Library

112 Days Late: I Built a Product Anthropic Had Already Shipped

A field report with the full data. I spent 22 days and 215 commits building a macOS app to manage Claude Code sessions — Anthropic had shipped most of it 112 days before my first line, and my one competitive check scoped itself around the mobile line and missed the desktop redesign entirely. Includes the 48-feature audit (18 replaced / 15 overlapping / 15 unique), the review economics (~190 agents for ~170 findings, 33% hit rate at one stage), six dated incidents including a symlink that destroyed an 823MB corpus, and a full data appendix with timestamps.

我的 app 有多少功能被 Claude 官方吃掉了 —— 一次逐条对账

自研 macOS 菜单栏 app 与 Claude Code 官方能力的逐条对账:48 条功能里 18 条已被替代、15 条部分重叠、15 条仍独占。核心发现是 2026-04 的桌面端重做吃掉的是 Mac 主线(多 session 侧边栏、每 session worktree、集成终端、预览窗格、环境选择器),而上一次竞品核查只查了手机线因此漏掉了它;剩下的护城河按含金量排是跨品牌抽象、看得见外部终端 session、审批台三条。含方法反思与未核实清单

Multiple Sessions or Multi-Agent? I ran 27 agents to answer that

An orchestration study with real experiments: 6 probes to pin down where a subagent's isolation boundary sits, and 13 agents building the same game three different ways, each build verified in a headless browser. The counterintuitive result — multi-agent buys width inside one task, multiple sessions buy length across time and total context; they're orthogonal axes. What actually decides whether parallel development works is an interface contract everyone follows (without one, 2 of 5 assemblies went straight to a blank screen).

An afternoon on Supabase: pgvector, a 44x query plan, and two RLS lessons

Supabase 实战英文复盘:pgvector 44x 查询优化与 RLS 双闸门

我的网站没有后端:拿 mingyuyang.com 讲透 Supabase

《Supabase 面试备考指南》实战篇:本站零后端接口的真相——resources 表即全站内容、PostgREST 自动 API、六行代码的 Google 登录、两张表两种 RLS 配方、匿名探针实证(公开表 78 行 vs 私有表空数组)、两把钥匙的管理纪律,外加四个真实踩坑和一段两分钟面试陈述逐字稿;所有代码原文摘自本站仓库

Supabase 面试备考指南:架构 · RLS · Auth · 一周学习路径

把「会用」升级成「能讲」:database-centric 架构、RLS 四种策略与性能优化六招、JWT 与新旧 API Key 体系、Realtime 逻辑复制原理与扩展限制、Supavisor 连接池两模式,配 28 道可折叠自测的高频面试题和 7 天学习计划;关键数字全部核对 2026-07 官方文档

长上下文如何失效

先看作者要挑战的那个流行假设:「上下文窗口越大,agent 就越强」。

RAG 全景图:核心概念与完整流程

从知识源到答案输出的 11 个阶段一次看全:切块、向量化、检索、重排序、上下文组装、生成与评估,外加 6 种高级范式、12 张概念速查卡与常见指标解释

DSpark 论文精读:置信度调度的半自回归推测解码

DSpark 论文精读

iOS App 第一次出海:归因与投放接入入门指南

整理自 Adjust 线上课堂:iOS 归因为什么难、ATT 与 SKAN 两套独立框架怎么理解和配置、Google(ICM)/ Meta(MAE/AEM)/ TikTok(SAN)三大渠道接入路径、归因窗口对齐建议与该看的指标,收尾附首次出海接入七步清单

Adjust Web to App Campaign 流程及设置:全渠道配置指南

从「为什么要做 Web2App」讲起,逐渠道拆解 Google、Facebook、TikTok、Kwai、Snap / Pinterest / Bing 的广告流程、Adjust 后台配置、跟踪链接拼接与数据回传设置,配动态流程图讲清每条链路的归因原理

Adjust Redownloads设置全解析：让数据更真实，用户价值更高

今天我们就来拆解 Adjust 的 Redownloads 功能，帮你轻松理解这些复杂的数据逻辑。

ODDL内部培训：移动广告归因与数据对接

ODDL 培训笔记

Adjust数据差异排查指南：从日志到归因

Adjust 数据排查指南

Adjust ODDL：海外广告投放与数据监控

Adjust ODDL 海外投放

D8 · Hooks

For the first seven lessons, every way you influenced Claude was "saying it to the model": CLAUDE.md, plans, Skills. Hooks take a different road — no…

D1 · The agentic loop and what Claude can actually read (/context)

No configuration tricks in this lesson. It answers just two questions: how Claude Code gets a task done (the agentic loop), and what it can and canno…

D2 · The context window & prompt caching

Every message you send makes Claude Code resend the entire conversation from the very beginning. This lesson explains why that works: the window's th…

AI 每日简报 · 2026年9月11日

AI 每日简报 · 2026年9月9日

AI 每日简报 · 2026年9月7日

托管 Agent 运行时:OpenAI Agents API 深度解析 | Managed Agent Runtime: A Deep Dive into the OpenAI Agents API

Agent Swarm 大规模并行智能体 | Agent Swarms at Scale

Harness Engineering:智能体运行时工程 | Harness Engineering: The Agent Runtime Discipline