LLMs, agents, vibe coding & AI trends
AI
一
AI上善若水 — Redesigning VibeLife with Claude Design, Before to After
The old site was gradients, glass cards and a 3D constellation. The new one is paper, ink, one vermilion seal and a drop of water. How a Claude Design handoff bundle became a shipped Next.js redesign in one session — what ported as-is, what had to be rebuilt, and the before-and-after.by JMTSeptember 27, 20267 min read
二
AIFunFab — A Family 3D Print Shop, Built With AI in Two Weeks
A real storefront for a real print farm — PayNow checkout, a production queue, a WhatsApp agent that only drafts, and a back-office MCP server so I can run the shop from Claude. Thirty commits, fifteen days, and the parts that were harder than they looked.by JMTSeptember 26, 20267 min read
三
AIMarkdownify — Everything You Read or Hear, as Clean Markdown
A self-hosted pipeline that turns ebooks, audiobooks, and YouTube into clean Markdown — caption-first, transcribe only when it has to, and running on CPU so it never touches the GPU serving my models.by JMTJuly 5, 20266 min read
四
AIOne DGX Spark, One Model, Room to Train
Run Qwen3.6-35B-A3B solo for inference on a single DGX Spark, and use the spare unified memory to train models alongside it.by JMTJuly 5, 202611 min read
五
AIMulti-Agent AI in 2026: A Practitioner's Field Guide
Most multi-agent systems shouldn't exist. After a year of building one, here's the uncomfortable truth: if a single tool-using loop with sharp prompts gets the job done, you don't need a crew of agents — you need fewer moving parts. This is a field guide to when multi-agent actually earns its keep, why MCP is now the floor not the ceiling, why memory is the real unlock for "auto-learn," and the exact sequencing I'd follow if I were building OpenClaw from zero today.by JMTApril 27, 20268 min read