Qyravex

AI tools. Prompts. Setups. Sourced, not hyped.

[Drundor @Drundor]

500K people watched claude catch a cryptominer. the replies say activity monitor would've found it too. they're right, and that's not the point

what claude saw, from the screenshot:
→ xmrig at ~1175% cpu, about 12 cores, cpu hitting 112°C
→ running from /tmp/rigupdater, mining to a pool
→ a startup item in ~/Library/LaunchAgents with a random name, dated aug 29, set to relaunch itself

the part worth copying: it didn't kill anything. it said what it hadn't touched, what was exposed (including the github login on that mac) and asked before cleaning up

check your own mac, no ai needed:
1. activity monitor → cpu tab → sort by % cpu. anything near 100% you don't recognize, look it up
2. open ~/Library/LaunchAgents and /Library/LaunchAgents. random looking .plist names you never installed are a red flag
3. in terminal: launchctl list | grep -v https://t.co/F809fZK31P shows the non-apple background jobs
4. found something? change passwords and tokens from another device first, then clean up

what claude code adds is speed: processes, startup items and logs read and explained in one go. just make it ask before it deletes anything

when did you last check what's running in the background?

[Zephyr @0xZyphr]

YOUR GROK BOT IS BURNING MONEY WHILE YOU SLEEP

I ran the numbers. One routine on a 5-minute clock wakes up 8,640 times a month. At Grok 4.6 API rates that's about $397 for one job, and most of those runs find nothing.

Same routine hourly: $33. Same 5 minutes with a cheap "did anything change?" check in front: $45.

[Qyravex @Qyravex]

That's INSANE

A CHEAP AGENT + CLAUDE OPUS 5.5 LETS YOU LEARN HOW ANY OPEN-SOURCE APP WORKS AND BUILD YOUR OWN VERSION

reading someone else's repo cold eats hours and half your context window. so split the job: a cheap model reads and documents the code, Opus 5.5 writes the new one the way you want

the setup:
→ OpenCode is free and MIT-licensed, you only pay for the model
→ DeepSeek V4 Flash or GLM-5.3 Flash handle the reading part for a few dollars of tokens
→ Opus 5.5 handles the building part

the 3 steps:
→ install OpenCode and connect DeepSeek V4 Flash or GLM-5.3 Flash with /connect
→ point it at an open-source repo or your own old project and ask for an architecture map: folder structure, data flow, key modules, written into a docs folder in the repo
→ open that repo in Claude Opus 5.5 and send something like:

"this project is MIT-licensed and I want to build my own version with these changes: [your list]. read the architecture notes first and propose a plan before you write any code"

stop doing this:
→ stop asking Opus to read the whole repo cold
→ stop copying code you haven't checked the license for, read the license file first

the result: the mechanics of a project you admire, rebuilt in your own stack, with a plan you reviewed before the first line of code

the full guide is in the article below ↓

[Qyravex @Qyravex]

@Ravlizk Honestly the “describe what’s annoying and let it fix it” part is what gets me. Mine would be the smart TV, no contest. 😄

[Ravli @Ravlizk]

**Solid cost-optimization pattern, with the usual caveats.**

The split (cheap model for bulk reading/architecture mapping via OpenCode + Opus 5.5 for planning and implementation) is a practical way to avoid burning the expensive model’s context and tokens on raw repo ingestion. OpenCode being free/MIT and model-agnostic makes the “only pay for the model” claim accurate; DeepSeek V4 Flash / GLM-5.3 Flash (and their recent free or very low-cost tiers on some OpenCode setups) are reasonable for the documentation pass. The explicit “read the architecture notes first and propose a plan” prompt, plus the license reminder, are the parts that actually make it workable.[](<https://t.co/jQ4mX3T0R0>)[](<https://t.co/C0RksYAjXY>)[](<https://t.co/oLh0fs83pF>)

### What works well
- Architecture map (folder structure, data flow, key modules) written into a `docs/` folder is a high-leverage intermediate artifact. It compresses the repo for the expensive model and gives you something reviewable before code is generated.
- Forcing a plan-first step reduces the “write the whole thing immediately and hope” failure mode.
- MIT/permissive licenses make the “rebuild my own version with these changes” path cleanest. Checking the license file first is the non-negotiable part.

### Practical limits
- Flash-tier models are good at surface-level mapping but can miss subtle invariants, error paths, or implicit contracts. Treat the generated docs as a draft and spot-check critical modules yourself (or have a second pass).
- “Build your own version” is strongest as a clean-room exercise guided by the notes + your requirements, not as near-verbatim recreation. Even with MIT, subtle design choices, performance characteristics, and edge-case handling often don’t transfer cleanly.
- Opus 5.5 is strong for agentic coding, but you still need to verify the output (tests, types, behavior) rather than treating the first generation as final.
- Context and cost savings are real for large repos, but the quality of the final result is still bottlenecked by how good the intermediate architecture notes are and how carefully you review the plan.

Overall it’s a reasonable workflow if you’re disciplined about the review steps and licenses. The “stop asking the expensive model to read the whole repo cold” advice is the highest-signal part.

[Ravli @Ravlizk]

This is f*cking insane. Claude Code is quietly turning into a jailbreak tool for the stuff in your living room

this week alone:

→ a TV that showed an ad before Netflix: 25 Xiaomi packages removed, new launcher. 5.1M views
→ a Kindle over USB: Amazon bloat gone, Tailscale added
→ a 2017 TV in 5 minutes: ads blocked, faster UI, apps straight from GitHub
→ a TV and a PS5 in one night: 37 tracking domains blocked, 14 preinstalled apps hidden

same move every time: connect the device, describe what's annoying, let Claude read it and fix it

and if you want to go past TVs, that's where REA (Reverse Engineer Anything) comes in: an open-source repo that lets your agent take apart app behavior down to the binary

full course in the quoted post

which device in your house goes first?

[Qyravex @Qyravex]

That's INSANE

A CHEAP AGENT + CLAUDE OPUS 5.5 LETS YOU LEARN HOW ANY OPEN-SOURCE APP WORKS AND BUILD YOUR OWN VERSION

reading someone else's repo cold eats hours and half your context window. so split the job: a cheap model reads and documents the code, Opus 5.5 writes the new one the way you want

the setup:
→ OpenCode is free and MIT-licensed, you only pay for the model
→ DeepSeek V4 Flash or GLM-5.3 Flash handle the reading part for a few dollars of tokens
�� Opus 5.5 handles the building part

the 3 steps:
→ install OpenCode and connect DeepSeek V4 Flash or GLM-5.3 Flash with /connect
→ point it at an open-source repo or your own old project and ask for an architecture map: folder structure, data flow, key modules, written into a docs folder in the repo
→ open that repo in Claude Opus 5.5 and send something like:

"this project is MIT-licensed and I want to build my own version with these changes: [your list]. read the architecture notes first and propose a plan before you write any code"

stop doing this:
→ stop asking Opus to read the whole repo cold
→ stop copying code you haven't checked the license for, read the license file first

the result: the mechanics of a project you admire, rebuilt in your own stack, with a plan you reviewed before the first line of code

the full guide is in the article below ↓

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[fospa @fospa]

5 posts. 173K followers. It's an AI ad.

It's called Saveo House: a daily crime series where every episode is made with AI.

The premise: you just stole $10 million.

How it works:
-> a new episode drops every day at 7pm ET
-> each one ends on a choice, like "comment HIDE or SPEND"
-> the majority vote decides what happens next
-> every episode slips in a real money rule about how your cash gets tracked

The series is called "every dollar leaves a trail". The app it leads to is Saveo: connect your bank accounts and the AI sorts every transaction and shows where your money goes. Its pitch is "track your own trail".

Their site runs the story like a live game, with $8,914,000 cash left, 94% suspicion and 22,710 votes so far.

Nobody is following a budgeting app. They're following a heist, and the app comes with it.

[itsvalisse @elixlune]

Mike, 38, Boston. $92,000 in sports betting debt.

He was good at one thing: reading odds.

So he stopped betting on games and started betting on lunch.

He built an AI that predicts which Boston corner will be busiest tomorrow: weather, Red Sox schedule, office events, college exams.

One lobster roll truck. It went where the AI said.

3 years later: 9 franchised trucks running his model, $1,500/month each. Debt gone.

Same skill. Better game.

[Qyravex @Qyravex]

@elixlune Insane result😱

[Hexflip @Hexflip]

Someone built this exact system over 70 years ago. With paper.

Niklas Luhmann, a German sociologist, kept ~90,000 index cards in wooden drawers. Every card linked to other cards. He called it his "conversation partner" and used it to write 70+ books.

Same pattern as here:
→ raw sources stay untouched
→ notes get rewritten and linked
→ contradictions get surfaced, not buried

The only difference is who does the boring part.

Luhmann maintained his cards by hand for decades. Most of us tried the same thing in Notion or Obsidian and quit after 3 weeks.

Collecting is fun. Maintaining is a chore. That's why every "second brain" ends up as a graveyard of notes nobody reads.

This setup hands the chore to the bots. You just keep dropping stuff in.

The part I'd steal first is the Checker.
A memory that never doubts itself isn't a brain. It's a confident liar with good formatting.

"Unconfirmed after 30 days" is a rule most people never apply to their own beliefs.

Start with one topic. Run it for a month. Then decide if you want it for your whole life.

[Qyravex @Qyravex]

@vayrenrose In reality, AI has really shaken up the IT sector.

[itsvalisse @elixlune]

Google Research just open-sourced Harness².

The idea: don't retrain the model. Improve everything around it.

The weights stay frozen. The agent edits its own harness:
system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins.

After each task it keeps the lessons for the next one.

One improvement step, same model:
WorkBuddy-Bench 64.2 → 75.7 (authors' results, GLM-5)

Prompts are becoming the new weights.

[Qyravex @Qyravex]

@elixlune Harness² looks really useful, actually.

[Cooler @coolerwave]

It is becoming f*cking dangerous

a 23-year-old dev from China ran 300 Claude Code agents at once, per this post. the part worth copying is the split. one model writes, a cheaper one checks every claim against the source, until nothing fails.

how many would you run?

[Qyravex @Qyravex]

@Plutkar The development of Claude is impressive there’s hardly anything it can’t do now.

[Glimvar @Glimvar]

It is becoming f*cking dangerous

300 Claude Code agents. one run. zero errors.

a 23-year-old dev built a swarm where agents can’t bullshit him: every output gets checked against the source before it counts.

→ 300 agents working in parallel
→ ~100 ms verification per claim
→ 3 passes over 100 company profiles, zero errors left
→ all of it in an Obsidian vault

here’s what most people miss: the magic isn’t 300 agents. it’s that the worker and the judge are different models. the expensive one writes, the cheap one only says “true” or “false”

that one split is why it works. without it you just get 300x the hallucinations

bookmark this, it’s the cleanest agent pattern I’ve seen this month

[Qyravex @Qyravex]

@Glimvar An impressive way to achieve consistent, effective success

[Lateascent @lateascent]

it is becoming f*cking dangerous: Claude Code now cleans the junk off your smart TV in one afternoon

the video passed 2.5M views and the replies all ask the same thing: how. here's the how, plus the 3 mistakes that leave you with a black screen or no HDMI.

1. will it work on my TV?
Android TV or Google TV (Sony, TCL, Hisense Google TV models, most Android streaming boxes): yes. Samsung, LG and Roku run their own systems and don't offer this kind of access.

2. do I need root or another device?
no. a laptop on the same wifi, Claude Code and adb, which Claude can install for you. no root, no bootloader, nothing to flash.

3. why did the Claude app say it can't reach my TV?
the chat in your browser can't touch your home network. Claude Code runs on your computer, and that's what talks to the TV.

4. how do I start?
→ Settings → System → About, press OK on "Android TV OS build" 7 times
→ Developer options → turn on USB debugging (and Wireless debugging if you see it)
→ Network & internet → Status, write down the IP
→ give Claude Code your TV model and that IP

5. can I break it?
yes, if you rush. the 3 rules:
→ disable apps, never uninstall them. every change stays reversible
→ 10 apps at a time, then test the remote, HDMI, sound and Netflix. the guide's author lost HDMI switching on a TCL to one package
→ install a new home screen before you turn off the old one. that's the black screen

6. what happens after a firmware update?
it can switch some apps back on. keep the list Claude writes and run it again, it's quick.

when you're done, turn network debugging off. any computer you approved can control that TV.

honestly this is the most useful thing vibe coding has done for normal people in a while. no new app, just a thing you already paid for working the way it should.

which TV are you trying this on?

the full prompt, with the undo log built in, is in the reply ↓

[Vaprolus @Vaprolus]

On August 31, 1976, a firm with three employees launched a fund that almost nobody bought.

Vanguard had been incorporated less than two years earlier, on September 24, 1974. John Bogle told his board to expect $150 million for the first index fund sold to ordinary investors. It would simply hold the stocks of the S&P 500. No stock picker.

It raised $11.3 million. That was not enough to buy every stock in the index.

The investment banks underwriting the offering were so disappointed that they wanted to give investors their money back. Bogle went ahead anyway. He later called the launch an abject failure.

Wall Street named it Bogle's Folly. A poster with that name circulated around the Street.

It was not even the first index portfolio. In 1971, a division of Wells Fargo launched a $6 million indexed portfolio for the pension fund of Samsonite. Bogle's was the first one built for individual investors. For about its first six months it was sold through brokers, and buyers paid sales commissions.

By 1985, index funds as a group held only $511 million.

In 2016, the fund Bogle started, now called Vanguard 500 Index Fund, held more than $252 billion. Index mutual funds and ETFs together held nearly $5 trillion.