本月精選
Ratios are easy to pick and mostly wrong. A scale earns its place when every step has a job on a real page: five sizes, a measure that holds at 320 and at 1440 pixels, and clamp() written so a designer can read it out loud.
最近發布的內容，最新在最上。更早的文章都收在文章庫裡。
Lab tools measure a robot on fibre. Field data measures your readers on a bus. Here is the smallest honest setup we could build: three metrics, one endpoint, no third party script, and the week our two sets of numbers disagreed by two and a half seconds.
Most projects that went wrong were already wrong on page one. This is the brief we now refuse to start without: who decides, what already exists, what counts as done, and the three questions that reveal a budget before anyone has to say a number.
Inverting a palette is an afternoon of work. The hard part is everything the palette does not control: photographs, screenshots, charts, embedded video, and the one brand colour that stops meaning anything once the background goes dark.
A schema that fails the build is worth more than a CMS that lets anyone publish a broken page. What we validate in Astro, what we deliberately leave alone, and the three rules that keep a Zod schema from slowly turning into a second CMS.
A variable font is not automatically cheaper than two static weights, subsetting saves less than people think, and the expensive part is usually the fallback nobody designed. What a typeface actually costs, measured in bytes, requests and reflow.
A budget is not a Lighthouse score and it is not a Slack message. Here is the one we now write into the contract: four numbers, the device they are measured on, and the sentence that decides what happens when someone asks for a video header in week nine.
每篇文章只歸屬一個主題，這樣一個系列可以從頭讀到尾，不用在其他內容裡繞路。
