Regarding model usage, let’s start with Gemini-2.5-pro. Although overshadowed by Sam’s release of 4o image generation, I believe its power is underestimated.
Gemini is surprisingly powerful in finding and fixing bugs, but its tool-calling initiative is low. Without guidance, it might not even understand the project’s files and code logic.
Therefore, I recommend Gemini’s flagship model as the first choice for optimizing a relatively mature project.
However, unless you provide it with the entire project code; when using it in an AI IDE, to help it understand the project, you can prompt any Claude model with: “Maximize tool calls to understand all code logic”, possibly needing 2-3 rounds of repetition depending on the project size.
The 3.7 series models are undeniably powerful in coding and tool calling. However, during collaboration, Claude-3.7 is like a runaway horse—it doesn’t care about the project structure. Based on its belief that “pulling one hair affects the whole body,” often when adding a small feature, without supervision, Claude can completely transform the project beyond recognition, requiring cautious use.
If seeking high creativity, pursuing Vibe coding, not afraid of errors, and only acting as a feedback provider and requirement proposer, you can solely use Claude-3.7-thinking.
If safety and high efficiency are needed, the combination of Gemini as the architect, Claude as the engineer, and the individual as the connector is, in my opinion, quite excellent.
关于模型使用方面，先说 Gemini-2.5-pro，被 Sam 发布 4o 图片生成”狙击”，我认为他的强大被低估了。
Gemini 在查找 bug 与修复方面出乎意料的强大，但是他的工具调用积极性不高，如果不去引导他，甚至不知道项目的文件以及代码逻辑。
因此，我建议，对于一个比较完善的项目进行优化。现阶段 Gemini 旗舰模型是首选。
然而，除非将全部项目代码全告诉他；在 AI IDE 中使用时为了让他了解项目，可以通过 Claude 任意一款模型 prompt：“最大化调用工具了解所有的代码逻辑”，甚至可能需要重复 2-3 轮，视项目大小而定。
3.7 系列模型不可否认的是代码方面、工具调用很强大，然而在协作过程 Claude-3.7 犹如脱缰的野马，他不会在意项目的结构。甚至基于他认为”牵一发而动全身”，往往每增添一个小功能，如果在没有监管的情况下，Claude 会将项目改得面目全非，需要慎用。
如果寻求高创意，追求 Vibe coding（氛围编程），不害怕出错，只作为反馈者与需求提出者，完全可以只使用 Claude-3.7-thinking。
如果需要安全性与高效率，Gemini 作为架构师，Claude 作为工程师，个人作为链接者的组合，我认为是很不错的。