The old site stopped in 2023
The last content update on my personal site was October 23, 2023. I pushed it the next day, and then it sat in a corner of the internet gathering dust for three years.
Going back to it felt like archaeology: Hexo 6.3.0, four course notes, nine JPEG covers, Bulma + jQuery + Fancybox, a repository around 25 MB. The About page still called me a “junior undergraduate”, the CV was the 2023 version, and the navigation contained a typo — “Intership-Experience”. It was like running into an old friend who greets you by the wrong name.
Honestly, that was my first attempt at a personal site. The engineering was immature and the design was still finding its feet. But some of it went beyond a funny story: zoom was disabled on mobile, so enlarging the text meant bringing my face close to the screen like a treasure map; the analytics scripts were commented out but the structure was still there; and every time I wanted to change something, I had to travel back to my 2023 state of mind. A site nobody maintains slowly turns into an out-of-date introduction of yourself.
A lighter foundation
The new site is built with Astro 7 and ships as static output. Markdown and content collections handle the content, with separate collections for projects, papers, and blog posts. Styles are hand-written design tokens, all fonts are self-hosted (no Google Fonts), and RSS, the sitemap, and the 404 page are generated by the build. Deployment goes through GitHub Actions to GitHub Pages.
The visual direction changed too: warm archival paper, near-black ink, a touch of burnt orange; Newsreader for headings, Public Sans for body text, IBM Plex Mono for labels; a three-layer glass navigation that falls back to a solid surface on Firefox. There is a 1.78-second “QL signal lock” intro — it plays on refresh, not when you move between pages. It carries no information; it just says this site is taken seriously.
English is the source. Simplified Chinese and Japanese are generated at build time, with every translated string cached in the repository, so CI needs no API key. Three static sites, each with its own RSS feed.
Letting the local DeepSeek do the work
Most of the execution on this rebuild was not typed line by line by me. The model runs locally on an air-gapped Mac mini (M-series chip, quietly humming along), connected through DeepSeek Harness into an agent that can actually do things: read the repository docs, edit files, run npm run check and the build, drive Playwright to screenshot and click through three browsers, run axe for accessibility, and hand me the results.
My job was closer to editing and signing off: is this content true, are the dates and roles right, do we keep this design, can this change be pushed. The model is good at finishing repetitive checks (like 90 page, viewport, and browser combinations in regression) and at producing a working patch once I describe the problem. But it does not know which parts of my experience are real, so anything factual gets my signature.
The practical benefit is sustainability. Before, changing the site meant rebuilding the environment and digging up old commands. Now the rules live in the docs and the agent follows them: run the checks, fix what fails, and wait for my nod. It does not slack off, and it does not need coffee — though I am not sure whether it drinks.
A few specific traps
- Three languages are not three pages, they are three static documents. Switching language is a full page load, so the browser jumps back to the top. A small script that records the scroll position fixed it; now switching languages actually keeps your place.
- Translations need a cache. When the English source changes, only the new strings go out for translation and the rest are reused; CI rebuilds from the offline cache and fails if a string is missing — much better than noticing a gap on the page.
- Accessibility has to be treated as a feature. Removing
user-scalable=nofrom the old site was only the start; focus styles, skip links,aria-current, contrast, and reduced motion all belong in the regression. - A blog needs categories. After three years of course notes, I gave the blog
life,course, andprojectcategories plus keyword search. Search reads the rendered text, so the Chinese page searches Chinese and the Japanese page searches Japanese.
Now
The site is live again. The four 2023 course notes are still there, labelled “Course notes”, and the public projects, papers, and experience have been checked against my CV. The old site’s images and 2023 PDF are still in the repository root — my first attempt at a personal site, technically immature, but hard to throw away.
Bringing a three-year-old site back to life costs less than I expected. The trick is writing down how to change things, how to verify them, and when it is allowed to ship — then letting an executor who never gets tired follow the rules. The judgement still has to be mine.