Been building a MapleStory-themed Vampire Survivors game solo and figured I’d share the workflow that’s kept me sane.
The short version: I use an AI assistant as my systems architect and a separate agentic coding environment (V8) as my hands. Neither operates alone — they work in a strict 3-phase loop.
Phase 1 — AI drafts a survey prompt. V8 goes off and queries the asset pipeline (MCP tools, CDN, codebase) and reports back. No code is written yet.
Phase 2 — AI reads the actual source files that will be touched. This step is non-negotiable. Prompts written from memory produce wrong schemas, wrong line refs, broken TypeScript — learned that the hard way.
Phase 3 — AI designs the implementation and drafts a precise execution prompt. V8 generates and saves the code. Then we verify with tsc --noEmit + bun run build before moving on.
A few rules that made this sustainable:
- AI never writes game files directly. Its job is read + review + draft prompts.
- V8 never generates code during Phase 1.
- Every new gate passes a 9-item checklist before it’s declared done.
- Asset pipeline flows one way: MCP → WZ JSON/PNG → render plan JSON → runtime object pool.
It’s slower to set up than just vibing with autocomplete, but once the guard rails are in place it scales surprisingly well. Been shipping gates consistently without regression.
Hope this gives someone a useful starting point for structuring AI-assisted game dev. It’s not perfect, but it’s working.