Games as Recipes: What AI Can (and Can't) Build

The one-paragraph answer

A video game is a recipe of about 45 reusable ingredients drawn from five kitchens: art, audio, narrative, code and polish. AI has *near-solved* a specific cluster of those ingredients — 2D imagery, voice, music, sound effects, prose, flavour text, boilerplate systems code — and has *barely moved* on another cluster — real-time netcode, combat feel, level design, and the iterative "is it fun?" tuning loop. A genre's AI-buildability is therefore decided almost entirely by which kitchen its core ingredients come from.

Games that are mostly *content + 2D art + writing + simple state* — visual novels, interactive fiction, point-and-click, hidden-object, clickers, card games — are now overwhelmingly automatable by a single non-specialist. Games whose identity *is* real-time feel and synchronised state — shooters, fighting games, MOBAs, souls-likes, battle royales, big open worlds — remain expert, multi-discipline craft where AI saves labour and none of the hard parts. Everything else sits on a smooth ramp between.

One note before any number: the AI-build difficulty scores here are our own editorial assessment of what today's tools can do — a judgement, not a measurement. The demand figures beside them are from the MythicLemon Games Catalogue, our own compiled dataset of roughly 82,000 released games. We keep the two clearly apart throughout.

A game is a recipe of 45 ingredients

We scored each of the 45 ingredients on a 1–5 scale of how much of the *heavy lifting* a generalist can hand to AI tools today: 1 = AI does it near end-to-end (prompt, curate, drop in); 3 = a strong assistant where the human owns the structure; 5 = AI barely moves the needle, and the outcome is expert, real-time craft. Then each genre is written as a *recipe* — a weighted list of ingredients, weight = how central that ingredient is to the experience — and its score is the centrality-weighted mean.

Group the ingredients by discipline and a clean pattern falls out. Audio (avg 2.2) and art (avg 2.6) are the most automated kitchens — voice, music, sound effects, backgrounds, textures and concept art all sit at 1–2. Narrative (2.8) is split: bulk writing is easy, but level, puzzle and encounter design are not. Code (3.0) scaffolds fast but tunes slowly. And polish/feel (3.75) is the hardest kitchen of all — game feel and balance are where the human craft concentrates.

The one-line version: *generation of artefacts* (pictures, sounds, words, scaffolds) is largely solved. *Synthesis under real-time, fairness and "feel" constraints* is not. Everything AI is good at, it's good at because the artefact can be judged in isolation; everything it's bad at requires judging the artefact in motion, against a player.

The genres, easiest to hardest

Roll the recipes up and you get a single ranking from *a generalist with AI tools can ship this* to *expert, multi-discipline craft*. The full table is below; four bands fall out naturally.

AI-buildable (under 2.4): visual novel (1.80), interactive fiction (1.83), point-and-click (1.94), hidden object (2.07), clicker (2.23), farming sim (2.38). Every core ingredient is in AI's strong zone; the remaining human job is *taste*.

AI-assisted (2.4–3.0): walking sims, card games, JRPGs, management and colony sims, 2D platformers. Systems-light, content-heavy — one skilled human directs and AI fills in. The JRPG is the tell: a writing/art/music genre wearing a turn-based-combat costume, which is why it scores far easier than its "hard RPG" reputation.

Systems craft (3.0–3.6): metroidvania, tower defense, city builder, horror, puzzle, deckbuilding, roguelike, survival. Their *assets* are cheap to generate, but their identity rests on level design, balance or performance — the things AI assists rather than authors.

Expert only (over 3.6): action RPG, racing, MOBA, immersive sim, souls-like, open world, FPS, fighting, battle royale. These stack 3D characters, animation, combat, environment art, performance and game feel — AI shaves the edges of each; none of the cores collapse. Fighting games are the extreme: only 17% of the catalogue is recent, a near-frozen, expertise-gated niche.

Demand, by contrast, is spread right across that difficulty range. Some of the biggest audiences — shooters, open worlds, survival — sit in the hardest tiers, while several of the most buildable genres are only mid-sized. Buildability and demand are independent axes, which is exactly why their *overlap*, not either one alone, is the part worth acting on.

Demand doesn't follow difficulty

Buildability alone is a trap: you can now build a visual novel in a day, and so can everyone. The useful question is where *easy-to-build* overlaps with *real, recent demand.* So we joined each genre to two things from the catalogue — how many titles carry the tag (proven demand) and what share were released since 2023 (momentum).

The momentum chart reorders the story. The fastest-growing genres aren't the hardest ones: deckbuilding (64% released since 2023), immersive sim (61%), farming sim (59%), walking sims (51%) and colony sims (50%) lead. Several of those sit squarely in AI's easy zone. A live, growing audience and a low build cost are pointing at the same shelves.

A word on the demand numbers: they are banded popularity estimates — a relative-popularity proxy compiled in our own dataset, not unit sales. There is no revenue figure anywhere here, and we never treat one as if there were. Read owners as *"roughly how crowded the audience is,"* not *"how much money changed hands."*

The opportunity quadrant

Put the two axes together — buildability across the bottom, momentum up the side, bubble size for proven demand — and the sweet spot is the cluster that is *both* easy for AI *and* still growing. Filter hard (build difficulty ≤ 2.7, at least 1,000 titles of proven demand, at least 40% released since 2023) and five genres survive: interactive fiction, hidden object, clicker, walking sim and card game.

Each is a genre where AI can now do most of the production and a real, recent audience keeps buying. Two honourable mentions sit just outside the filter: farming sim (the single hottest momentum among easy genres, just under the volume bar) and the JRPG (a content/art/music genre in a hard-RPG costume).

Then the contrarian read. Deckbuilding is the catalogue's fastest-growing genre (64% recent) but scores 3.18 — *systems craft* — because its soul is *balance*, the AI-hardest content work there is. That gap is the defensible niche: let AI generate every card and every piece of art, and spend your entire human budget on the one thing it can't do — making the numbers sing. Highest demand × hardest-resisting core = the place a skilled human still wins.

The four things AI still can't do

These recur as the 4-and-5 ceilings across every hard genre. If your game's identity rests on one of them, AI changes your *cost*, not your *difficulty*:

Game feel / juice (5) — the millisecond coupling of input, animation, camera, sound and screen-shake that makes a jump or a hit feel *good*. Judged in motion, iteratively, by a human. No tool generates it.

Real-time authoritative netcode (5) — rollback, lag compensation, anti-cheat, server authority. The hardest sustained engineering problem in games, and the place language models produce confident-but-subtly-broken code most often. This single ingredient is why competitive multiplayer is the last frontier.

Level / puzzle / encounter design (4) — spatial pacing and fair-but-surprising challenge. AI generates *plausible* layouts, not ones tuned to teach, escalate and reward. The content looks done and plays hollow.

The "is it fun?" loop (4) — balance, tuning, and knowing what to cut. Taste under playtest data — the work that actually consumes a real game's schedule. The AI revolution in game-making is, overwhelmingly, an asset-and-boilerplate revolution: it compresses the *production* of a game by an order of magnitude and the *design* of one barely at all.

How this was built — and what it does and doesn't claim

Two things are scored independently and then joined. Per-ingredient difficulty is an editorial 1–5 judgement of current tooling — our domain assessment, revisited as tools move. Per-genre buildability is the centrality-weighted mean of a genre's recipe. Demand and momentum are joined live from the MythicLemon Games Catalogue: title counts, banded owner-estimate bands and the share released since 2023.

The honest boundaries: the scores measure automatability of the parts, not fun — "buildable" is not "good." The difficulty numbers are a judgement, not a measurement, and the fast-moving frontiers (3D characters, skeletal animation, adaptive music — all 4 today) could each drop a notch within a year and pull the 3D-action tiers down with them. The demand numbers are banded estimates, not sales. Read the score as *where the labour collapses*, never *where the craft does.*

Master ranking — every genre, easiest to hardest to build with AI

AI-build difficulty is our editorial assessment of current tooling (lower = more automatable). Titles, owners and momentum are from the MythicLemon Games Catalogue; owners are banded estimates, not sales.

The 45 ingredients, by kitchen

AI-build difficulty: 1 = AI does it near end-to-end, 5 = expert real-time craft. Our editorial assessment of current tooling, early 2026.

The opportunity quadrant — AI-easy and still growing

Genres with AI-build difficulty ≤ 2.7, at least 1,000 titles of proven demand and at least 40% released since 2023. Owners are banded estimates.

FAQ