Personal site for Vito Lin. Hand-written HTML and CSS, one typeface, no JavaScript on the page and no npm packages anywhere. Posts are Markdown; a build script that uses only Node's standard library turns them into static HTML.

posts/*.md            ← you write here
data/profile.json     ← identity, hero line, about text, links
data/github.json      ← generated: repos and pull requests
templates/base.html   page shell
assets/style.css      the whole design
scripts/build.mjs     Markdown → HTML, plus RSS and sitemap
scripts/fetch-github.mjs      GitHub API → data/github.json
scripts/import-linkedin.mjs   Shares.csv → posts/*.md drafts
index.html writing/ about/ 404.html feed.xml sitemap.xml   ← generated, do not hand-edit

cd ~/Desktop
mv vito-lin-dev.github.io vito-lin-dev.github.io.old
# unzip the new archive here, so that ~/Desktop/vito-lin-dev.github.io/index.html exists

If you already pushed the old version, keep the same repo and force the new tree in:

cd ~/Desktop/vito-lin-dev.github.io
git init -b main
git add .
git commit -m "feat: rebuild site, refined minimal"
git remote add origin https://github.com/vito-lin-dev/vito-lin-dev.github.io.git
git push -u origin main --force

Then check Settings → Pages is on main / / (root), and Settings → Actions → General → Workflow permissions is Read and write.

node scripts/build.mjs            # build published posts
node scripts/build.mjs --drafts   # include draft: true, for previewing
python3 -m http.server 8000       # preview at http://localhost:8000

Push to main and the workflow refreshes the GitHub data, rebuilds, and commits the output. You never edit generated HTML by hand.

Create posts/my-slug.md. The filename becomes the URL: /writing/my-slug/.

---
title: "Why cuTile's tutorial imports are broken"
date: 2026-08-11
summary: "A stale import path in the official examples, the compile error it produces, and the two lines that fix it."
linkedin: https://www.linkedin.com/posts/vitothedev_...
---

Body starts here. Markdown subset: `##`/`###` headings, paragraphs, **bold**,
*italic*, `code`, [links](https://example.com), bullet and numbered lists,
blockquotes, images, fenced code blocks. Lists do not nest.

Only title and date are required. summary is what people read in the list, the RSS feed and Google results — write it for a stranger. draft: true keeps a post out of the build.

First work out which kind you have. Open one of your posts and look at its URL: /posts/... is a short feed post, /pulse/... is a long-form Article. They land in different files in the export.

1. On desktop LinkedIn: Settings & Privacy -> Data privacy -> Get a copy of your data.
   • Short posts only: choose Want something in particular? -> Posts. Ready in roughly ten minutes.
   • If you have Articles too: request the larger archive. It takes up to 24 hours and arrives in two emails.
2. Download and unzip. Shares.csv holds every short post with its date, URL and full text. Articles/ holds one HTML file per long-form article.
3. Convert whichever you have:

node scripts/import-linkedin.mjs ~/Downloads/Shares.csv    # short posts
node scripts/import-linkedin.mjs ~/Downloads/Articles      # long articles

Each becomes a posts/*.md file with draft: true, its date, its LinkedIn URL, and a provisional title. Reshares and one-liners are skipped. Article imports keep code blocks, lists, links and blockquotes; they also drag in a stray "Published on ..." line and sometimes a wrong date, so check the top of each file.

4. Then edit. This is the part that matters, and no script can do it:

   • Give it a title that states the finding, not the topic. "Why cuTile's tutorial imports fail" beats "Some notes on cuTile".
   • Write the summary. One or two sentences, aimed at a stranger deciding whether to read.
   • Paste in the things a feed post could not hold: the actual error output, the actual diff, the version numbers, a link to the PR.
   • Cut the feed scaffolding — the hook line that only made sense mid-scroll, the "thoughts?" sign-off, the hashtags.
   • Remove draft: true.

The download link expires after 72 hours, so keep a copy of the CSV somewhere safe.

Do not publish all ten at once. Go oldest-first over two or three weeks, so the site reads as alive rather than dumped.

Your site is the canonical version. LinkedIn gets a shorter, self-contained version — not a copy, and not a teaser.

1. Write the full piece in posts/, push, confirm it is live.
2. Write a LinkedIn post that delivers the insight on its own, so that someone who never clicks still got something. Around 150–250 words. First two lines carry it: the feed truncates the rest behind see more.
3. Do not paste the URL in the post body. Say "full writeup on my site" instead, and keep the actual link in your profile's Featured section, website field and About section — those are not feed content, so no reach cost.
4. Paste the LinkedIn post URL into the post's linkedin: front matter and rebuild. The article header then links back as Discussion on LinkedIn, closing the loop.

On the reach penalty, the evidence conflicts and you should know that. Several 2026 analyses claim outbound links in the body cut reach by roughly 60% and that the old "link in the first comment" workaround is now caught as well. A larger study of 1.3 million posts puts the drop at about 19% for a single body link and also finds the first-comment trick no longer helps. Nobody outside LinkedIn can measure this cleanly, and the numbers move. What every source agrees on is the part that is actually within your control: the post has to be worth reading without the click. Build that habit and the algorithm question mostly stops mattering.

Two structural things that beat any posting tactic:

• One narrow topic. Rust, CUDA, and the toolchain between them. Consistency on a single subject is what makes you findable and quotable; range makes you invisible.
• Write the post the moment you fix something. Your PR descriptions to NVlabs/cuda-oxide are already the first draft of ten articles. That is a supply of material almost nobody else has.

The layout language follows stephango.com: one narrow column (40rem), no hero — the page opens with the latest post — hairline dividers between sections, and flat lists (2026 · 08 Title) instead of cards. System font stack, so nothing loads from a font CDN. Colors are Flexoki by Steph Ango (MIT, stephango.com/flexoki): paper #FFFCF0 / ink #100F0F in light mode, inverted in dark mode following the system setting, with Flexoki cyan for links — the only colour on the page. To adjust, edit the token block at the top of assets/style.css.