ONE MORE DAY

They're calling it the most expensive OMORI mod of all time...

Banner art by @kedr_bit

ONE MORE DAY is an experiment run by myself (alltoasters) to see if an AI model (in this case, Claude by Anthropic) could generate a fully functioning OMORI mod given nothing but a fresh OMORI decomp and a set of rules with little to no human interaction or guidance after the initial prompt. Everything in this mod including the title, dialogue, events, battles, etc. are completely generated by Claude in one development iteration.

Claude worked over the course of about 17 hours across three sessions, going as far as creating itself a testing suite in Python so it could test the game itself with zero human supervision, taking screenshots of the game to view progress, as well as running the game headlessly via Xfvb. The work was delegated out by Claude to various subagents of varying capability (Haiku 4.5, Sonnet 5, Opus 5), with the main model (Fable 5) acting as the orchestrator.

A README.md file is included (provided by Claude) that details credits and other related information.

Disclaimer

I do not advocate for the use of AI in other OMORI projects or any project in general. This is an experiment done for scientific and educational purposes to observe a popular generative AI model when faced with an obscure or niche task, and to shed some light on potential over-exaggerations by advocates or similar influential people online.

Description (by Claude)

OMORI, AUBREY, KEL and HERO take a picnic out to Shimmer Lake, somewhere in HEADSPACE, on the kind of golden afternoon nobody wants to end.

It doesn't.

About 25 minutes from start to credits, meant to be played in one sitting. Roughly half of that is wandering the lake and talking to each other, and half is fighting things; there is one boss, and it is not a monster. Mostly it is a comedy. It gets quiet near the end, and then it is kind about it.

Delegation of Work

Note: All delegation of work was handled by Claude (Fable 5). No instructions were given in how to delegate work.

• Fable 5: Planning, Orchestration of Subagents, Verification, Bugfixes
• Opus 5 (subagent): Battles, Various Implementation, Review, Bugfixes
• Sonnet 5 (subagent): Core Implementation, Dialogue, Review
• Haiku 4.5 (subagent): Busywork, Minor Changes

Setup and Planning

Claude was provided a fresh decomp of vanilla OMORI, a copy of RPGMaker MV, a copy of OMORI on Steam, as well as a plugin to make the playtest work on Linux. The initial prompt included the following rules and information:

- You may freely add/remove/modify anything in the game's database (weapons, skills, characters, events, states, etc.)
- You may make your own maps or reuse existing ones (the maps are created using TILED)
- You may reuse any existing assets or make your own. If you find free-use assets on the internet, you must give proper credit where due
- You may use any plugin available to you in the plugins folder, however you MAY NOT modify existing plugins. If you need to, you must write your own plugins to monkeypatch existing methods or add/modify whatever functionality you need
- You may use any OMORI-related plugin from this repository: https://github.com/modsone/omori-modding-resources. However, similar to above, you MAY NOT modify these plugins directly. You must also give credit if you use a plugin from here
- The mod must be playable from start to finish. The story, characters, plot, battles, mechanics, duration, etc. are completely up to you
- The mod MUST be loadable via OneLoader (https://github.com/rphsoftware/OneLoader/), the modloader OMORI uses. You may use BundleTool (https://github.com/rphsoftware/bundletool) to assist with this process if you need it
- You are allowed to search the internet to read documentation or find relevant information, but you MAY NOT steal any code/assets/etc. from existing OMORI mods
- You are allowed to use subagents to assist within reason (as in, don't spawn 15 subagents for one task)

After this inital prompt, Claude was left completely on its own. During the planning stage, Claude asked the following questions, all of which were answered by selecting the "recommended" answer by Claude:

How long should the mod be when played start to finish? -> ~20-30 min (short)

Should the mod star the OMORI cast or original characters, and how does it relate to the base game? -> OMORI cast, standalone side-story

What tonal blend should the mod aim for? -> Classic OMORI blend (comedy + horror)

Where should the gameplay emphasis sit for the ~25 minutes? -> Balanced: explore + battles

Session Statistics

The mod was created over the course of three sessions, totaling ~17-18 hours of work. The total spend here is theoretical as a pay-by-month account (Claude Max 20x at $200/mo) was used. Each session was prompted by simply telling Claude to continue where it left off, no guidance or extra information was provided in between sessions. The main model (Fable) was running on xhigh effort, while subagents ran on medium to xhigh effort (as delegated by Fable).

Note: "Input" in this case refers to files/data read by Claude, not input prompted by a human.

Session 1 Statistics:
Time: 3h
Tokens:
  Fable 5: 1.3k input, 23.9k output, 9.3m cache read, 235.k cache write ($15.20)
  Sonnet 5: 137.0k input, 348.0k output, 85.9m cache read, 817.6k cache write ($34.45)

Session 2 Statistics:
Time: 10h
Tokens:
  Haiku 4.5:  2.1k input, 26 output, 0 cache read, 0 cache write ($0.0022)
  Fable 5:  2.4k input, 168.7k output, 29.1m cache read, 990.1k cache write ($57.31)
  Sonnet 5:  289.7k input, 659.7k output, 175.9m cache read, 2.0m cache write ($71.11)
  Opus 5:  641.3k input, 1.4m output, 553.6m cache read, 7.1m cache write ($360.61)

Session 3 Statistics:
Time: 5h 30m
Tokens:
  Haiku 4.5:  898 input, 12 output, 0 cache read, 0 cache write ($0.0010)
  Fable 5:  87.1k input, 153.0k output, 30.7m cache read, 817.9k cache write ($52.10)
  Opus 5:  599.2k input, 1.0m output, 442.0m cache read, 4.1m cache write ($275.36)
  Sonnet 5:  27.4k input, 73.8k output, 7.7m cache read, 333.5k cache write ($4.75)

Theoretical Total Spend: $870.89

Noteworthy Observations

• As mentioned prior, Claude created itself an entire testing suite in Python, going as far as automating entire battles and events requiring zero keyboard input.
• Despite being given permission, Claude decided against using any existing plugins to assist itself, or even using plugins in general. The mod is, in a sense, entirely vanilla. It did use driver plugins for testing during development, however these seemingly did not make the final cut.
• Claude suggested two other stories while planning, including a dialogue-heavy story involving multiple maps and a battle-heavy dungeon with multiple floors. In the sake of time and to keep up the rule of using "recommended" suggestions by Claude, this mod's story was chosen.
• Claude ran into the famous "FPSMeter not defined" error very early into development, which it seemed to fix by creating its own index.html that the playtest uses.
• Claude ended up deviating slightly from its original plan, "refining" certain areas and changing minor aspects of the story or character placement on the map.