dungeon hunter

Inspiration

We wanted to prove that AI-assisted development can ship a complete, polished game — not just a tech demo. Traditional game dev takes months. We asked: what if Codex, paired with GPT-5.6, could collapse that to days? Dungeon Hunter is our answer: a full pixel-art dungeon crawler built end-to-end with AI — from the first line of engine code to every single sprite.

What it does

Real-time pixel-art dungeon crawler with 6 heroes, 10 dungeons, 8 bosses (72 unique spells across 3 phases), real-time Canvas combat with Sekiro-style shield-break mechanics, a branching campaign (8 quests, 3 acts), item system with 5 rarity tiers, skill upgrades with branching choices, town exploration, cloud saves with Supabase, and an endless wave mode. All rendered in a retro 480×270 viewport with 2x pixel scaling.

How we built it

Stack: Next.js 16, React 19, HTML5 Canvas, TypeScript, Supabase, Prisma, JWT auth. Every component was shaped through Codex agent mode sessions — from architecting the 3,400-line raid engine (real-time physics, boss AI state machines, projectile systems, hazard AoEs, NovaWave expansions) to extracting clean module boundaries (bossSpells.ts, skills.ts via GameContext pattern), wiring Supabase cloud saves, and designing the branching campaign system. All visual assets were generated with GPT Images 2.0 . Every hero spritesheet (6 heroes × 3-4 animation sets × 4 directions = ~72 frames), every building texture (castle, manor houses, dungeon gate, smithy, market), every terrain tile (grass, roads, water, lava, volcanic rock), every NPC sprite (King Aldric, Lord Castellan, villagers, blacksmith, merchant), and every decoration (fountains, fences, signposts, planters, barrels). Output sprites were chroma-key processed for transparent backgrounds and loaded into a custom 128px-cell directional sprite engine built with Codex.

Challenges we ran into

Canvas + React lifecycle: React StrictMode double-mount corrupted our canvas zoom transform — solved by switching from ctx.scale to ctx.setTransform. Engine monolith: The raid engine grew to 3,400+ lines. Codex helped us refactor it into clean modules using GameContext/SkillContext interfaces. Sprite pipeline: Generated sprites needed consistent 128×128 cell sizing, 4-direction layout, chroma-key processing, and per-hero custom draw functions — we built a pipeline where Codex parsed sheet layouts and Imagen-2 regenerated frames on the fly. Asset consistency: Keeping pixel art style coherent across 6 heroes, 8 bosses, buildings, and terrain required iterative prompt engineering and regenerate cycles with GPT Images 2.0

Accomplishments that we're proud of

• 8 unique bosses with 72 total spells — each with 3 escalating phases (Giant Slime splits, Lich raises the dead, Leviathan summons maelstroms)
• Sekiro-style shield-break mechanic — bosses have a shield bar (40% maxHP) that must be chipped to expose a 5s vulnerability window with 1.5x damage
• 6 heroes with 18 upgradable skills — skill level 3 offers 2 branch choices (e.g., Charge → Battering Ram or Shadow Rush), giving meaningful build diversity
• Every visual asset AI-generated — hero spritesheets, buildings, terrain tiles, NPCs, decorations — all created through Google Imagen-2 and integrated into a production game
• Full campaign with branching narrative — 8 quests across 3 acts with a meaningful choice: Undead path or Ocean path
• 10,000+ lines of working game code — not a prototype, a shippable game with cloud saves, JWT auth, endless mode, and a map editor ## What we learned Codex isn't just a code generator — it's a pair programmer that excels at refactoring and architecture. The biggest wins came from iterative dialogues: "extract boss spells into a separate module", "implement Sekiro-style shield break", "add blood orb system for Wolf Pack". GPT Images 2.0 's pixel art capabilities let us iterate on game art instantly — describe a "frost knight in pixel art, 128×128, 4-direction idle sheet", get it in seconds, and have Codex wire it into the engine immediately. The tight loop of prompt → generate → integrate → test is a fundamentally new way to build games. ## What's next for Dungeon Hunter
• Multiplayer co-op raids
• More dungeon biomes and bosses
• Weapon crafting system
• Deeper GPT-5.6 integration for dynamic quest generation and NPC dialogue
• Mobile touch controls (Joystick component already built)
• Leaderboards and weekly challenges

Built With

• canvas
• gpt
• gpt-5.6
• nextjs
• supabase