A personal "morning paper": a single page that shows what six tracked people posted in the last 24 hours across X, YouTube, and one Substack — short clips, offhand tweets, replies, and posts you might otherwise miss. Built around the explicit rule: keep the small details, don't filter for "substantive" content.
Live URL: https://out-mu-pink.vercel.app
Every page load runs a Vercel Python serverless function that fetches the feeds in parallel, parses RSS/Atom XML, filters to the last 24 hours, and renders the editorial HTML on the fly. The CDN caches the response for 5 minutes (with stale-while-revalidate up to 15 minutes), so most visits are instant.
A Refresh button in the masthead bypasses the CDN cache (via ?t=<timestamp>) and swaps fresh content into the DOM in place — no full page reload.
Visitor → / → rewrite to /api/news → Python function
├─ parallel-fetch 11 feeds
├─ filter to last 24h
├─ render HTML
└─ return with Cache-Control
Configured in api/news.py (the PEOPLE constant). Channel IDs are pre-resolved.
| Person | Surfaces |
|---|---|
| Andrej Karpathy | X (@karpathy), YouTube |
| Theo (t3.gg) | YouTube, X (@theo) |
| NetworkChuck | YouTube, X (@NetworkChuck) |
| Nate Herk | YouTube, X (@NateHerk) |
| Elon Musk | X (@elonmusk) |
| Dean Ball | Substack RSS, X (@deanwball) |
X feeds use Nitter (https://nitter.net/{handle}/rss) with xcancel.com as fallback. YouTube uses the Atom endpoint https://www.youtube.com/feeds/videos.xml?channel_id={id}. Dean Ball's blog uses his Substack feed directly.
~/Desktop/newsDebrief/
├── README.md # this file
├── deploy.sh # `vercel deploy out --prod --yes --scope godfreyps-projects`
├── sources.json # legacy: kept for reference; function has its own PEOPLE list
├── out/ # what gets deployed to Vercel
│ ├── api/
│ │ └── news.py # the serverless function (fetch + parse + render)
│ └── vercel.json # rewrites `/` → `/api/news`
└── .vercel/ # vercel project link metadata (gitignored if repo'd)
The Vercel project is godfreyps-projects/news-debrief.
Editorial / Notion-Linear minimalism, applied via the taste-skill skills (redesign-existing-projects + minimalist-ui).
- Warm bone background (
#F7F6F3), charcoal ink (#2F3437) - Serif display heading (system serif: New York / Iowan Old Style)
- Sans body, mono for metadata (timestamps, eyebrow, footer)
- Type pills with muted-pastel backgrounds (yellow for blog, red for YouTube, neutral for X)
- Real red YouTube logo (inline SVG) for YouTube items; unicode
𝕏and✎for X and blog - Relative time chip per item (
3h,12m,now) with absolute CT timestamp on hover - Section stagger fade-in on page load (
prefers-reduced-motionrespected)
The frozen template (CSS + per-item HTML) lives in two places that must stay in sync:
out/api/news.py— the function. Source of truth for what's served.~/.claude/commands/daily-news-debrief.md— the Claude Code slash command, kept around for manual rebuilds and as documentation.
If you change the design, update both.
cd ~/Desktop/newsDebrief
bash deploy.shRequires vercel CLI logged in. The script bakes in --scope godfreyps-projects (Vercel CLI requires explicit scope in non-interactive mode).
Edit the PEOPLE list in out/api/news.py. Each entry is (name, [(kind, key), ...]):
("x", "<handle>")— Twitter/X handle without the@("youtube", "<channel_id>")— must be theUC…channel ID, not the @handle("blog", "<full RSS URL>")— for Substack and similar
To resolve a YouTube @handle to a channel ID:
curl -sL "https://www.youtube.com/@<handle>" | grep -oE '"channelId":"UC[A-Za-z0-9_-]+"' | head -1Mirror the change in ~/.claude/commands/daily-news-debrief.md so manual runs match.
A /schedule skill routine fires daily at 8am CT (0 8 * * * America/Chicago). The routine pings the live URL with cache-busting so the morning's first visit is fresh, and serves as a daily reminder. (Not wired up yet — see Task #6.)
- X-RSS reliability is bad in 2026. RSSHub (
rsshub.app) is gone (returns Google 404). Nitter (nitter.net) returns valid RSS for some handles and empty/200 or HTML errors for others, varying by IP.xcancel.comperiodically 403s. The function falls through both providers and renders a per-person fetch-error row when both fail; the rest of the page keeps working. - YouTube RSS endpoint is intermittent. Has 404'd for all channel IDs simultaneously across the entire RSS service at least once. When this happens, every YouTube section renders a fetch-error row.
- 24h window is recompute-on-render. The function uses
now - 24h, so if you visit at 8am CT and again at 6pm, the windows are different (and items can roll out of frame). - No persistence. The function doesn't store anything — every render is fresh from the source feeds. There's no archive of past mornings.
- Hobby-tier serverless. 10s timeout per invocation. Parallel fetch with 6s per-feed timeout keeps total under 7s in practice.
Started as a Claude Code session in early May 2026 to mimic the OpenClaw-style "agent that builds you a news debrief" pattern, but using only Claude Max + Claude Code (no API key, no third-party agent infra). The first iteration was a static HTML deploy by a slash command; iteration two added editorial design via taste-skill; iteration three migrated to a server-rendered function with a refresh button so the page stays live throughout the day.