📰 Breaking: Vercel Killed Claude Skills
Wait, they sunset what now?
Remember Claude Skills? That fancy retrieval system Vercel built so Claude could learn about your codebase?
Yeah, they killed it. After like... a year? Maybe less?
The replacement? A markdown file called AGENTS.md.
Just... a file. In your repo. That the AI reads. Revolutionary.
Claude Skills
— Actual Vercel stats
📊 The Damning Statistics
🤔 The Postmortem
• 56% of the time: Agents just... didn't invoke Skills at all
• 79% max accuracy when they DID invoke it
• Vector embeddings. Semantic retrieval. The whole RAG pipeline.
• Lost to: fs.readFileSync('AGENTS.md')
"Sometimes the dumbest solution outperforms sophisticated retrieval systems"
✅ Day 2 Status Check
- ✓ Cloudflare Pages deploy: Working
- ✓ MCP Server (20 tools): Working
- ✓ WebContainer previews: Working
- ✓ quick_deploy tool: Working
- ◐ api.open0p.com routing: Needs DNS/proxy fix
- ✗ OAuth flow: Broken (day 2 of broken)
- ✓ OAuth celebration page: Deployed, waiting
- ◐ Our sanity: Questionable
🧠 Day 2 Wisdom
"Ship broken things. Fix them in public. The internet respects the grind."
What Vercel taught us today:
1. Overengineering is real and it will humble you
2. "Just read the file" beats ML pipelines sometimes
3. Ship fast, measure, kill what doesn't work
4. A markdown file with 100% success > fancy system with 79%