跳到正文
RCreddit.com·

I build my game with coding agents. The scarce resource is the decisions I still have to make.

AI 摘要

一位独立开发者正在构建一款由AI驱动的生命模拟游戏,其中AI代理负责编写代码,而开发者则专注于产品方向和技术决策。该项目目前包含大约168k行的后端Python代码、108k行的前端源代码文件以及95k行的后端测试代码。这种开发方式有助于管理不断增长的代码库,开发者强调,稀缺的资源是他们仍然需要做出的决策。

为什么是这条

该报告首次详细介绍了独立开发者如何利用AI代理编写游戏代码,与传统开发不同,其侧重于高层决策。

时间与来源

时间显示为 UTC

显示时区:UTC

本地时区尚不可用,暂时显示 UTC。

发布当时偏移:UTC+02026年9月8日 17:27 UTC

收录当时偏移:UTC+02026年9月9日 17:00 UTC

发布
2026年9月8日 17:27
收录
2026年9月9日 17:00
来源类型
开发者社区
档位
社区
信源状态
正常

档位是按信源手工设定的编辑判断,不是逐条打分。

讨论趋势

暂无对比
最近 24 小时与此前 24 小时的快照均值对比 · 7 天曲线

百分比基于采集到的讨论信号,不代表新增评论数或独立参与人数。曲线仅用于同一话题在不同时段的比较。

正文

I'm a solo developer building an AI-driven life simulation game. AI writes the code; I decide the product direction and make the technical calls. The part I want to share is how I keep that arrangement manageable as the repository grows.

For scale, my current checkout has about 168k lines of backend Python, 108k lines in the selected frontend source files, and another 95k in backend tests. Those are text-line counts including comments and blanks, not SLOC or a productivity benchmark. More code can also mean more maintenance.

My biggest constraint is how many decisions stay in my head after a task ends. Here are the concrete conventions I use:

Route context by the task. The root AGENTS.md is a map of responsibilities and reading requirements. Changing the simulation loop points to its runtime contract; changing UI points to frontend conventions. Local instructions live beside their modules. I don't ask every task to digest every historical document.

Separate decisions from implementation. I keep a document of product and collaboration decisions, including rejected directions. Code and schemas describe what exists. Active contract documents describe what should be true. If they disagree, the agent has to show the conflict; silently declaring either one obsolete is not a resolution.

Make repeated corrections executable where possible. My frontend has checks for design tokens and UI structure, plus generated protocol checks. The benefit is that the next task can discover a violation from tooling instead of requiring me to remember the last conversation. These checks don't decide whether the design is good.

Define completion beyond the diff. Behavior-changing work needs an expected outcome, a data source and time window, a pass criterion, and a follow-up schedule. A merged change and a verified effect are separate claims. A check that was skipped remains skipped.

One less comfortable rule: when a mechanism needs a second layer of patches, pause and ask whether it should still exist. AI can keep making a local solution more elaborate while leaving me with a system I no longer understand.

Compared with keeping instructions only in chat, the tradeoff is maintaining these repo contracts. They can become stale too. I don't have a controlled before/after measurement of time saved, so I'm sharing the workflow rather than a speedup claim.

For people maintaining larger projects with coding agents: which repeated human correction have you successfully moved into a check, and which still needs your judgment?

Disclosure: AI-assisted writing, based on my actual repository and development decisions.

来源·reddit.com