作者展示了利用多模型协作完成游戏全流程的实操路径与成本,为独立开发者提供了可参考的AI工作流范本。
3 months of building a fishing and exploration game almost entirely with Claude Opus 5.5, and the first full zone is now finished and playable.
Nightshoal is being built in Godot with AI handling a huge part of the production pipeline.
The game now has:
fishing + exploration
RPG-style leveling and talents
boat equipment with different rarities
quests and a main story
Dredge-style inventory management
dynamic weather
nighttime dangers like pirates and sharks
moral dilemmas and sea events
an optional cozy mode
This week, Zone 1 and the demo were completed while work started on the Black Market and a large new city for Zone 2.
The workflow is pretty interesting too.
ChatGPT generates visual mockups, then Opus 5.5 uses a /goal + /loop workflow to gradually bring the actual Godot scene closer to the reference.
Jev decides how much reasoning effort each Opus task needs.
A decisions.md file keeps a persistent record of design decisions, with [USER] decisions taking priority over [PROPOSED] AI decisions.
For 3D assets:
ChatGPT → multi-view reference → TripoAI → Opus 5.5 → Blender MCP → Godot
Some assets skip TripoAI completely and are built or textured directly by Opus.
Total AI/tooling cost after roughly 3 months:
~$900–950
And unlike a lot of AI game experiments, there's now an actual finished demo you can play.
来源:Hakimi Eiqbal · x.com