Ten Days, 2.6 Billion Tokens: How I Shipped an R15 Game to itch.io

"Honestly, I wouldn't pay 1200 for it myself. But 1200 marked down to 900? That I'd buy."

If I ranked this month's projects by tokens burned, first place wouldn't be Live2D. It'd be a game called NEURAL FLESH.

I opened the first session on September 27 and it went public on itch.io on October 7. In between: 22 Claude Code sessions, roughly 2.6 billion tokens in total. The final "package + publish" session alone ate 398 million.

This post isn't about how I made the game. It's about how it went from "I want to put this on itch" to "you can actually buy it." Because finishing the game is only half the job. Shipping it is a whole separate hell.

As for what the game actually is (dark fantasy, a prologue plus 8 chapters, branching story, tactics battles on a 9×7 grid), I've already covered that in the earlier posts on the game, the script, and the tactics system, so I won't repeat it here.

# Fast-forwarding through the first nine days

Just the first line I typed at the start of each session tells you most of what those nine days looked like:

• 9/27: "Take the concept from @../electron-flesh but build our game with threejs"
• 9/29: "Can you lay out the choice branches for me in html? Overall it feels way too linear right now, there's not much room to play around"
• 9/29: "Honestly 20–30k characters per playthrough is too short, the shared main route is only about 10k"
• 10/2: "Feedback so far: there's no tutorial once you get into the tactics battles"
• 10/3: "Right now headpats look like you're ripping her scalp off"
• 10/3: "The tactics battles aren't fun enough right now... I think most people will just hit auto in the late game"

Four chapters became eight. I sent out a whole row of Sonnet writers to draft each route in parallel, then sent reviewer agents to read through the transcripts line by line, about 90,000 characters per route. One route came back from revision cut from 9,100 characters down to 6,600, because the writer had used the same signature move, "three knocks," 14 times. After the edit: zero.

The Japanese voice acting was generated with TTS running locally, 6,447 lines in total. By the end I'd given up on doing anything else and just kept asking "progress?"

That night, five of my messages in a row in the transcript were "progress?"

# "I want to go on itch, I'm changing strategy"

A little after 1 a.m. on October 7, I decided to skip Steam: sell the Japanese version on DLsite and the English version on itch.

The AI first checked the rules (butler allows 2 GB per upload, the web uploader 1 GB), changed the game so the language on first launch depends on the storefront, and queued up four builds: itch full version, itch demo, DLsite, and DLsite demo.

Halfway through, it noticed the Mac build was missing the Electron/Chromium license files (the Windows build had them), added them to the packaging script, and rebuilt. Then it opened the DLsite build to confirm it launched in Japanese the first time, and the itch Mac build to confirm it launched in English.

Great work. And then I said:

I think let's not package yet

It stopped. Then I said:

delete the steam release

So it pulled out the Steam packaging and upload steps and deleted the two 1.2 GB Steam builds along with them.

I was the one flip-flopping. It just did what it was told.

# Store assets: an 8 KB cover image

Next up was promo material: a trailer, Shorts, the itch cover, YouTube thumbnails, and promo images for X, one set each in English and Japanese.

The approach was to record gameplay directly from a hidden Electron window, script six self-playing shots (prologue, the sanctuary headpat, the transplant surgery, 3D exploration, two tactics battles), then composite captions and layouts with canvas.

A few of the small gotchas along the way were pretty representative:

• The sanctuary shot recorded a brand-new game: before reloading, the game wrote its empty save back to disk, overwriting the save the AI had slipped in
• The cover PNGs were only 8–30 KB: it captured frame 0, where the fade-in animation hadn't started yet and the background opacity was 0. In other words, it exported an almost fully transparent image
• The Shorts cover was 3.6 MB: over YouTube's 2 MB limit for custom thumbnails, so everything got a separate JPG export
• At one point it told me the DLsite version had 480 CGs, then corrected itself: it was 495. DLsite's cap is 500, so it squeaked in just under

None of these are deep bugs, but every single one is a case of "looks finished, actually isn't."

# itch: the price field said $1,007.99

Around noon that same day, I dropped a URL:

https://itch.io/game/new ↗ fill this out for me

The AI got stuck at first on itch's security check and login page, because logging in means entering a password, and that's something it's not allowed to do for me. I asked, "can't you control the itch browser? turn on debug mode," and in the end we switched to my own Chrome where I was already logged in, and it took over filling in the form from there.

And then:

The price field got filled in as $1,007.99: the original $1.00 wasn't fully selected.

Thankfully it went back and checked. Separately, the short description came out at 121 characters, one over the limit.

Saved as draft, added screenshots, filled in the metadata, downloaded butler 15.32.0, and pushed four folders totaling about 3.5 GB. The Windows build was 984 MB compressed, Mac 953 MB.

While it was at it, it also themed the itch page: a chapel background, a pixel-font title, a red buy button. Along the way it hit another snag: itch's editor is controlled by React, so setting values programmatically doesn't register. It has to actually click and actually type.

# Pricing: would I buy this myself?

While the upload was running, I asked a question that actually mattered a lot:

dlsite feels like 1200 is about right, or should we lower everything? what do you think about these two launch prices and the content. personally I wouldn't spend 1200 on it, but original price 1200 discounted to 900 I would buy

The AI's initial suggestion was US$12.99 on itch and ¥1,980 on DLsite, both based on typical indie game pricing. The problem is that this isn't a typical indie game. It's a game made with AI. The CGs are generated, the voices are synthesized, the script was written by a row of writer agents. Players are going to have a different number in their heads. None of the AI's suggestions were prices for an AI game.

So the price kept getting pushed down: itch went from $12.99 to US$7.99, and DLsite from ¥1,980 to ¥1,100 before tax (¥1,210 with tax). We'd originally agreed on 25% off for launch, but when it came time to fill in the forms, both went straight to 50% off: 14 days on DLsite, and the itch sale price became $3.99.

(Half of $7.99 on itch would be $4.00 because of rounding, but the page ended up showing $3.99. That's the kind of detail I probably wouldn't have noticed myself.)

# DLsite: the last button was left for me

DLsite also got filled out in my Chrome. I registered the circle myself, and it handled the product submission: genre set to ロールプレイング, age rating R15, main image plus 7 sample images uploaded (the uploader only takes one file at a time, so they went up one by one).

Along the way there was a very Japanese-website bug: in the keyword field, the second keyword had its first character eaten, so 「マルチエンディング」 turned into 「ルチエンディング」.

At the end, it stopped on the "final confirmation" page:

The DLsite product submission is all filled in. I've stopped on the "final confirmation" page; all that's left is for you to press 「登録する」 to submit it for review.

Over on itch, it went public at 13:51. It even went out of its way to fetch the page again while logged out, to confirm that a regular visitor could see the price, the Buy Now button, and both free demos.

# The very last thing: too promo-y

After launch I had it write a devlog ("can you make it shorter"), and then a tweet:

feels too promo-y, can you add some personal emotion like in the devlog

mention that it's r15 lol, make it more casual

AI can fill out the forms for you, but it can't write your feelings. That line is one I specifically asked it to add.

# What I learned

1. Shipping is no less work than making the game. Four builds, two platforms, two sets of assets in English and Japanese, three cover sizes, and every one of them has its own limits and pitfalls
2. When it "looks done," look again. The 8 KB cover, the $1,007.99 price, 480 vs. 495 CGs: every one of those was caught only by going back to check
3. Passwords, submit for review, go public: keep those three buttons for yourself. It annoyed me when the AI stopped at the login page, but looking back, those were always supposed to be my buttons to press
4. Price it with your own wallet. "Would I buy this?" is more direct than any market analysis. The market rates the AI compares against aren't necessarily the market for a game like yours

Ten days and 2.6 billion tokens, in exchange for one public itch page.

Looking back, what burned the most tokens wasn't writing code. It was the back-and-forth: chapter counts changed again and again, Steam on, Steam off, calling a halt halfway through packaging. AI can execute every decision really fast, but I still have to think the decision through myself first. If I ship another one, I'll lock in the storefronts and pricing on day one, instead of changing strategy on the very last day.