A solo dev building a fishing game entirely with AI documents how the workflow matured by week 3 — from “just typing what I want” to a network of dedicated AI sessions wired into Blender and Godot via MCPs. The post is really about process discipline, and it reads like a checklist born from concrete failures.

The six rules:

  • Reference first — ChatGPT generates concept art and UI mockups; the human decides the look before any geometry exists. “Without human eye AI will just make stuff not fit together.”
  • Scripts over generated meshes — Rodin’s 3D output needed too much cleanup, so a dedicated Blender MCP session writes models instead. The script is the source, so proportions stay editable — and low-poly is all the camera distance needs.
  • One session per job — modeling, integration, and UI each get their own session; a general chat window dispatches commands between them.
  • Screenshots are the referee — nothing counts until it’s rendered from the actual in-game camera. Four new islands checked from overhead looked fine, but in game they were “flat sandy pancakes floating on the water. From above, a pancake looks perfect.”
  • Blind scored reviews — criteria written before the change, twelve things scored out of ten, ship only at 8+. One critic judged screenshots without being told which version was newer — that kills “I changed loads, so it must be better.”
  • Every decision tagged with who made it — marked USER or AI-suggested, because an upgrade nobody ever proposed once sat in the design docs and got quoted back as the author’s own idea.

The summary line: “AI does the work, I’m the art director, and I’m the only human in the loop.”

The standout is the acceptance-testing discipline — in-game screenshots as the referee plus pre-written blind scoring directly attack the two classic failure modes of AI gamedev: artifact rot nobody looked at, and improvement theater with no baseline.