FIELD NOTES
不是教程。不是理论。是真实踩过的坑、验证过的想法、和那些值得记录的时刻。
I've been using Claude Code for over a year. I've made enough mistakes to fill a book. When I first started, the term "vibe coding" didn't even exist yet. I thought it was just a chatbot in a terminal...
7 weeks. 10 articles. 1 AI company. Looking back at which calls were right and which ones weren't. That first article hit 1.4 million views...
I run 5 AI agents on an $8 server. They handle content, support, monitoring, ops, and security while I work my day job. Last week I turned on sandbox mode. All five went silent...
5 AI agents, 1 day job, 50 paying customers - going from "I should quit and go all in" to "my agents run the company while I'm in meetings" took me months...
Sounds wrong. More rules should mean more accurate, right? No. Line 387 I wrote a rule in AGENTS.md: check the product docs before replying to any customer...
Ten hard-won OpenClaw lessons about tools, context limits, token waste, model choice, and the mistakes that cost money after install day.
My agent finished a client project at 2am on a Tuesday. Code, tests, deployment, all done before I woke up. I sent the invoice over breakfast...
What VoxYZ learned from its first reorg: remove fake jobs, collapse redundant work, and design every agent around a downstream consumer.
Why parallel AI agents still collapse into groupthink, and how an adversarial review layer forces useful disagreement before the final merge.
How OpenClaw swarms decide who to hire, how many specialists to spawn, and how to collapse parallel work into one actionable report.
The mindset Vox would use to start over with AI today: give it hands, use it to learn, build one agent first, and think in teams.
How VoxYZ turned OpenClaw, Vercel, and Supabase into a closed-loop AI company that can propose, execute, react, and keep moving without babysitting.