I tried a different workflow this week: let the AI interview me first, then spin my answers into a story. Two big moments stood out—catching the vibe coding wave with Claude Code, and feeling the AI heat in San Francisco during Tech Week.
Vibe Coding Sends Productivity into Orbit
Trying vibe coding wasn’t a spontaneous whim. I’ve long used Copilot mainly for inline completion, occasionally switching to its chat mode, and I relied on Codex to build a few standalone features. After relentless recommendations from a teammate, I finally put Claude Code through its paces. The real trigger was a batch of side projects that needed polished, end-to-end experiences. Algorithms alone weren’t enough; I needed front-ends that didn’t feel like demos. I used to make do with Gradio, but it quickly hit its limits. With models now competent at front-end work, I decided to lean in.
Claude Code lived up to the “vibe coding” hype. The leap wasn’t just model quality; it was how the product orchestrated the whole workflow. The exact prompts that hit a wall in Copilot’s agent mode—running Claude Sonnet 4.5—went through effortlessly in Claude Code. Engineering scaffolding, context handling, interaction flow: the surrounding mechanics determine how much of the raw model power actually surfaces.
Claude Code: The First Tool to Hit My Usage Limit
For the first time ever I maxed out a $20 monthly plan. Three intense hours in Claude Code pushed me to the session limit and ate 16% of my weekly quota. Copilot and Codex never came close. The productivity jump was palpable: I shipped 5,700 lines of code in a single evening. In my most familiar domains that used to take a week. If a mid-level engineer in the Bay Area earns that output for roughly ten minutes of pay, is AI underpriced or are humans just expensive? Even in China, a mediocre outsourcing day-rate still runs north of ¥400.
High achievers getting even faster is only half the story. We also need to rethink our skill mix and time allocation. The old “wooden barrel theory” is being rewritten—if AI can patch a weak plank, stop obsessing over it. A back-end engineer who used to grind their way through front-end just to ship a solo product probably shouldn’t bother anymore; the ROI has collapsed. One vibe coding friend even used Claude Code to draft, fill out, and submit the paperwork for a Japanese visa in the browser. These tools are already capable far beyond their marketing copy.
Raising the Bar for Prompt Craft
Claude Code’s consistency comes from more than raw horsepower. I read Anthropic’s best practices post, then put together a CLAUDE.md with context, constraints, examples, and task lists spelled out clearly. Once instructions were clean, digressions and meandering responses dropped sharply. Still, it’s not a silver bullet: a few stubborn UI nits were solved faster by hopping back into Codex. A well-stocked toolbox matters, and multi-model prompt engineering is the new baseline. (I also skimmed the recently hyped Speckit—same message.)
This run reminded me of the curve ahead. The more polished the wrapper, the easier it is to overlook instruction design as real labor. But if you want AI to dive into the deep end, you have to articulate boundaries, carve out self-check slots, and ground the model in real business context. AI programming is quietly becoming an arena for the well-prepared; prompt craft is likely to regain respect.
San Francisco Tech Week: Density Meets Heat
My Tech Week calendar was packed, though I only made it to two events in the end. First up was Agentic AI: Building the Next Wave of AI Agent by Xpectrum AI. Flashy branding aside, the content felt like LiveKit wrapped in an agent shell—the same call-handling tricks we were hacking on months ago. The founder’s pitch was flat, so my friends and I ducked out early. Still, the room was full, proof that “AI agents” remains a money magnet.
The second event, Agents Everywhere: From Workflow to Connectivity, was the real draw. Four companies—E2B, Composio, AGI, and Pokee—and my second run-in with Bill this week. His take on MCP mirrored my own observations: APIs are a tough business, and marketplaces like RapidAPI haven’t exactly thrived. Whether a company exposes its crown-jewel API is ultimately a business decision; plenty of critical capabilities will stay closed, making a “tool marketplace” hard to sustain. But “AI-native tools” are undeniably a thing—most APIs today aren’t designed for agents, and aligning them to AI workflows leaves a ton of engineering work on the table.
Why San Francisco?
San Francisco has barely 800,000 residents—smaller than Shanghai’s Minhang district—and you can drive across town in under an hour. Yet the city felt supercharged all week. AI ads everywhere, coffee shop chatter circling the latest model, funding rumor, or product demo. Founders, VCs, and builders collided in the same physical space, and the density was overwhelming. If you’re curious, take a spin through the Tech Week calendar: https://www.tech-week.com/calendar/sf.
It felt less like attending events and more like absorbing a collective mood. Passion for AI, hunger for opportunity, and hands-on product grit were everywhere. Compared with other tech cities I’ve visited, San Francisco treats “bold experiments” and “quick failures” as defaults. (Yes, that line sounds like it was written by an AI—couldn’t resist.)
Closing Notes
Two events won’t deliver an epiphany, but being physically present in this era matters. For me it was a nudge: if I get to live here right now, I should keep listening, observing, and documenting. I’ll keep folding these notes into AI Weekly so you can feel the heat of the room, not just read the headlines.
Vibe coding doubled my throughput; Tech Week reminded me that AI isn’t distant thunder—it’s the background hum in every corner café. Expect more snapshots in the weeks ahead, and I hope you find your own rhythm and collaborators along the way.
Bonus
Prompt:
帮我在content/post/AI-weekly/ 建一个新的文章。我本期想讲的是vibe coding（我这周尝试了claude code，之前一直在用codex和copilot）和在三番参加tech week的感受。我想尝试一个新做法，就是你针对我定的主题先列出来一些问题，相当于采访我。我回答你的问题，并且补充一些我想讲的东西。最后你把这个东西再转化成一篇博客文章
Response:
太好了！我先按两个主题分别准备一些问题，方便你集中思考和补充。你可以逐条回答或自由发挥，之后我会根据你的回答整理成文章。
Vibe coding & Claude Code
- 你最近是怎么开始尝试 vibe coding 的？有没有什么契机或灵感？
- 和之前用 Codex、Copilot 时比，Claude Code 给你最直观的不同是什么？
- 在实际项目里试用 Claude Code 的体验如何？有没有特别惊喜或让你抓狂的瞬间？
- 你觉得 vibe coding 对创意、效率或心态有怎样的影响？有没有适合或不适合 vibe coding 的场景？
San Francisco Tech Week
- 你参加了哪些活动或 meetup？有没有最难忘的场景或对话？
- Tech Week 的氛围和你之前参加的科技活动相比，有什么独特之处？
- 在那段时间里，有没有某个趋势、话题或人给你带来新的启发？
- 这次经历对你之后的工作/学习/生活有什么影响或计划上的调整？
回答时可以随意扩展，比如想分享的具体故事、反思或者任何不在问题里的内容。准备好了就告诉我，我们再一起把这些内容整理成文章。