Building This Site with AI in a Day: From Résumé and Notes to an Automated Deploy

This site is itself the product of a day of AI-assisted work (with Claude Code). The starting point was just a PDF résumé, a few scattered tech notes, and a talk write-up; the end point is this bilingual site that goes live the moment I push. Here’s the workflow, the discipline, and the gotchas — in case you want to do something similar.

Related: A Practical Guide to AI-Assisted Development.

The stack

#

• Claude Code — the AI pair, doing most of the mechanical work, translation, and consistency checks.
• Hugo (extended) + the Blowfish theme — static site generator; Blowfish installed as a Hugo Module.
• GitHub Actions → GitHub Pages — push to main and it builds and deploys automatically.

The workflow

#

1. Positioning first, then the README

#

The first step wasn’t code — it was getting the positioning straight. A GitHub profile README is a business card, so I first decided the audience (job search / overseas), the angle (backend + applied ML/CV), and the depth (a minimal card, detail left to the site). I used a “drill down every decision branch” process to settle each choice before building — so I wouldn’t finish and then realize the direction was wrong.

The result is a minimal card: a one-line tagline, grouped tech badges, three contact links, and everything else routed to this site.

2. Scaffold the site with Hugo + Blowfish

#

• Install Blowfish via Hugo Modules (hugo mod init → import the theme → hugo mod get); upgrading later is just hugo mod get -u.
• Bilingual via Hugo’s built-in i18n: language definitions in languages.toml, content with .en.md / .zh-tw.md suffixes.
• Use Blowfish’s profile homepage layout, surfacing the résumé highlights (backend, ML/CV, talks, publications).

3. Automate the deploy with GitHub Actions

#

One workflow: push to main → install Hugo extended → hugo mod get → hugo --gc --minify → publish to GitHub Pages with the official upload-pages-artifact / deploy-pages actions. After that, publishing a post is just a push, live in a minute or two.

The one manual step: in the repo, Settings → Pages → Source = “GitHub Actions”. Set it once and forget it.

4. Turning existing docs into posts — discipline is the point

#

This is the part that most needs a human in the loop; the AI shouldn’t decide it alone. I converted the résumé, a talk write-up, self-study notes, and work notes into bilingual posts, holding to a few rules:

• De-identify (most important): for anything from work, describe only the domain and the technology — no employer name, no internal service names, no business specifics. For example, an internal Feign development guide whose example was a sensitive business case got its entire example domain swapped to a neutral “e-commerce order,” keeping only the generic architectural pattern.
• Strip private links: reference links pointing to a personal knowledge base are broken for readers and leak internal IDs — remove them all.
• Strip tracking params: when pasting course/external links, cut the long gclid / utm_* tail and keep just the clean URL.
• Bilingual: every post ships in both English and Chinese.

After each deploy I run a quick audit (e.g. grep -ri over the published output for any leftover sensitive terms) and only call it done once the live site is clean.

Gotchas

#

• Match the theme’s Hugo version: Blowfish has a tested maximum Hugo version; a newer Hugo throws a compatibility warning. Pinning both local and CI to the theme’s supported version is the simplest fix.
• Future dates get skipped: Hugo doesn’t build “future” posts by default. When I dated a post “today,” my machine’s timezone (UTC+8) meant that “today at UTC midnight” was actually still in the future, so the post wasn’t built. Fix: give the date an explicit timezone (or use the previous day).
• Config file naming: language definitions go in languages.toml; a filename like languages.en.toml makes Hugo treat the .en as a language qualifier and drop the keys. The .<lang> suffix is only for menus.<lang>.toml.

Takeaways

#

AI did the mechanical work — scaffolding, the workflow, translation, formatting, consistency audits — quickly and reliably, which is what made “live in a day” realistic. But two things stayed firmly my responsibility: the positioning judgment, and policing the disclosure boundary — what can be public and what must be de-identified. The AI can flag it and do it, but you make the final call.