每隔几个月，就有新模型登顶跑分榜。追逐榜单很诱人。但我日常使用的模型并不是纸面上最强的。它是契合我实际工作形态的那一个。
我大部分工作是在现有代码库里做小幅编辑。一个快速、便宜、擅长续写的模型，胜过一个擅长一次性表演的模型。
什么塑造了我的选择
- 编辑循环的速度
- 长会话的成本
- 续写的质量，而不仅仅是生成
- 它读取现有结构的能力
选择契合你工作形态的模型。跑分衡量了某些东西，但不是键盘前最重要的东西。
2026-07-21aiinmydraft
2026 年我实际在用的模型
2026 年有很多强模型。但我日常使用的那几个并不是跑分最强的。它们是契合我工作形态的那几个。
更多更新
general2026-10-05
Search Experience Acceptance Criteria That Test Real Behavior
Teams usually find gaps in search experience when an exception exposes an unclear decision. A better starting point is to connect the query box to helpful ranking, filters, result context, empty states, and recovery. This matters because a technically fast…
search experienceacceptance-criteriapractical guide
阅读
general2026-10-04
Homepage Structure Acceptance Criteria That Test Real Behavior
A credible homepage structure implementation has a narrow promise: lead with an evidence-backed promise, audience, primary action, proof, and a clear route to detail. Treat a collection of slogans forces visitors to infer what the product does and why they…
homepage structureacceptance-criteriapractical guide
阅读
general2026-10-04
Design System Acceptance Criteria That Test Real Behavior
The hard part of design system is not adding another tool or screen. It is deciding how to standardize the repeated tokens, components, states, accessibility rules, versions, and exceptions already in use, while accounting for one concrete failure: a…
design systemacceptance-criteriapractical guide
阅读