Mito

21 | AI builder & researcher | testing systems | coding

[Mito @0xMitto]

@theSethian Because if Microsoft stops, someone else keeps building, thats the answer every lab ends up giving

[Mito @0xMitto]

@tidyowl Deleting a session takes its worktree with it, thats one warning I’d put in bold

[Mito @0xMitto]

@0xgloob Runbook links that open the right repo with the prompt ready is the first deep link use that actually makes sense to me

[Mito @0xMitto]

@ottollm 56 KB for domes, a rover, a rocket and 36 potato plants is wild, my screenshots are bigger than that lol

[Mito @0xMitto]

@Lumitri_x The kids Netflix tablet is exactly the device nobody thinks about and the one that never gets an update lol

[Mito @0xMitto]

hackers stole $5.3 million per second from binance. for 20 minutes straight

i gave claude code one sentence. it found where every dollar went, and $3.6 billion of it is still moving right now

"hey claude, follow the funds drained from binance's hot wallets"

no api keys. no insider access. only the public blockchain anyone can open

29,450 transactions. 10 chains. 376 attacker wallets, all wired to one operator

this is 4x bigger than the $1.5 billion bybit hack. the biggest crypto theft in history, and it took less time than your lunch break

how they did it:

-> days before: operator wallet funded through tornado cash
-> then 50 authorizations signed in advance, sitting there like a loaded gun
-> 20:05 utc: one sweep hits binance hot wallets 14 and 16
-> 20:25: done. 10 chains emptied

where the money went:

$3.79 billion in eth
18,270 btc
$628 million on tron
$296 million in xrp

then it got shredded through uniswapx, aave, justlend and btc peel chains, smaller and smaller until it's almost invisible

claude matched the pattern to the bybit hack. same playbook, same crew

trm labs names them: tradertraitor, part of the lazarus group, high confidence. the fbi hasn't said a word yet

tether, circle and ripple froze $2.7 billion in time. binance says user funds are safe

the other $3.6 billion is still out there, spread across hundreds of wallets

if a group of hackers can empty the biggest exchange on earth in 20 minutes, ask yourself how safe your exchange account really is

[Mito @0xMitto]

@PandaAgents Already sending this to my dad, he has tapped later on every update since 2019 lol

[Mito @0xMitto]

@masamunecode One eye going red the second time it looks at you means it remembers me, I’m not sure I like that

[Mito @0xMitto]

@Lumitri_x Motion mode in the hallway is gonna start beef with everyone you live with

[Mito @0xMitto]

I'm never leaving my phone face down on the table again... Opus 5.5 just showed me what it does when the screen is off

I asked it to build one thing: a tool that reads Android's hidden system log and explains every line in plain English. Took one evening

I plugged the phone in, locked the screen, put it on the table and didn't touch it

8 seconds later the tool had a list. My selfies. Scans of my documents. Screenshots of my private chats. A spreadsheet I downloaded months ago and forgot existed. Over 20 files, opened one after another, with the screen black

I thought it was old history. Then I scrolled to the bottom and saw a new line appear while I was reading: another selfie, opened 1 second ago

Android shows you a green dot when an app uses your camera or mic. Reading your files gets nothing. No dot, no popup, no notification. The only place it shows up is a log nobody ever reads

8 seconds was all I recorded. The phone is on 24 hours a day, every day, for years

I don't know how many times my selfies and documents have been opened like this. I just know I never saw it once

If you want to see yours, plug it in, turn on USB debugging and ask Claude Code to read the system log and list every app that opened your photos, documents and screenshots while the screen was off

[Mito @0xMitto]

It's starting to get f*cking dangerous

A 23-year-old Chinese dev spun up 300 Claude Code agents side by side and wired them into a fact-checking layer, so none of them could feed him nonsense

Open his live dashboard and you see:
- 300 agents working and verifying in parallel
- Each output matched to its source in roughly 100 ms
- 100 company profiles, 3 rounds of checks → not a single error left
- Everything piped straight into Obsidian

The secret is the division of labor: an expensive model does the work, a cheap one rules on whether it's true. Nothing stops until every claim holds up

Analysts would bill weeks for a job like this. He finished it in one go

This is the Second Brain + agent swarm combo that keeps going around. Save the full breakdown

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Mito @0xMitto]

This is getting f*cking dangerous

A 23-year-old developer in China just launched 300 Claude Code agents at once and put a verification layer on top so not one of them could lie to him

Here's what his live dashboard shows:
- 300 agents writing and fact-checking simultaneously
- Every answer checked against the original source in about 100 ms
- 3 passes over 100 company profiles → 0 errors remaining
- All of it saved directly into an Obsidian vault

The whole trick is splitting the roles: one model does the work, a cheaper model checks whether each claim is actually true. The loop keeps running until everything passes

An analyst team would charge weeks of work for this. He did it in a single run

It's the same Second Brain + agent swarm setup everyone keeps sharing. Bookmark the full breakdown

[Mito @0xMitto]

This is getting f*cking dangerous

A 23-year-old developer in China just launched 300 Claude Code agents at once and put a verification layer on top so not one of them could lie to him

Here's what his live dashboard shows:
- 300 agents writing and fact-checking simultaneously
- Every answer checked against the original source in about 100 ms
- 3 passes over 100 company profiles → 0 errors remaining
- All of it saved directly into an Obsidian vault

The whole trick is splitting the roles: one model does the work, a cheaper model checks whether each claim is actually true. The loop keeps running until everything passes

An analyst team would charge weeks of work for this. He did it in a single run

It's the same Second Brain + agent swarm setup everyone keeps sharing. Bookmark the full breakdown

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Mito @0xMitto]

OPUS 5.5 + GROK BOT cleared a 48-task backlog in 7 hours for $5.76, and it's f**king unreal

31 pull requests, 0 broken builds, and not a single line merged without my ok

Google says AI already writes over a quarter of its new code. this is the same thing for one person, for 12 cents a task

Opus 5.5 is the brain: reads all 48 tasks and plans every single one before anything gets touched

Grok Bot is the hands: takes them one by one. fix → test → pull request → next

how it actually holds up overnight:

-> test fails? Grok Bot retries until it's green. 39 test runs, 0 broken builds
-> task needs no code? it leaves a note instead of faking a fix. 14 of those
-> anything risky gets parked for you. 3 of those
-> nothing merges on its own. ever

48 tasks at an hour each is 6 full workdays for a developer. this was one night and $5.76

that's $0.12 per task. a junior dev costs about $8,000 a month

save this before your next sprint ↓

[Avid @Av1dlive]

https://t.co/sVlvM5N7Xb

[Mito @0xMitto]

REA + Opus 5.5 can rebuild any feature from any app in one weekend, without copying a single line of their code, and it's f**king unreal

i pointed it at the Cmd+K search in my notes app. 312 lines later i had my own version, 14 out of 14 tests green, and 0 lines of their code

this is the same "clean room" trick Compaq used in 1982 to clone IBM's BIOS legally, and it helped turn them into a billion-dollar company. one team studies, another team builds, only a spec crosses the wall

REA traced the feature through 9 functions in 4 files, then Opus 5.5 wrote down every behavior and 11 edge cases. no code, just what it does

then a fresh agent built mine from that spec only, and kept fixing it until every test passed

a freelancer would charge you thousands and 2 weeks for one feature like this. this took a weekend

one 35-line prompt does all of it inside Claude Code ↓

step 0 → saves progress to a file, so you can stop and pick up later
step 1 → asks you 3 questions: which feature, what it should run on, how deep
step 2 → sets up REA for you and waits for your ok
step 3 → traces the feature from your click to the result
step 4 → writes a spec: behavior only, every line with proof, zero code
step 5 → builds your version from the spec only, tests first
step 6 → runs the tests until all green, then shows spec → code → test side by side

your own apps only. no bypassing licenses or logins, never copies their code, assets or text

full prompt is free in the image, copy it before you pay for another subscription ↓

[Mito @0xMitto]

REA + Opus 5.5 can find exactly why any app is slow, down to the line of code, and it's f**king unreal

i ran it on an app i open 20 times a day. every click took 3.2 seconds, and 2.5 of them came from one function someone forgot to fix

Amazon famously found that every 100ms of delay cost them 1% of sales. this app was wasting 2,500ms on one click

the culprit: it was drawing the same list of 1,400 rows 3 times every time i opened a project. 79% of the wait, in one file, on one line

3.2 seconds × 20 clicks a day is 6.5 hours a year staring at a loading screen. a performance consultant charges thousands to find this. this took one evening

one 36-line prompt does all of it inside Claude Code ↓

step 0 → saves progress to a file, so you can stop and pick up later
step 1 → asks you 3 questions: which app, which slow action, what you want at the end
step 2 → sets up REA for you and waits for your ok
step 3 → records the slow action 3 times, no guessing
step 4 → maps every step from your click to the screen, in milliseconds
step 5 → finds the biggest bar and explains why it's slow
step 6 → writes a report with proof and the smallest fix

your own apps only. measure first, no guesses, never touches the original install

full prompt is free in the image, copy it before you wait another 3 seconds ↓

[Mito @0xMitto]

built a trading terminal with DOTS + JEV in 2 hours, turned it on and went to bed

woke up to $12,161

2 years of charts till 4am never got me close

overnight it checked 1,608 coins and bought only 4

3 went up, one did +204%

the 4th went red and it sold at -26% instantly. i would've held that to zero

2 hours of work beat 2 years of my trading

1,608 coins, 4 trades, 1 night, $12,161

[Mito @0xMitto]

built a trading terminal with DOTS + JEV in 2 hours, turned it on and went to bed

woke up to $12,161

2 years of charts till 4am never got me close

overnight it checked 1,608 coins and bought only 4

3 went up, one did +204%

the 4th went red and it sold at -26% instantly. i would've held that to zero

2 hours of work beat 2 years of my trading

1,608 coins, 4 trades, 1 night, $12,161

[savip. @savipww]

https://t.co/dBrZixM1dC

[Mito @0xMitto]

SPIDER + JEV + DOTS made me $12,161 in one night and i didn't press a single button...

220 tokens scanned, 4 trades, 3 wins, 1 loss

spider reads → jev picks → dots trades

Spider reads the whole market: whales, volume, socials, wallets

Jev is the gate. every token goes through 7 hard rules: 15+ minutes old, $12k+ liquidity, 150+ trades, no wallet holding over 5%...

then 5 soft checks: is the dev still loaded? is the hype louder than the chain? is the move already spent?

one fail and it's dropped

Dots is the hands. it places every order in 0.3s and never more than 30% of the bank on one ticket

00:44 → first entry, out at +88%
02:17 → $SILK, filled in 0.3s, closed green
04:14 → $FLY, "slipping, stop is close". cut at -26%
06:11 → $BITE, spider read it 2 hours early. out at +204%

the loss is the important part. no holding a bag hoping it comes back. cut and move on

220 scanned, 4 trades, 1 cut, $12,161 by 07:00 ↓

[savip. @savipww]

https://t.co/dBrZixM1dC