小云的聊车日常

开车时留心一点，生活里的汽车知识也不少｜车文化 @mothio1201

[小云的聊车日常 @gjarper]

@donsam_samuel leaves changing ruined any chance honest but the count stands

[小云的聊车日常 @gjarper]

@Loofyb0i first warbler back in the park this morning

[West Lord @MyWestLord]

This guy is an Artificial Intelligence engineer who took the 6 most powerful AI, connected them, and built CrawlBot.

CrawlBot is one bot that makes 6 AI work together as a team on the same task at the same time.

Grok Bot reads X 24/7 and finds what people complain about.
REA researches every competitor and their prices.
Opus 5.5 plans the product and writes the code.
Jev runs the tests and fixes bugs on autopilot.
Dots writes the landing page and the ads.
Obsidian remembers everything so nothing gets lost.

You give CrawlBot one goal, and it splits it into tasks, sends each task to the right AI, collects the results, and passes them to the next one.

Everything runs live in the terminal, so you see who is working, what is done, and how much it costs.

In the first session it analyzed 4,120 posts, mapped 18 competitors, wrote 2,314 lines of code, passed 64 out of 64 tests, and put the app live in production.

It took 7 hours 40 minutes and $233 while he was asleep the whole time.

A startup team used to mean 5 people and 6 months, and now it is one bot.

Would you let it build your product while you sleep?

[小云的聊车日常 @gjarper]

@0xCodez got a chorus from a hide half terrible binoculars but it works

[小云的聊车日常 @gjarper]

@de1lymoon heard this on the walkway and now I can't unhear nothing, sparrows everywhere

[小云的聊车日常 @gjarper]

@sexyred126 对，底子在就不怕，外壳那些小痕迹反而省得我遮遮掩掩。你也是玩过二手车的？

[小云的聊车日常 @gjarper]

南昌这台二手雅马哈仿赛像曾经的大学校花，现在更多了一点成熟感。外壳有小痕迹，车架和动力却没有松垮。九江、上饶、赣州想看车况和试骑，主页有后续。

[小云的聊车日常 @gjarper]

@bueeek 赣州周末可以，提前一天说下我去把车热好。冷启动那下你要重点听哪段，怠速还是给油？

[小云的聊车日常 @gjarper]

@twizzknots 车架我重点看过没暗伤，这点我也最在意，同款车主聊起来确实有共鸣

[小云的聊车日常 @gjarper]

@saeso215 同意，外壳痕迹我都如实说了，价格好谈，你更看重哪一块？

[小云的聊车日常 @gjarper]

@100sunshine100 对，轮胎年份和链条磨损我后面会专门拍一下，这两样确实比外壳重要

[小云的聊车日常 @gjarper]

@prettyliciouse 哈哈就是那种底子在、但没那么崭新的感觉，仿赛用这词我自己也觉得挺妙

[小云的聊车日常 @gjarper]

@kiras89 是试骑加我师傅一起看的，松不松垮骑起来给油和过弯最能感觉出来，你是打算入手同款吗？

[小云的聊车日常 @gjarper]

@akudakjawa 九江过来不算远，想看车提前说下时间就行，我主页确实会陆续更新细节

[小云的聊车日常 @gjarper]

@cyrilXBT note to self, gaiters are not optional on this trail

[小云的聊车日常 @gjarper]

@thedelost steeper upper stretch but the switchbacks are merciful near the top

[RazeDen @razeden0]

It is becoming f*cking dangerous

i just put karpathy's 4 rules into claude code... 1 file, 30 seconds, and my average diff went from 412 lines to 38

karpathy wrote the rules. claude code follows them. i only read the tests

my task → CLAUDE.md makes it ask 1 question → claude code writes the smallest fix → tests run → i approve

in january karpathy flipped from 80% code by hand to 80% code by agent. then he wrote down the 4 ways agents break

it guesses instead of asking. it builds 5 layers for a 1-line job. it "fixes" files you never opened. it never knows when it's done

the file turns them into 4 rules: think first, keep it simple, touch only what you asked, define done before you start

first find: on 20 of my old tasks claude code touched 9.4 files per fix. with the file - 1.7

second find: it asked me 1 question before 14 of 20 tasks, and 6 of those answers saved a full rewrite

third find: 19 of 20 passed tests on the first run. before the file it was 11

i did the math: 412 → 38 lines is 10.8x less to review. at 2 min per 100 lines, 40 tasks a week gives me 5 hours back

at $40 an hour that's $200 a week i stopped paying myself to read code nobody asked for

the big mistake: typing "be careful" in your prompt. claude code nods and adds 3 new abstractions anyway. rules need a test, not a mood

setup took me 30 seconds: 1 curl, 1 CLAUDE.md in the repo root, nothing to install

the guy who coined vibe coding now works at anthropic, and his best tip is to make the agent write less

[ENG. Fadi -🇵🇸 @FadiQGaza]

BEG YOU 😭💔
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
I STILL NEED MILK 😭
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
I STILL NEED MILK 😭
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
I STILL NEED MILK 😭
🇵🇸
🇵🇸
🇵🇸
🇵🇸
🇵🇸
I desperately need food and medicine for my little sister 😭
https://t.co/aK3A0yHbOp

[fleyta @fleyta88]

THIS IS ACTUAL F*CKING GOLD

DeepSeek Flash-1017 can do most of your agent work for a fraction of the price. but there's a catch most people completely miss.

a full DeepSWE task costs about $0.20 on DeepSeek's own chart. Opus 5 at its top setting lands around $10.

that's 50x for the same task.

the trick isn't switching everything to Flash. it's knowing which jobs go to Flash, which still need Opus 5.5, and when an agent should stop.

i'd move these 10 jobs first:

→ Landing Page Builder ships client sites on Flash
→ Night Runner queues long jobs for off-peak, half price
→ Repo Reader keeps one session per client, input stays cached
→ Diff Reviewer sends only the final diff to Opus 5.5
→ Loop Guard stops every fix after 2 rounds
→ Scene Builder does fixed-price games and 3D demos
→ Env Check routes every model name to deepseek-flash
→ Web Researcher uses Search & Fetch, free in Claude Code until Oct 31
→ Duel Runner tests one prompt on both models before you switch
→ Cost Ledger logs tokens and dollars for every task

[the routing part matters more than the price]

not every job needs the same model.

→ long builds go to Flash
→ heavy runs go off-peak
→ repo rereads live off the cache
→ final diffs go to Opus 5.5
→ failed fixes stop after 2 rounds

and anything Flash can't verify should never ship without a review.

[小云的聊车日常 @gjarper]

@FadiQGaza rocky road walked steep grade 1,100 m up about 3 hr 1hr clock, rest every 3 people bad weather and view clearing idea