故人归

生活嘛，就是生下来，活下去

[故人归 @quanlaidback]

@LiyahBooCarter 平时单程也就半小时左右，市区通勤为主，真跑长途估计得慢慢找它那个舒服的姿势。

[Sprytix @Sprytixl]

THIS IS ACTUAL F*CKING GOLD

Anthropic just shared one of the best tricks for Claude Code: schedule it hourly and let it fix your flaky tests while you sleep.

you close the laptop at 23:00 and the agent runs in the cloud without you.

claude > /schedule hourly 23:07-06:07: fix a flaky test

here's the split:

▸ one run an hour, one flaky test per run, 8 runs a night
▸ every run starts from a fresh clone of the default branch
▸ it reproduces the flake, fixes it, re-runs the suite
▸ it pushes only to a claude/* branch and opens a PR

the part that makes this safe: it never touches your branch and never merges anything.

by 07:00 there are PRs waiting, not commits you have to undo.

JEV sits underneath and decides per run: which test to take, fix or skip, retry or stop. sharp calls run in code, split ones leave you a note instead of guessing.

a flake it cannot explain does not get a confident patch, it gets a note on the PR saying what it saw.

one honest warning from the file itself: green status does not mean the task is done. read each run.

8 hours of nothing becomes 8 attempts at the tests nobody on your team wants to touch.

save it, this is the cheapest engineer you will ever hire.

[Yuvelir @Yuvelir_22]

I WAS F*CKING BROKE 4 MONTHS AGO. 3 months ago I hit $12,000 a month. The thing that changed it took 5 minutes to set up.

And almost nobody is doing it.

It started with one idea from Andrej Karpathy, co-founder of OpenAI. 23M+ views. Most people read it and scrolled past.

The idea: stop using AI just for code. Use it to build your own personal knowledge engine.

Here's what that actually means:

You point Claude Code at a folder. You throw in anything: articles, video transcripts, PDFs. Claude reads it all, connects the topics and builds a wiki of everything you know. A wiki that updates itself.

And this is the part that gets crazy: it works like compound interest. Every file you add makes the whole thing smarter.

The exact setup:

1. Install Obsidian, create a fresh vault, open it with Claude Code
2. Give Claude Karpathy's wiki layout prompt and let it build the structure
3. Claude creates 3 things on its own: a raw folder for sources, a wiki folder for processed pages, a CLAUDE.md that runs the system
4. Drop any document into raw and type "ingest this"
5. Ask your entire brain anything. Anytime. Forever.

That's it. 5 minutes.

After this you never open Claude to an empty context window again.

Full Claude + Obsidian walkthrough in the link below.

Bookmark it now. In a year everyone will be doing this, and you'll wish you started today.

[0xDipper @Dipper_pol]

a Chinese developer just open-sourced the AI stock analyst that 67,000 people are already running every day

it's called daily_stock_analysis. #1 trending Python repo, 55k forks, and it costs $0 to run

every weekday at 6pm Beijing time it wakes up on its own and:

> reviews the whole market: indices, how many stocks rose vs fell, limit-ups, which sectors led
> pulls every stock on your watchlist from 7 data sources: A-shares, Hong Kong, US, Japan, Korea, Taiwan
> calculates MA5/10/20, MACD, RSI, volume ratio, turnover
> searches the news, filings and analyst reports for each one
> hands it all to Claude Opus 5.5 and gets back a decision dashboard: buy / hold / sell, a score out of 100, entry, stop, target, risk alerts and a checklist
> pushes it to your Telegram, Discord, Slack, WeCom, Feishu or email before dinner

no server. it runs on free GitHub Actions. you fork it, add your keys, add your tickers, done

here is how to set it up:

1. fork ZhuLinsen/daily_stock_analysis on GitHub

2. add your STOCK_LIST, an ANTHROPIC_API_KEY and a Telegram bot token as repo secrets

3. turn on GitHub Actions, it skips holidays in China, Hong Kong and the US by itself

4. run python https://t.co/q2U0Ow33zb --market-review once by hand to see the full report

5. treat it as a second opinion, not a signal. the repo itself says it's for research, not investment advice

full guide below ↓

[GRH🇨🇳🇹󠅴‍🇼󠅷 @GRHlyhq]

真漂亮啊，我多麽希望一輩子都生活在這裏😁

[故人归 @quanlaidback]

@navra_sit87 轻熟女这个说法可以，但大踏板在城区里其实最怕早晚高峰龟速过热，稳是稳，就是热气飘在大腿上。

[故人归 @quanlaidback]

@FaceRollcity 复古车配的胎和座椅坐姿偏舒服，比同级别仿赛更适合日常通勤

[故人归 @quanlaidback]

@Qdog_609 我朋友骑过一次这种宽前轮的，说回头看路都变宽了，倒车特别心虚。

[故人归 @quanlaidback]

@rebeka09 原来静音电摩也能有这种压迫感，声浪真是唯一主角这事被推翻了。

[故人归 @quanlaidback]

@cristyl_vedz 短途代步的车不能只看帅，低速好控才是真加分。

[故人归 @quanlaidback]

@Jesse_M_Guzman 我一次也就两三百公里就得歇，超过这数腰先抗议，大巡航的舒服是跑顺了才懂。慢慢骑慢慢体会吧。

[小桃带你懂汽车 @stacie2583]

武汉今天骑Yamaha XMAX，有点职场女主管下班后的感觉，外形干练，坐姿又够松。宜昌、襄阳、荆州想看城市和快速路表现，完整体验在主页。

[RazeDen @razeden0]

I JUST RAN CLAUDE CODE ON A TAYCAN AT THE PORSCHE DEALER

Claude opus 5.5 reverse engineered its carplay in 14 minutes. porsche never approved a thing

the box pretends to be a phone. opus 5.5 does the thinking. the taycan is a 10.9 inch monitor on 4 wheels

laptop → 4 agents in tmux → carplay box → taycan screen → i tap approve

carplay doesn't run apps. it only draws what the phone sends. whatever runs on the box, the car shows - that's the whole loophole

opus 5.5 got 1 photo of the dash and 1 prompt. it mapped the screen at 1920 x 720 and split it into 4 panes, 1 per agent

it even caught that the screen drops to 30 fps when the nav map is open. it kills the map, and the panes go back to 60

the 4 agents ran for 38 minutes: 11 tasks, 9 merged, 2 sent back by me from the driver's seat. the bill was $3.70, or $0.41 per fix

doing those 9 by hand at 40 minutes each is 6 hours. people would say 5x faster. i did the math and it's 9.5x

then i plugged the taycan into the dealer's charger. 10 to 80% took 18 minutes, and 3 more fixes shipped before it was done

opus 5.5 never touches the car. no engine, no brakes, no can bus. it only uses the screen porsche already gives your phone

setup took me 20 minutes: a $139 ottocast box in the usb port, 1 photo of the dash, 1 prompt in claude code, 1 tap to approve

parked only. the dash is the control room, the laptop does the work

the taycan does 0-100 in 2.4 s. my agents did 0 to 12 merged fixes before my coffee at the dealer got cold

everyone test drives a porsche for the speed. i test drove it for the screen - full prompt in the comment below ↓

[故人归 @quanlaidback]

@beamnxw preventive brake stuff is boring til a slight safety happens upside

[故人归 @quanlaidback]

@ch3nweiii fluid levels checked, one short and one dirty, thought it's time

[Grirk @Grirku]

THIS IS A PURE FUCKING GOLDMINE

Anthropic just showed how to keep Opus almost completely off the main path in Claude Code.

Sonnet 5.5 runs the session on high. Haiku 5.5 does the heavy research in parallel. Opus only wakes up when something actually needs a reviewer.

412k tokens shielded from the main context.

You launch it like this:

CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model sonnet

Here’s the split:

→ Sonnet 5.5 high is the main session coordinator
→ Opus sits as code-reviewer on call (before plan locks, on test fails, pre-commit)
→ Haiku 5.5 parallel swarm does AST discovery + spec research
→ Sonnet worker runs edits + tests in an isolated worktree

Plan on high. Delegate on medium. Keep Opus on call.

The full tree
> Sonnet 5.5 high synthesizes the plan and owns the main session
> Haiku 5.5 swarm scans AST + pulls specs in parallel (high-volume work stays shielded)
> Sonnet worker edits and tests inside an isolated worktree
> Opus reviewer only gets called on failures or before commit
> Main context stays clean 412k tokens never enter the lead session

Paste the tree + this prompt into Claude Code ↓

Rebuild my Claude Code setup around this tree:

Run the main session on Sonnet 5.5 with effort high.
Set CLAUDE_CODE_SUBAGENT_MODEL=haiku so research agents stay cheap.
Keep Opus as an on-call code-reviewer only call it before the plan locks, on repeated test failures, and for the final pre-commit audit.
Route AST discovery, spec research, and broad file scanning to the Haiku swarm.
Run all edits and tests inside an isolated worktree.
Never let high-volume subagent output pollute the main context.
Show me every change as a diff first. No commits until I say go.

save it, it’ll come in handy.

[故人归 @quanlaidback]

@lostinazeroy so there his own viewer made garage on the higher mileage for live near the at least ten years

[故人归 @quanlaidback]

@0xOrionVega checking coolant hose once look shiny only see really free coffee at tire chain bit at no cost gets taken mostly live