用 Jev 按推理难度给任务分级，13 周做出钓鱼探索游戏
开发者用 Claude Opus 5.5 做游戏，并用 Jev 扫描每个任务、判断该给 Medium 还是 Max 推理档位，再交给管理会话分配；13 周 AI 成本约 900–950 美元。
Week 13 of building a fishing and exploration game almost entirely with Claude Opus 5.5. The first of 5 zones is now almost complete, and exploration has evolved into a full gameplay loop alongside fishing. This week: > 29/40 collectible ships in bottles completed > rare ships can fire cannons and even crack their bottles > new curios and manually placed rewards > engine upgrades with unique sounds > improved island environments > collection cabinet + island dioramas polished > more models built with Opus 5.5 + Blender MCP and TripoAI The AI workflow changed too. The project now has 350+ decisions stored in Decisions.md, tagged as [USER] or [PROPOSED], so Claude can't rewrite the history of what was actually decided. Tasks are also being routed by required reasoning effort. A separate system called Jev scans each task and decides whether it deserves Medium → Max effort before the manager session assigns it. Opus 5.5 is now handling essentially all of the AI development work, with TripoAI for some 3D assets and the OpenAI API for textures. Total AI/tooling cost after 13 weeks: ~$900–950 The demo is getting very close to finished.
这条是引用着另一条写的：x.com/hakimieiqbal/status/2112153626588164368