Blog
Data pieces, builder interviews, and process notes sourced from the directory.
GeneralAI quietly fixed translation, but nobody really talks about itOne post for the other side of the ledger. The old Google Translate gave you English words in English order that meant nothing, and you came away more confused than you started. That problem is now mostly solved, and it happened without a headline.
TutorialsHow to fix Claude Code desktop crashing without running Windows RepairIf the desktop app dies every time you use the browser and only Windows Repair brings it back, the install is fine. An orphaned process is holding a file lock, and one taskkill command fixes it in a second.
LessonsHow to run a Vibe Coding session that doesn't go off the railsAnybody can fire up Claude Code and start typing. The people who get clean results are running their sessions with a bit of discipline. Here are the habits that keep a vibe coding session from turning into a slow-motion mess.
NewsClaude topped the AI satisfaction rankings, but the survey mostly measured an older ClaudeYouGov put Claude first at 56.2, ahead of Gemini and ChatGPT, and best of any assistant among former users. The survey ran from March to July. Opus 5 shipped on July 24, which makes this a score for the old models.
LessonsWe stopped writing the code. Now we’re starting to stop checking it tooFirst we let AI fetch the answers, then write the code, and now we hand its work to a second AI to review. That last delegation is different from the other two, because checking was where a gap in your understanding used to become visible.
GeneralClaude is supposed to be smarter than ever. So why does it feel worse?Opus 5 and Sonnet 5 sit near the top of the rankings, and longtime users say working with them is exhausting. Both are true, and Anthropic wrote most of the complaints into the release notes as intended behavior.
NewsHow Claude's text watermark actually works, and why people are cancelingAnthropic explained the mechanism two weeks after switching it on. It only changes where a random number comes from, which kills the quality complaint. It also means the people angriest about it are the least marked.
TutorialsHow to write a great CLAUDE.md file (with examples)A CLAUDE.md is the highest-leverage file you can add to a Claude Code project, and most people either skip it or bloat it. Here's what to put in it, where to put it, how to write instructions Claude actually follows, and how to keep it lean.
GeneralThe next big AI business may be a tool that makes machine writing impossible to detectClaude now watermarks its text and detectors flag most of what gets posted. Into that sits an unbuilt business worth an absurd amount: a tool that writes in any specific human voice, undetectable, no watermark. Here's why it burns down the market it sells into.
GeneralIs the real reason AI companies are adding watermarks to keep AI content out of training data?Every Claude output since August 2 carries a hidden mark that survives copy-paste. The stated reason is transparency for users, who cannot detect it yet. Anthropic can.
ToolsClaude Code was miles ahead but Codex and Kimi caught upA year ago there was no argument about which coding agent to use. I've spent weeks on Codex and the newest open models, and the gap closed. The harness got copied, the models matched it, and the constant switching is doing the rest.
Build logsWhy you should put Cloudflare in front of your websiteCloudflare blocked 22,471 threats. Here's what's actually hitting a new site, and the setup step most people skip.