如果 Codex 设置正确,它的精准度和对细节的关注简直是**绝了**。
3 分•作者: ditchfieldcaleb•大约 14 小时前
最近我一直在用 Codex 开发一款塔防游戏,部分是为了学习游戏开发是如何运作的,部分是为了看看只用 Codex 能做到什么程度,完全不手动编写代码。我有了我的 AGENTS.md 和 CODESTYLE.md 文件,以及另外六个全大写的 .md 文件等等,并且正在进行一些重构,以保持代码库的整洁和文件大小的减小等。<p>然后我在最新重构的 ExecPlan 中看到了这个:<p>---<p># 观察结果<p>- 观察:重构后,在正确重新捕获基线后,截图像素完全一致。<p>证据:sha256sum 截图/before-implementation-x.png 截图/after-implementation-x.png 报告了 before/after 对 1、2 和 3 的哈希值匹配。<p>---<p>这太疯狂了!我从未告诉 Codex 对应用程序的 before/after 截图进行 <i>sha 比较</i>,但我确实在我的 PLANS.md 中有说明,要截取游戏 Web 应用程序的 before & after 截图,以确保我们避免前端回归(它使用 GPT-Image-2 进行分析)。因此,对于不影响前端的更改,当然在游戏开始时以 <i>相同的时间戳</i> 截取的截图之间不应该有任何差异。<p>但进行明确的 SHA 比较——这简直是……我 <i>永远</i> 没想到的事情。太棒了。
查看原文
Lately I've been working on a Tower Defense game with Codex, in part to learn how game development works and in part to see how far I can get using <i>just</i> Codex, no manual coding at all. I've got my AGENTS md & my CODESTYLE md & six other ALLCAPS md files etc, and am working on some refactoring to keep the codebase clean & file sizes low, etc.<p>And then I see this in the ExecPlan for my latest refactor:<p>---<p># Observations<p>- Observation: The refactor made the screenshots pixel-identical after the baseline was recaptured correctly.<p>Evidence: sha256sum screenshots/before-implementation-x.png screenshots/after-implementation-x.png reported matching hashes for before/after pairs 1, 2, and 3.<p>---<p>Which is crazy! I've never told Codex to do an <i>sha compare</i> on before/after screenshots of the app, but I do have instructions in my PLANS.md to take before & after screenshots of the webapp for the game to make sure we avoid frontend regressions (it uses GPT-Image-2 for analysis). So for non-frontend impacting changes, of course nothing should be different between screenshots taken at <i>identical timestamps</i> into the game start.<p>But doing an explicit SHA compare - that's just...not something I would've <i>ever</i> thought of. Wild.