The Build Step As Data Compiler: 2.4MB In, Thirteen Fields Out

The author of a random Pokemon generator put the `@pkmn/dex` package in devDependencies, ran a single script against it at build time, and emitted plain static JSON the app imports like any other module ^[1][2]. The site makes zero calls to PokeAPI at runtime, and the 2.4MB dex package never reaches a browser ^[3][4].

What ships instead is five files: `pokedex.json` with 1,025 species carrying exactly 13 fields each, an 18x18 type effectiveness matrix, 685 moves, 310 abilities, and a per-species legal move list ^{[5][6][7][8][9]}. The reasoning for not fetching is structural rather than aesthetic. PokeAPI answers questions one at a time, and a generator does not ask questions one at a time ^[10]: the moment a user toggles "Gen 3 only, no Legendaries, starters excluded" and presses the button, the app needs the entire filtered pool in memory immediately ^[11]. Per-click fetching buys spinners, rate-limit exposure, and freeloading on a free public API ^[12].

The more durable payoff is that the messy definitional work lands in one file. The raw dex includes battle-only formes, cosmetic variants and fan-designed CAP entries, so the script hardcodes the exclusion policy: skip anything with a non-empty `forme`, skip `num <= 0`, skip anything flagged `isNonstandard` as CAP, LGPE, Custom, Unobtainable or Future ^[13][14]. Some categories do not exist upstream at all. There is no `isStarter` field, so the script carries a hardcoded set of 27 starter names and stamps a boolean onto each match ^[15]. Legendary and mythical come from tags ^[16]. Generation is not stored either; it is derived from National Dex number ranges, 151 and below being Gen 1 ^[17]. The client reads pre-computed booleans and fields and reasons about none of it, and changing the pool is a one-file diff plus a rebuild ^[18][19].

That is the part that generalizes. Anything you currently resolve at runtime because a rule is ambiguous - eligibility, tiering, region mapping - is a candidate for compilation, and the win is not only bytes. It is that the ambiguity gets a single owner and a version history instead of being re-litigated in three components.

Compilation does not make everything small. After trimming, the learnsets file still weighs 870KB, down from a 3.1MB raw chunk ^[20][21], roughly a 72 percent reduction and still an order of magnitude heavier than everything around it ^[1]. Because exactly one feature needs it, it is the only file with no static import: it loads via `await import()` the first time someone generates a moveset, and the chunk is cached afterward ^[22][23]. Moves and abilities sit in separate lib files rather than one shared battle-data module, so each route carries only its own weight ^[24]. Sprite URLs are not stored at all, since the filename is the National Dex number and the URL can be assembled at runtime ^[25].

Two things to watch if you copy this. The 13-field contract is the pressure point: the moment the UI needs a fourteenth field, the schema, the generator script and the rebuild are all in the loop, which is cheap in a hobby repo and less cheap behind a release process. And the only described refresh path is re-running the build ^[19], so any dataset that changes on someone else's schedule needs a scheduled rebuild and a way to notice when the upstream package moves. The pattern trades freshness for determinism. That is usually the right trade, but it should be a decision rather than a side effect.