Building a custom blog from scratch in 2026 is a deliberate choice. Most writers just pick a hosted platform and move on

Building a custom blog from scratch in 2026 is a deliberate choice. Most writers just pick a hosted platform and move on. I decided to rebuild my tech blog with Astro because I wanted to own every layer of the stack and to learn patterns that actually make a static site faster, not just different. The result is a bilingual site that serves Japanese and English content, ships zero JavaScript by default, and keeps every ounce of complexity on my machine rather than the visitor’s browser.

Here are the five patterns that made it work.

Content Collections with Zod

Astro’s Content Collections do more than organize Markdown files into folders. They enforce a contract between your content and your code. I attached a Zod schema to every article collection, which means the build step validates frontmatter before a single page renders.

The schema requires a language field that accepts only two values: ja or en. There is no ambiguity about which language a reader will get. I also added a pair field that links translations together. If I publish a post about Astro in Japanese and later translate it into English, both files share the same pair ID. This makes building a language switcher trivial because the relationship is explicit in the data, not inferred from file names.

Dates arrive from Markdown frontmatter as strings, so the schema coerces them into real Date objects automatically. That eliminates string-mangling logic inside my page templates. The most important benefit, though, is the failure mode. If a file is missing a required field or uses an invalid language code, the build crashes immediately with a clear error. I fix it in my terminal instead of discovering a broken layout or a silent 404 after deployment.

Article Conversion Pipeline

I did not write every post in this new format from day one. Years of content lived on Zenn and Dev.to, each platform with its own quirks and proprietary syntax. Rather than copy-pasting and fixing by hand, I wrote a TypeScript script that converts entire batches of articles into standard Markdown.

Zenn uses custom callout syntax for tips and warnings. My script turns those into semantic HTML aside tags so they render consistently across the site. Dev.to relies on Liquid tags for embeds and special blocks. The pipeline translates those into plain Markdown links that work anywhere.

Some posts contain asides meant only for the original platform, like a disclaimer about Medium paywalls or a Zenn-specific image path. I wrap those in HTML comments so the converter can strip them out during migration. The script processes files line by line, but it respects code boundaries. When it detects a fenced code block, it skips the transformation rules entirely. Corrupting a syntax sample would undermine the whole purpose of a tech blog, so the line-by-line parser treats code blocks as untouchable zones.

Running one command now republishes years of writing without breaking a single link or callout.

Build-Time OGP Image Generation

Social sharing images are usually an afterthought. You either design them manually or install a heavy runtime service that generates cards on demand. I wanted neither. Every Open Graph image on this site is produced during the build so that visitors receive nothing more than a lightweight img tag pointing to a static PNG.

I use Satori, which takes JSX markup and renders it to SVG. The output is crisp, predictable, and easy to template. The real optimization came from font handling. A full Japanese web font can easily top five megabytes. Loading that during the build, let alone asking a browser to fetch it, would be absurd.

Instead, I use Google Fonts subsetting. The script inspects the title text for a given post and requests only the exact glyph set needed to render that string. If a headline uses forty unique Japanese characters, only those forty characters travel across the wire. The build stays fast, and the rendered image never shows broken tofu blocks because the subset is precise. Nothing is left to runtime chance.

Dark Mode via Tailwind Tokens

I refused to decorate every element with dark: utility classes. That approach scales poorly and litters your markup with noise. I redefined the color tokens themselves so the same class name resolves to different values depending on the active theme.

I use CSS custom properties for every surface and text color. In light mode, --color-white maps to #ffffff. In dark mode, the identical variable name points to a near-black value. My HTML stays completely agnostic. A card can use bg-ui-surface and text-ui-primary without caring about the time of day. The theme switch changes the variable definitions at the root, and the entire interface responds instantly.

The one risk with this approach is the flash of light content before stylesheets load. I solved it with a tiny inline script in the document head. It runs before the first paint, checks localStorage and the system preference, and sets the correct data attribute immediately. Because the script blocks the render for only a few milliseconds, the visitor never sees a jarring white burst before dark mode kicks in.

Island Architecture and Zero-JS

Astro’s core premise is that a page should start as static HTML. JavaScript enters only when an interaction genuinely requires it. I took that seriously.

I avoided React for the global menu and the theme toggle. Both are handled with a small amount of vanilla JavaScript that lives in a single module. There is no hydration overhead, no virtual DOM diffing, and no framework runtime to download.

The only heavy library I use is Mermaid.js for rendering diagrams from text. Instead of importing it globally, I wrapped it inside an Intersection Observer. The observer watches for diagram containers. When a user scrolls within a few hundred pixels of one, the script dynamically injects the Mermaid module and renders the diagram. If a post contains no diagrams, that library never touches the network. The initial page load stays light, and the browser only pays for what the reader actually sees.

The Build-First Mindset

The thread running through every pattern is straightforward: if you can do the work during the build, do it there. Validate your data with Zod before the site deploys. Convert proprietary platform syntax ahead of time rather than at request time. Render social images into static files instead of spinning up a server. Resolve theme colors through tokens instead of shipping logic to every client. Defer heavy JavaScript until the user actually needs it.

Pushing complexity leftward into the build step keeps the runtime predictable, the payload small, and the maintenance burden manageable. The site remains fast not because of any single trick, but because there is simply less happening inside the visitor’s browser. That is the real payoff of choosing a static architecture in 2026.