OpenAIWhat they did
The poster writes a weekly devlog of a fishing and exploration game built almost entirely with Claude Opus 5.5. A manager session hands tasks to other Claude sessions. In week 13 they split those sessions by effort level, from medium to max, and had a tool they call Jev read each task and pick the level it deserves. A Decisions.md file tags each of more than 350 design decisions as [USER] or [PROPOSED], and user decisions win, so the model cannot claim the poster decided something they did not. 3D models came from TripoAI and from Opus working in Blender through Blender MCP; Opus fixed up a Tripo model of the rarest ship, for example. Textures were made either in Tripo or by Opus calling OpenAI image generation with an API key the poster gave it.
How it went
Zone 1 of 5, the demo zone, is nearly done apart from polish and bug fixes, and a free demo is on itch.io. A week later the poster called zone 1 and the demo finished and moved on to a black market and the zone 2 city. That week brought curios, engine sounds, a relit 3D collection cabinet, island dioramas and 29 of 40 ships in bottles; the rarest one cracks its own bottle when it fires. The poster felt no difference in quality or token use from the new effort routing but says that is a subjective impression. The Steam release is waiting on ID verification.
Worth knowing
Monthly spend is about $900 to $950: three Claude Max plans at $200, a $100 Astra plan, OpenAI credits, ElevenLabs and Tripo. Tripo has paid for the devlogs since week 8 and gifted a Max plan, so that cost is missing from the total. The harness behind the Claude sessions is not named.