正在打开每日关注
正在打开每日关注
以 Zentra 仪表盘为例，比较 Gemini 与 Claude 在多轮视觉改稿和指令遵循上的差异。
Claude Opus 5 actually did a goood job https://t.co/OYCINSMjXj
Tbh, Gemini 3.7 Flash still isn't there yet 🥶 I don't see any meaningful improvement over Gemini 3.1 Pro or the other Gemini models. I tried vibe coding a dashboard in Google AI Studio. It got some things right, but constantly failed at others. I repeatedly asked it to redo sections to match the design, but it just wouldn't. No visual changes at all. Although its speed is impressive compared to Kimi K3 and even Claude, it lacks intelligence to create things the way user requests. This has been a problem with Gemini for a long time. I've built quite a few apps with it, and while it can get things done with a few more iterations than other models, eventually it just stops following requests. You can keep asking. It simply doesn't make the changes. Google really needs to step up its game. Studio: https://t.co/LGlMZyvE87 Design: https://t.co/FaLvG3F4tE
是什么 这是一条关于 AI 辅助前端开发的经验对比。引用内容称，作者在 Google AI Studio 中用 Gemini 3.7 Flash 做 dashboard 的 vibe coding：模型初步完成部分任务，却在被反复要求按设计重做时几乎不改变视觉结果；原推文只用“Claude Opus 5 did a good job”给出相反的正面判断。链接正文停在 Google 登录页，无法核验 Claude 的具体任务、提示词和改动。
核心设计与论点 争议焦点不是首版速度，而是模型能否理解参考界面、定位待改区域，并在多轮反馈后持续执行。作者认为 Gemini 3.7 Flash 相比 Gemini 3.1 Pro 等模型没有明显提升，还会在完成一部分后停止遵循请求。截图中的 Zentra 仪表盘以支付漏斗、数据卡、点阵图和洞察卡构成复杂评测对象。
关键洞察 这类体验应拆成初始生成、视觉一致性、交互完成度和持续改稿能力来观察；“看起来完成”不等于真的按设计迭代。把模型比较落到可观察的前后差异，比单看生成速度更有决策价值。
边界与风险 这是单次、主观的模型比较，没有原始设计、代码 diff、任务脚本、成功标准或重复实验。截图能证明成品界面，不足以证明具体模型归因，因此更适合作为选型线索，而不是性能基准。