ClassYang（互关/FollowBack） @classyang91

Blue check AI folks, let's link up and talk shop. 蓝 V AI 圈的朋友，认识一下，一起聊 AI。 I post AI jokes, takes, lists, and plain-English AI news. 我发 AI 段子、观点、清单和大白话资讯。 Your current build? 你在做什么？👇 This checkmark has to earn its keep somehow.

Per Anthropic, you can switch back anytime. Its own caveat: it "does not eliminate risk." Touching prod infra? Still review it yourself. Looking back, the most dangerous command in my terminal was always "y".

Fine print: Claude Code, 1,053 testers (paid), test setup, nothing risky ran. Worse, we faded: ~17% caught early on, ~5% after 50+ pop-ups. Starting next week, most Claude apps open new sessions in auto mode for Pro, Max and Team. Enterprise, API: opt-in for now.

Anthropic measured how closely we read the "allow this?" pop-ups. Oof. Per Anthropic, in Claude Code we say yes to 97%, but reject 39% of its plans. We read the plan. We skim the pop-up. Anthropic's lab test slipped in one rogue command. Humans caught 13.6%. Auto mode: 89%.

The critics win in exactly one case: you stop reading the code. If you can't explain a line, you don't own it, and the first description becomes the accurate one. If typing every line made you an engineer, court stenographers would be shipping operating systems.

"You just tell the AI what to type." Fine. Describe it differently. Their version: typed a prompt, accepted the diff, took a long lunch. Other version: chose what to build, vetoed the risky approach, cut a dead feature, kept prod up. Nobody paid engineers by the keystroke.

Expectation: free time to think. Reality: thinking is the only job left. AI took the easy parts, so every hour is decisions, reviews and judgment calls. That's the tiring part. Last night I dreamed I was reviewing an AI's pull request. I approved it. In my sleep.

AI took the busywork. So why do I need a nap? Expectation: AI writes the email. Reality: I reread it so it won't sound like AI. Expectation: one helper. Reality: tabs waiting for me to click "continue." Expectation: finish early. Reality: finish early, start the next thing.

Total savings this month: one streaming service I canceled to pay for all of it. What's the one line on your AI bill you can't justify out loud?

Every AI subscription I have was bought to fix the last one. Chat assistant: supposed to be enough. Coding agent: copy-pasting got old. Extra credits: the agent eats like a teenager. Second assistant: to check the first. AI notes: to remember which assistant said what.

Most seniors learned on small bugs and torn-apart drafts. That boring work was the training. Early in your career and it feels harder? It's not you. The bottom rung moved. If AI had fixed every bug I wrote in year one, I'd still be a junior. Just a very well-rested one.

25,000 fresher offers for FY27: that's the number TCS gave in April. In FY26 it hired 44,000 freshers. Its CEO: "Clarity on demand will lead to more hiring." (Per TCS, via BusinessLine.) TCS didn't link it to AI. Still, it got me thinking about how anyone gets good at this job.

The guide it shared is free and open source (MIT, over 1,000 stars). Per its own description: a guide, a starter vault, agent skills and scripts. Don't save it. Open it tonight and set up one thing. github.com/undefined-ui/s… Full disclosure: I found it in my bookmarks.

Your bookmarks folder is where AI tutorials go to die. Hitting save feels like progress. It's a gym membership with a bookmark icon. A post that opened with "Don't bookmark this" got 500+ bookmarks.

He says it ran it all: managed RAM, handled failures, stitched the frames back together. Find the tool. Staff the job. Watch resources. Handle what breaks. Ship. That's a project manager. I've had "learn to delegate" on my review for years. Claude picked it up in one prompt.

A developer says he asked Claude Code to remake a 63-second video in one animation style. It noticed he had Codex installed and figured it could use Codex's built-in image generator. So, he says, it launched 161 Codex sessions to generate the frames.

Small print: the repo's own test, per its README. 4 games, 2 blind reviewers, 1 human playtest. Not a study. Its takeaway: "the extra 29 documents bought traceability, not a better game." github.com/Donchitos/Clau… Anyway, I'm canceling the kickoff meeting for the kickoff doc.

A 26K+-star AI game-studio repo (49 agents) tested its own process. Same brief, different rigor settings. Minimal: 1 doc (the brief), then code. Standard: 30 docs before the first line of game code. Two blind reviewers and a human playtest ranked the standard build last.

"Do this" → Goal: a one-page launch brief. Context: use the attached cost and timeline. Done when: budget, launch date and decision are explicit. That's not a prompt. It's a handoff doc. Claude was never the intern. My last prompt was "pls fix." The intern was me.

keyboard_arrow_left Previous keyboard_arrow_right Next

Stop treating Claude like an intern. "Fix this" → "Blank email crashes login. Show a validation error. Don't touch the CSS." "Make it better" → "Shorter, less passive-aggressive. My boss reads it." "Better" just buys you exclamation points.