Blog
Data pieces, builder interviews, and process notes sourced from the directory.
ToolsWhat MCP actually is, and which servers are worth addingThe Model Context Protocol connects Claude Code to outside tools, and it's the part of a setup most likely to cause you a problem. How the three scopes work, what a stdio server really does, and how to be selective.
ToolsWhen to use a Claude Code subagent, and when not toA subagent has its own context window, so delegating a big read keeps eleven files out of yours. Three things that surprise people, including that rewind cannot undo a subagent's edits.
TutorialsEvery way the Claude desktop app crashes on Windows, and what the log saysThe file-lock crash has siblings. The GPU crash from the in-app browser, the 'This app can't open' package flag, the service that blocks self-repair, and the CLI freeze, each with its log signature and a fix shorter than Windows Repair.
GeneralIs it time to nationalize AI?Dario Amodei wants the frontier slowed by regulation, government referees, and embedded evaluators. Hacker News read the essay and asked the question it skips: if the danger is real, why is this still a private company? Why the answer is still no, and what's actually being proposed.
LessonsHow to read code you didn't writeReading code is a separate skill from writing it, and building with an agent means your codebase is now bigger than your reading experience. The approach that works, which is not starting at line one.
GeneralShould Google pay publishers for AI citations?Search used to trade content for traffic. AI Overviews kept the content and stopped sending the traffic. This site drew 500 clicks from 38,492 impressions last quarter, and Google already counts every one of the impressions it doesn't pay for.
TutorialsClaude Code won't reopen on Windows? Check Task Manager for ssh.exeThe third orphaned-process crash in the Windows desktop app. If you use Claude Code on remote servers, a hung ssh.exe survives the app closing and holds the pipe the next session needs. End it in Task Manager, then stop it happening with one ssh flag.
GeneralThe risk behind letting Claude or ChatGPT manage your inboxYour inbox is the account that resets every other account. An email with hidden text can instruct the assistant reading it, researchers have already used that to exfiltrate Gmail, and both Anthropic and OpenAI say there's no clean fix. What I'd hand over and what I wouldn't.
LessonsPrompt injection: how a README can hijack Claude CodeA sentence in a README, an issue or a web page can address your coding agent instead of you, and a coding agent can run curl. What Claude Code does about it, what turning off permission prompts removes, and the posture that survives a real workday.
TutorialsClaude Code error messages, explainedThe common Claude Code errors with the exact message text, what actually causes each one, and the fix. Includes a desktop app crash that has no writeup anywhere else.
NewsFable 5.1 is out: smarter, cheaper, and less trigger-happyAnthropic released Fable 5.1 and Mythos 5.1 with big agentic gains, a 75 percent cache-read cut, and safety classifiers that finally fire less on legitimate work. The one thing the announcement doesn't mention is subscriptions.
NewsClaude Code's permanent 25% limit raise is actually a 17 percent cutAnthropic announced weekly limits rising 25 percent from September 14. Everyone is currently on a 50 percent promotional increase, so the lived change is minus 17 percent.