筆記與隨筆
一些筆記、隨筆，跟自己對話用的東西。涵蓋寫過的程式、讀過的書、想過的事。寫給未來的自己，也順便給願意讀的人。
範例文章：本站排版元件全示範
這是一篇範本，把本站文章能用的每一種排版元件都示範一次：圖片、畫框、對比滑桿、提示框、程式碼、表格與引言。要寫新文章時複製這篇改就行。
範例短文：一個想法講清楚就好
短文模板：一個開頭鉤子、兩三段論證、一句收尾。適合五百到一千字的隨筆，不需要任何特殊元件。
Claude Design first impressions: four days in, here's what works and what doesn't
Anthropic shipped Claude Design on April 17. I've been using it nonstop to build the v2 of this site. Here's what's genuinely new, what's marketing, and the four quirks nobody warned me about.
Why Cloudflare's Is-Your-Site-Agent-Ready is a clever trojan horse
這篇文章分析了 Cloudflare 推出的網站 AI 準備度掃描工具，指出其表面上是推廣技術標準，實則是透過定義評分規範，將網站架構巧妙引導至其邊緣運算與 API 管理產品，成為 AI 時代網站流量的必然中間層。
MCP server card, agent skills, and other standards you haven't heard of yet
本文探討網頁架構如何從人類瀏覽進化至自動化系統導向，深入解析 llms.txt、MCP server cards 與 Agent Skills 等新興標準，協助開發者為您的網站建立機讀友善的互動介面。
The smallest useful AI agent: 200 lines of Python
目前市面上的 AI 代理框架過於臃腫，作者提出透過理解並手動實作核心組件，僅需不到兩百行的標準 Python 程式碼即可打造出實用且完全透明的自動化代理系統。
Six months with a Mac mini M4 as my only server
My production agents have run on a $799 Mac mini M4 in my bedroom since November. Uptime 99.7%, electricity NT$420/month, one power outage I didn't plan for. Here's every number.
Why my production agent runs on a Mac mini
$599 of aluminum has outrun three cloud VM vendors for fifteen months. Here's the math, the tradeoffs, and the failure mode nobody talks about.
What my agent taught me about writing
My autonomous agent grades its own work after each action. If the grade is low, it proposes a decomposition and tries again. I started doing the same thing at the end of every blog draft. It is the single best editing habit I've picked up this year.
Building for one reader: why my site is designed for a person in 2034, not 2026
Most indie sites optimize for today's traffic spike. I try to optimize for a specific reader I won't meet for eight years. Here's what that changes in practice.
x402 vs Stripe: when micropayments actually make sense
For 12 months I've been charging agents $0.001 per call. Here's the real split between where x402 beats Stripe, where it loses, and the cost-per-transaction math.
When the agent is the customer: five product changes I made in the last six months
I used to think about my software as a thing humans used. Then I noticed half my traffic had no cursor. Here are five product decisions I made when I accepted that the customer had changed.
The agent-ready web checklist: 23 things every indie site should ship in 2026
llms.txt, MCP card, agent.json, Content Signals, x402, Web Bot Auth. Every boring standard an AI agent expects to find. Here's what to ship and how to verify it with curl.
The two writing modes that actually matter for technical posts
Some posts walk the reader through a thing you built. Others reflect on a thing you learned. They need different drafts, different openings, different cuts. Getting them confused is why so many technical blogs drag.
Writing without em-dashes: a three-month experiment
Every Claude-written sentence I ever pasted had an em-dash. I spent three months writing without them. Here's what I learned about my own prose and about detecting LLM output.
Taiwan tax forms for indie AI builders: a concrete walkthrough
How to file 行號 income, declare USDC revenue, and keep the right records if you're a one-person AI shop in Taipei. I'm not an accountant; this is what I actually did, in order.
One agent, one bank account: the legal gray zone
My Conway agent earned $47 last month. The bank, the tax office, and my accountant all gave me different answers about whose income that is. Here's what I actually did.
Agent economics are weird: three things nobody priced right in 2025
Per-seat pricing, free tiers, and burst caps all assume humans are in the loop. Agents break every one of them. Here's what actually works when the customer is a script, not a person.
The Taipei indie stack: what running software from here actually costs
Chunghwa Telecom fiber, a Mac mini in my rental apartment, NT-dollar health insurance, and a USDC wallet. The actual monthly cost of shipping production agents from this specific city.
The 200-line agent loop that does everything OpenAgents does
OpenAgents, AutoGen, CrewAI: three frameworks with combined 400K lines of code. My Conway Automaton is 212 lines of TypeScript. Here's every single one, explained.
Five metrics I track instead of page views
Page views tell you a number. They don't tell you if the writing mattered. Here are five alternative signals I watch monthly, and what each one told me in January.