Currently, AI still encounters bugs when facing complex projects. Moreover, modifying parts of the bug often gets more troublesome.

In fact, I have always used the simplest method: the method of repeating it three times.

Many people write subagents, skills, or complex projects like superpower, oh-my-claude. However, I feel that repeating it three times is still simpler and more convenient.

The method is simple: after the AI completes the task, you say to let the AI review the code according to the requirements. According to my experience:

In the first round, the AI will find several issues; you can just let it modify them without thinking.

In the second round, the AI will find several issues; you need to glance at them and let it modify the serious ones.

In the third round, the places that need special attention will come up. The AI will always find some issues, some of which are not important at all. So, at this point, you need to ask another AI to review it.

For example, for coding done by Codex, you can let Claude Code review it. The reverse is also true.

Basically, the evaluations provided by the second AI in the third round will be mostly minor issues. Then you can stop. If there are still some serious problems, you can continue the review.

The results I have achieved over the past few months show that I have not encountered any situations requiring code fixes during the project operation phase.

Through repeated practice of this most straightforward method, I also understood

1) Even the best programming AI, like Codex or Claude Code paired with Opus, will still encounter bugs when writing complex projects.

2) Even the best programming AI will sometimes find new bugs unnecessarily.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。