I Had AI Write My Game Script, and "Three Hundred Years" Showed Up 308 Times

Not enough words? Just add a few more AI writers. Sure enough, the word count got there, and so did a pile of repeated dialogue.

In the script for the game I'm making, the phrase "three hundred years" appears 308 times.

308 is the total across every file of the first draft after it was expanded to 8 chapters. It really is a key term in the worldbuilding, but even so, that's way too many. After one round of cuts, a single route still had it 62–100 times.

The script was written by several AI writers working side by side. When the word count fell short, I added more writers, and the same dialogue doubled right along with it.

# There are 17 writers

It's a dark fantasy game that branches based on your choices, currently a prologue plus 8 chapters.

The first version was 4 chapters, with 9 AI writers splitting the work by chapter and by character, each writing their own part of the script in parallel. A coordinating AI was in charge of deciding when each event triggers.

But when it was done and I counted, it was too short. One playthrough came to 23,000–29,000 characters, which I estimated at under an hour, short of the 2-hour target.

The next version expanded to 8 chapters with 17 writers. First, 9 of them wrote the common story, then 8 wrote the individual routes, flashbacks, and epilogues. Each one only touched the files they were responsible for, and at the end the coordinator stitched them together.

By the count on October 9, after revisions, one playthrough came to 58,000–73,000 characters, about 2.5 times the first version. That's counting the source-language script, not the Japanese translation.

# The same kind of exchange 8 times in Chapter 1

I read it from start to finish myself, and my takeaway was: "The core story beats all look the same. The overall writing quality dropped. So this is what happens when you write for volume."

Going through Chapter 1, 8 of its 17 scenes followed the same pattern: the protagonist notices something, the terminal's voice explains it away with a mechanical reason, and the protagonist answers back with something poetic. The explanation changes from "air currents" to "resonance in the ventilation ducts," but the shape of the exchange is identical.

The characters' signature lines and gestures kept repeating too. "Knocking 3 times on the door" showed up 160 times across the whole draft, and "three hundred years" hit 45 times in a single file. Scenes kept ending the same way, too: the terminal falls silent, or someone gazes at something for a long time.

# Rules for the rewrite

Before rewriting, I settled one thing first: "What has changed by the end of this scene?" Either someone learns something new, or someone makes a decision. Sentences that repeat explanations got cut: 5–9 lines for a normal scene, and at most 14 even for a big one.

The same pattern of exchange could appear at most 2 times per writer's script. For signature lines, I set rules for who can use them and how many times. That said, lines that get paid off in later scenes can't be cut. Search before deleting, keep scene numbers and branch conditions untouched, and aim for a 20–30% reduction.

In the first revision pass, the 17 writers' scripts went from 116,711 characters in total to 85,162, a 27% cut. That's the total across every branch, which is a different thing from the length of a single playthrough.

For example, "...For the first time in three hundred years, someone is cooking." became "...It's been a long time since anyone cooked, huh." In the same revision pass, I also redrew the images that didn't match the lines.

# Having AI read it 3 times

Alongside the rewrite, I had AI readers go through it from start to finish 3 times.

The approach: replay the events day by day under the same conditions as the game, lay out the script in the order it would actually be displayed, and then have the AI read that. The main-route script used for the third read was about 90,000 characters, title included.

What the first read caught was scenes that were written but never actually appear in the game:

• Defeating a boss immediately bumps the chapter number by +1, so climax scenes that were supposed to play "within this chapter" never do. On one route, about 10 scenes, including the climactic song, vanished entirely
• Chapter 7 got stuck and couldn't be entered at all

Feedback on the writing itself was concentrated in the second read. Too many hints before the big reveal, so by the time the truth actually came out, all that was left was confirmation. In the final farewell scene, the beat where everyone's name gets called out one by one happened 4 times.

I capped hints at 3 and kept the name-calling to just once. "Three hundred years" dropped to 11–20 times per route and 75 times across all files. That said, I kept lines where the number itself matters, like the one about the expired cookies, where "three hundred years" is the punchline.

The third read produced 57 comments. I fixed 46 the same day, partially fixed 7, and kept 4 as-is after judging them to be by design. There was also 1 where the AI misread something, and I left a correction in the review log.

# Counting again after the rewrite

Besides the readers' feedback, I also checked things with a program that counts through the script.

The check on October 9: references pointing to nonexistent scenes, and scenes with no way in, were both at 0, and missing Japanese and English translations were also at 0. But having a way in doesn't guarantee a scene actually shows up in the game. The chapter-number mismatch was only discovered once the script was laid out in its actual playback order.

There's also a check that combines the main choices into 90 combinations. Previously, 27 combinations got the wrong route; after changing the rules, that's down to 0. But that's only within those 90 combinations. I haven't tried every possible way to play.

# Pitfalls in translation and voice

Lines are translated into Japanese and English one at a time via a dictionary. There's a pitfall here too: if the source line is identical, then even if you want to translate it differently depending on the speaker, say "Sis," "Sister," or "Elder Sister," the dictionary merges them into a single entry. The first read caught 76 of these collisions.

Voice is tied to the lines too, so rewritten lines need their audio regenerated. Only the lines that changed get redone automatically.

The voice check at this stage flags suspicious clips by length, speaking rate, pauses, and pitch. In the end I still have to listen myself, but I haven't finished listening to all the voice lines regenerated after the rewrite.

# What's still left

I haven't actually played through it from start to finish and timed it. Based on the character count, the estimate is 2.5–3.5 hours. The target is 2 hours, so it's probably too long. The next step is to go check that for myself.