QuizCaster: Shipping an AI-Generated Game With No Game-Dev Background

What QuizCaster is

QuizCaster is a free educational roguelike — Matt's first game ever — built by two people (Matt and his collaborator Jason) with the core loop coming together in a couple of weeks and about two months of total work. It's inspired by Vampire Survivors for its dead-simple control scheme (you really only move and dash with the space bar) and by the roguelike genre for its replayability: randomized modifiers, dynamic weapon drops, and level-up beats that make each run feel different. Every 15 seconds the action pauses and serves a question. It ships with K–5 math, spelling, and vocab plus SAT and AP prep decks, and you can bring your own flashcards by importing from Quizlet or Anki or pasting AI-generated sets. At the time of the talk it had around 500 weekly active users and more than 10,000 quizzes answered — including one dedicated player in Brazil with 24+ hours logged.

The learning science

Matt frames a study game as having three jobs: motivation (will kids choose to play it?), learning outcomes (are they actually learning?), and cognitive load (is it overwhelming?). On motivation, the key move is that learning is integrated, not gated — answers influence your rewards rather than acting as a wall, and roguelike variable rewards keep players coming back because they never know what upgrade they'll get. An early lesson from beta testing: kids started running away from XP orbs because they associated leveling up with "now I have to do math," so the team decoupled the two — you gain XP from combat, and math appears on its own 15-second timer. On learning outcomes, QuizCaster uses spaced repetition: questions you get right move toward the discard pile, questions you miss come back sooner and then at widening intervals, and a mastery page shows performance over the last 5, last 10, and all-time attempts. On cognitive load, the minimal control scheme is deliberate — kids are already learning while playing, so the game shouldn't also demand complex inputs.

The visual pipeline

By Matt's estimate, about 75% of the work was visual asset generation. The main tool is Pixel Lab (chosen over Retro Diffusion, which they liked for high-detail stills but found weaker on animation), with a whole API and MCP for generating characters, objects, and map tile sets. A key trick: go from a text description to a reference image first (via Gemini or a similar generator), then from that reference to the pixel art — the model does a much better job than jumping straight to pixels. Anything coming back from Pixel Lab runs through Python (PIL) cleanup scripts that catch outlier pixels, plus Aseprite (also MCP-accessible) for hand fine-tuning. Matt stresses codifying repeatable patterns into skills — one per Pixel Lab endpoint, capturing which endpoint to hit and how many tokens it costs. Two consistency rules matter: lock the pixel density so a pixel is the same size across objects, and lock the color palette. His single biggest regret is choosing pixel art at all — the margin for error is thinner than with SVGs, and he'd reach for SVG-based art next time so the AI wouldn't have to sweat low-res rendering.

The audio pipeline

Audio came almost entirely from two tools. JSFXR is a free, local, instant procedural synth (an npm package) used for effects like XP pickups, level-ups, and UI clicks. ElevenLabs handled the more "generated" sounds and all the music — and notably, the lowest $6 plan was more than enough, never hitting limits. Matt's prompting tip: describe the real physical source ("a flaming meteor hitting the ground and puffing up dust") rather than abstractions like "plasma energy," which AI handles far better. Because a model can't hear its own output, the team built a skill that visualizes the waveform, finds anomalies like stray buzzing from ElevenLabs, chops them out, and balances levels.

Game dev tips, balance, and QA

Beyond "core loop first," Matt covered balancing 35 weapons across schools like arcane, fire, ice, void, shock, and poison — an industry practice called red boxing. Human playtesting is still king, but to find outliers they built AI playtesters: invulnerable bots with movement patterns optimized per weapon, running 100–200 parallel runs to surface what wasn't balanced. On QA, he's known for feature flags — a way to test features in production and instantly roll back a bad release without a staging environment — alongside a cheat menu for isolating weapons and jumping to specific moments, and a system map (built by Jason) that visualizes every screen, state, and trigger. The through-line: QA is an underrated skill in AI development, and you should never let a model grade its own work — use a separate, blind adversarial reviewer that checks the product against the standard, since coding agents tend to rubber-stamp themselves.

Marketing with AI

Marketing runs on AI too. A scheduled Claude job produces a devlog every Monday at 9am from the previous week's PRs, summarizing what shipped and suggesting clips — published live at quizcaster.com/devlog and cross-posted to itch. The team writes shorts for YouTube, TikTok, X, Instagram, and Facebook, and runs paid ads primarily on Reddit. Matt's key habit: don't start a script until you've digested which formats and hooks are already working on each platform. Traffic splits across kids, parents, and self-studiers, and QuizCaster's edge is modularity — paste any flashcards and it works, which few games offer.

The Gauntlet Loop

Asked about the "Gauntlet Loop" (unrelated to Gauntlet AI the school), Matt was candid: he didn't use it to build QuizCaster and probably wouldn't. His critique is that it's fundamentally referential — you point at an existing game and say "build something like that" — and that the results are often architecturally messy, token-intensive, and expensive (one example ran about $632 in tokens). The part he agrees with completely is the blind adversarial reviewer: a separate agent that verifies the work against the standard, which he considers essential in AI development generally.