ASOBIMON

Asobimon was only my second Playdate project, and I built several early systems for quick iteration. During the closed beta, the problems with that approach became clear. The most demanding menus dropped to roughly 15 FPS on hardware, half of Playdate’s 30 FPS target.

The menu system had grown into a single large grid implementation, with too much UI work happening up front. I separated it into focused views, moved shared behavior into a reusable grid manager, and distributed module loading across frames. The refactor brought those menus back to 30 FPS.

I applied the same approach across the wider game: distributing startup work across frames, releasing Expedition rooms outside the current room neighborhood, running garbage collection deliberately after large transitions, and reducing unnecessary sprite and rendering work. On Playdate, small inefficiencies become visible quickly.

When I added trainer battles, I needed each NPC to carry a repeatable party with specific levels, genders, natures, abilities, genetic values, and dedication values. I solved that with Asobicodes: fixed-width 40-digit strings that contain everything needed to reconstruct a monster. Every field is zero-padded to a known width, so the game can read the code without separators. I now use the same format in saves, the Deck, procedural parties, special trades, network payloads, and round-trip tests.

Adding a move originally meant manually creating its normal cell, selected cell, and description image. Even a balance or text change could require rebuilding those assets. I replaced that workflow with runtime-generated move views built from shared assets, structured move data, and reusable UI components. I then built Python importers that validate monster and move CSVs and update their Lua definitions.

As the trainer logic got more complicated, playing battles one at a time became too slow to tell me much. I built a headless simulator that runs Trainer AI against Trainer AI, then added an in-game screen for setting up batches and reviewing the results. Tests can hold teams constant, compare profiles and skill tiers, rotate tactical scenarios, and keep each trainer’s decision history.

One of the first useful results was also a little embarrassing: in early batches, the skill-100 Trainer AI often lost to skill-1 Trainer AI opponents. The advanced trainer was evaluating setup, statuses, abilities, resources, switching, items, and future turns, while the simpler trainer mostly chose immediate damage. The move pool did not yet offer enough reliable payoffs for all that additional planning.

That showed me the problem was the available content, not just one Trainer AI bug. I added richer moves, statuses, abilities, and tactical options, then used the simulator to check whether they actually improved higher-skill play.

Instead of trying to save live Lua objects, the save system converts the current session into plain data that a later build can validate, migrate, and rebuild.

Monsters are stored as compact identity codes alongside mutable state such as health, mana, experience, status, provenance, online identity, and held items. Inventory entries become item IDs and quantities. World progress is reduced to scene, position, cutscenes, defeated trainers, and collected items. Expedition saves also retain the generated run state needed to reconstruct the current floor.

The main RPG and Expedition use separate save slots, but both pass through the same serialization and protection pipeline.

I had worked on multilingual iOS apps before Asobimon, so I began moving UI text, battle messages, and content descriptions behind localization lookups before I had translators. That preparation became useful when a Discord community member offered to translate the game into German. Spanish and Japanese began as proofs of concept; Italian and an expanded German translation followed later.

The runtime supports five language tables. Only the selected locale is loaded, and every lookup follows a fallback chain: the current language first, then English, then the original key or supplied source text. An incomplete translation therefore produces readable English instead of an empty label or broken menu.

Named placeholders such as {name}, {move}, and {damage} allow translators to change sentence order without changing the battle code that supplies those values.

Asobimon currently defines 109 species and allows the player to store up to 432 individual monsters across 24 Deck pages. On a Playdate with only 16 MB of RAM, those monsters could not all remain as complete animated game objects.

A battle-ready monster carries much more than an identity and a set of stats. It can reference a sprite sheet, animations, instantiated moves, an ability, audio, UI state, and temporary combat data. Keeping all of that for hundreds of Deck slots, plus the full Asobipedia catalog, would spend memory on systems most stored monsters were not using.

I separated the collection into three representations: shared species definitions, compact stored records, and fully hydrated runtime monsters. The game moves between them depending on what the player is doing.