Bought for RMB 4.1m, Now Quoted at 2.7m: Buying Versus Renting
A Fangshan home illustrates principal, interest, sale proceeds and renting costs—with author-provided prices clearly separated from reproducible mortgage assumptions.
WRITING · 37
Articles
Essays, guides and learning series, in one place.
What Are Developers Actually Using for AI Coding?
A source-checked guide to developer surveys, product milestones, and today’s AI coding workflows—from IDE assistants and coding agents to cloud tasks, app builders, and desktop workbenches.
VLESS、Hysteria2、AnyTLS、TUIC：这几种代理协议该怎么选？
用通俗的方式讲清四种代理协议的区别、适用场景和订阅使用方法，也解释为什么多协议订阅提供的是备选线路，而不是把网速自动叠加。
Validate Before You Build
Most people build first and search for users later. A better order is to test real commitment with a tiny experiment before investing serious time and money. Four cases and a practical seven-day method.
From Java Backend to AI Agents: A Complete Path to Production Delivery
A sanitized livestream question becomes a practical transition guide: role boundaries, RAG, tool use, evaluation, production engineering, interviews, portfolio design and a 12-week roadmap for Java backend developers.
把 Dify 装进自己的服务器：Docker Compose 私有化部署实战
按真实实施顺序讲清 Dify 私有化部署：固定版本、.env 与 Secret、域名反代、HTTPS、数据库与对象存储、向量库、备份恢复、升级迁移和常见故障排查。Dify 系列教程第十篇暨收官篇。
能跑不等于能上线：把 Dify Demo 变成可运营的生产应用
从草稿与版本、DSL 交付、真实日志和逐节点排障，讲到标注回复、质量成本指标、权限隔离、灰度发布与回滚边界。Dify 系列教程第九篇。
别让模型只会聊天：用 Dify Agent 学会选择和调用工具
拆解 Workflow 与 Agent、ReAct 与 Function Calling、工具节点与双向 MCP 的边界，并用资料研究助理讲清参数描述、最小权限、停止条件和 Prompt Injection。Dify 系列教程第八篇。
别把循环当复制粘贴：用 Dify 批量处理一组数据
用一个批量内容卡片工作流讲清数组怎样进入 Iteration、items/index 如何逐项处理、顺序与并行怎样选择、失败结果如何对齐，再用渐进式润色解释 Loop 的状态、终止条件和最大次数。Dify 系列教程第七篇。
别急着把文档塞给模型：用 Dify 做一次真正的 RAG 实战
把已发布的 Dify 教程送进知识库，从清洗、800/80 分段、经济索引和倒排检索讲到向量、混合检索、Top K 与 Rerank；同时如实记录 Sandbox 额度限制，以及为什么召回测试必须先于工作流测试。Dify 系列教程第六篇。
把外部世界接进来：让 Dify 真正调用一条 API
国家信息小助手不再让模型凭记忆回答事实：这一篇用参数提取器识别国家名，通过 HTTP Request 调用需要 Key 的 API Ninjas Country API，再用 Code 解析 JSON、格式化人口与面积，并为歧义输入、鉴权失败和网络异常补齐兜底路径。Dify 系列教程第五篇。
别只让数据路过：用变量给 Dify 应用打地基
国家信息小助手的四条分支各自为政，用户问完一个国家转头就忘。这一篇用变量聚合器把互斥分支收敛成一个出口，再用会话变量 + 变量赋值器让应用第一次记住跨轮对话的查询历史，顺带把三类变量的边界讲清楚。Dify 系列教程第四篇。
给 Dify 应用装上"分岔路"：什么时候写规则，什么时候让模型判断
国家信息小助手只会回答首都，遇到问候和无关请求就露怯。这一篇给它装上问题分类器和 If/Else 两种分岔方式：语义理解交给模型判断，能写死的条件交给确定规则，用一次真实改造讲清楚两者的分工边界。Dify 系列教程第三篇。
从零搭出第一个 Dify 应用：用云版完成一次 Chatflow 上线
连不上公司内网就先别等——用 Dify 官方云版的免费额度，从注册到发布，真实走完一次 Chatflow 应用的搭建全过程：五种应用类型怎么选、LLM 节点最容易踩的坑在哪、发布后 Web App / API / Embed / MCP 四条分发路径分别是什么。Dify 系列教程第二篇。
Before You Learn Dify: Map the AI App Platform Landscape First
Before touching Dify, map the AI app-building landscape first: what really separates Workflow, Chatflow and Agent, where Dify sits next to Coze, LangChain/LangGraph, n8n, RAGFlow and China's Bailian, Qianfan and WorkBuddy, and how to choose. First piece in a Dify learning series.
The Models Arrived Together. The Agents Had Already Moved On
A September 2026 comparison of leading global and Chinese models and agent ecosystems: ranking tiers instead of decimal-point theater, correcting two common myths—SpaceX, not xAI, acquired Cursor, and Groq is not Grok—and contrasting full-stack agent platforms with models that compete for a place inside existing developer workflows.
MCP Python SDK 2.0 Is Not Just a Rename
MCP Python SDK 2.0 changes far more than class names. Using everyday analogies—phone calls versus postcards, a service counter asking for missing paperwork—this guide walks through stateless requests, the unified Client, and Resolve-based interactions, with four diagrams and a migration order designed to avoid silent failures.
Align Claude Code's 5-Hour Usage Windows: Warm Up Opus and Sonnet with Scheduled Routines
Use Claude Code's built-in Scheduled cloud agents to start five-hour windows with a minimal request before real work. Includes the observed Opus/Sonnet pool boundary, an eight-routine example schedule, setup steps, and four real-world gotchas.
Code Is No Longer the Bottleneck — It Just Moved
Reading Claude's official "The AI-native SDLC playbook," the line that stuck was: code is no longer the slowest step. Once AI sped up writing code, the bottleneck moved to approvals, testing, and release — the stages still running at human speed. A plain-language walkthrough of the six SDLC stages, what intent.md/spec.md/plan.md/CLAUDE.md actually do, a loop diagram, and three things any solo developer can borrow today.
HTTP Status Codes Quick Reference: What 401, 403, 404, and 500 Actually Mean
A practical HTTP status-code reference for users and vibe coders: identify the problem family by the first digit, then distinguish common codes such as 400, 401, 403, 404, 429, 500, 502, 503, and 504. Includes useful 2xx–5xx codes, vendor-specific codes, and a repeatable debugging checklist.
Codex CLI Mouse Wheel Cycles Input History? Restore Full Conversation Scrolling
In Ghostty, the mouse wheel may cycle Codex input history instead of scrolling the conversation. Use --no-alt-screen to restore native terminal scrollback, with steps for resuming, making it permanent, and undoing it.
Stop Repeated Codex Approvals: Complete Config and Rules Guide
Codex only: approvals, sandbox boundaries, writable roots, project trust, persistent command rules, validation, rollback, and a copy-ready prompt.
Stop Clicking Yes in Claude Code: Complete Settings and Hooks Guide
Claude Code only: complete settings, auto and allow/deny rules, a fail-closed PreToolUse hook, validation, rollback, and a copy-ready prompt.
Stop Being a Yes Engineer: Complete Permissions for Codex and Claude Code
Reading, testing, installing — every step asks again. The answer is not disabling all safeguards, but separating operating boundaries, routine passes, and hard red lines. A complete sanitized setup with Codex config.toml, prefix rules, Claude settings, programmable hooks, delegated-agent mode, verification, and rollback.
One Month, 10.6 Billion Tokens
I pulled my June Claude Code bill: 10.6 billion tokens — ten thousand readings of a classic novel. But hold the awe: 95% is cache reads. The real output is 32M tokens, roughly 240 novels' worth. The model-colored calendar tells stories too: Fable 5's wild four days, how expensive a cache-less world is, and the only three numbers worth reading on a token bill. Interactive report included.
Drive Claude Code From Feishu: Keep It Working While You Are Away
Drive your locally logged-in Claude Code from Feishu: send tasks in any project, watch progress live, resume across turns; it asks you when it needs a decision and continues once you reply. Runs on your subscription (no API cost), long-connection needs no public IP. Includes step-by-step Feishu app setup, the p2p-permission gotcha that cost me hours, and full copy-paste code.
Your Agent Demo Dazzles. Production Falls Flat.
A PM asked me how to deploy agent orchestration, naming LangGraph, Dify, Coze as the mainstream. My first reaction: these three are not even in the same race. From what orchestration actually is, to why 90% of agents die between demo and production, to the three-tier map, the control-vs-autonomy spectrum, China's two compliance walls, and a selection table you can use today — with GitHub stars verified live on 2026-06-26.
I Used Fable 5 Until It Went Dark — Ahead of Schedule
Fable 5 arrived, on-plan until the 22nd. I'm on the top tier and normally can't spend 60% of a week — this time I ran six sessions plus multi-agent workflows and drained the weekly quota in three days. Then it went offline before the 22nd. The full illustrated log, plus a thought: a model isn't software, it's fresh produce.
The Bar Didn't Drop. It Moved.
The hard part used to be execution: syntax, environments, errors. Now it's three other things — stating the problem, breaking down the work, judging the result. Same AI, and the gap between people is being amplified, not shrunk.
The "Basics" Were Never About Code
A designer who can't write a line of code shipped a tool her team actually uses — in two days. She did exactly three things: described the pain clearly, said "right" or "wrong" to each iteration, made the calls at decision points.
Segment by Scenario, Not by Person
The same person is a geek when coding, a novice when facing legal paperwork, and an expert demanding pixel-level control when writing copy. Segmentation should never cut by person — it cuts by scenario.
Prompts Are Code
Prompts are code — the runtime just happens to be a model, not a compiler. Well-structured instructions get misexecuted less. The rigor you put into your prompt is the rigor AI puts into your request.
What Survives a Power Cut
Networks drop, context overflows, tasks drift. Cut work into atomic blocks and you resume wherever it broke. Judge a system not by how fast it runs downwind, but by how much survives a power cut.
Build First, Learn Later
The traditional path is "learn properly, then build" — and students die in week three, somewhere in the syntax chapters. Now everyone gets something working in lesson one. Don't aim to understand it all at once; aim to break through, step by step.
Claude Login · Full Guide
Most students get stuck at login. Every common failure point and the matching fix.
Claude Login · FAQ
Symptom-based troubleshooting list for Claude login & Claude Code.
The Bottleneck in Voice Input Isn't Recognition
Voice tools grind recognition from 93% to 95%; nobody spends the extra 0.2 seconds on polishing. But what users want has never been "accurate" — it's "I speak, and what comes out is something I can ship."