists makes a huge difference to the consistency of the tactical depictions.
A concrete example: how I searched for a “smoldering talent”
This is probably the clearest example, so let me write out the actual exchange.
The protagonist James discovers still-unknown talent, and I worked with the AI to search historical fact for the “player to be discovered.” The first names the AI came up with were real English players who were languishing in the non-league at the time — Iain Dowie, Eddie McGoldrick, and so on. Historically accurate, but honestly, neither I nor Japanese readers would find that resonant.
So I reissued the request: “I want names with more name recognition. Players at the level Japanese football fans would recognize from seeing them at the World Cup, the Premier League, or the Champions League.”
Then they just kept coming. One after another, late bloomers who were essentially unknown until their early twenties and then broke out spectacularly in later years. Each one came back with details down to “where they were and what they were doing in the early 1990s” and “when they broke through.” Most of this was knowledge I didn’t already have, and I selected from it and worked it into the story, seasoned to taste. I won’t name names here, but every one of them is someone Japanese football fans would surely have gasped at least once watching the World Cup, the Premier League, or the Champions League. Who ends up wearing a Wandle shirt is something I’ll save for the book itself.
And this is where I ran a single thread through to give the work its identity. James doesn’t collect players just because he already knows the “results.” He spots talent that hasn’t yet emerged into the world with his own eyes, and persuades them. Then he “sets them free to the world even one season earlier than in real history” — I built that into a story of passion.
(As an aside: I agonized quite a bit with the AI over “once the time comes for them to shine in real history, should they be released from Wandle to their actual club, or kept at Wandle through to the end of the story?” I was having, if I may say so, a genuinely “fairly complicated” conversation with it.)
…Reading this far, you might think “so AI did everything for you and made it easy.” But next time I’ll write about how it wasn’t 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.