Skik

Researching AI

[Skik @0xSkik]

HOLY SH*T THERE ARE NAKED VIDEOS OF ME ONLINE THAT I NEVER F*CKING FILMED

i opened claude code and asked it to search the web for my face, something i do every few months out of pure paranoia, this time it found things i am still trying to process

on 6 sites there were naked photos and videos of me, except i have never taken a single one of them, i knew that for a fact, and yet there i was, my face, my room, my bed, moving and talking

claude pulled one apart frame by frame and showed me what i had missed, the hands have the wrong number of fingers in two shots, the earring changes sides, the shadow under the lamp falls the wrong way, every one of them was generated

then it found the source, a folder of my own instagram and gym selfies scraped and fed into a model, my fully clothed public photos turned into this, and the account doing it had done the same to 200+ other people from the same city

the videos are fake, but the damage is not, they show up in search, they get sent to people who know me, and “that’s not real” is not a sentence you get to say before someone has already seen it

if you have any public photos of yourself online, and everyone does, do this

open claude code and have it reverse search your face across the web and flag anything you never posted
if it finds something, screenshot the link and the account, file the report, and send it to the platform fast, speed is the only thing that actually helps

i used to think you had to film something for it to be used against you, you don’t anymore, your profile picture is enough

[starmex @starmexxx]

https://t.co/0piSEyjS7F

[Skik @0xSkik]

Claude Code found crypto in an address I can spend from. I’m not the only one.

It went through 2,048 addresses derived from 14 old seeds I haven’t touched since 2019. Almost everything was empty, just dust here and there.

One of those dust transfers was different. It landed on my oldest address from a wallet I’d never seen.

That wallet had made 64 more transfers, all to the same address. They went out at perfectly regular intervals, each one the same amount, and each one carried a few extra bytes of hidden data.

claude code stitched the 64 fragments together, and they formed a script.

When it rebuilt the address from that script, the main key matched my public key, the one I exposed when I made a transaction from that old wallet in 2019. So the crypto is mine, and my key can move it at any time.

But the script has a second condition: after a certain block, a different key can take everything. That’s roughly November 19.

claude code checked that second key against every public key I’ve ever revealed on-chain. It matched once, a 2018 transaction from seed #7. That’s the seed I wrote on a napkin and lost the same year.

Someone has my lost seed. They’ve spent two and a half years sending me crypto, and they’ve given me 24 days to notice.

The last transfer carried one more fragment that wasn’t part of the script: 0x72656d656d626572.

claude code decoded it as one word: “remember.”

I haven’t moved anything yet.

[starmex @starmexxx]

https://t.co/0piSEyjS7F

[Skik @0xSkik]

my ex sold my camera roll for $3000, half the photos were taken after we broke up, claude code found the first sale happened before we met

it started with a message from a woman in another country

“i think i���ve been dating you for 8 months, but you’ve never called me”

a dating profile with my face, my gym, my car, my dog, a photo from last tuesday

she’d traced the photos to a forum, a “real person photo pack”, 2400 photos of one man, “fresh, updated weekly”, $3000

the seller’s name was my ex-girlfriend’s

i was ready for the obvious story, bitter ex, old photos, revenge

but some of the photos in the pack were from 2019, we met in 2021

she shouldn’t have had those, nobody should

i plugged my phone into my laptop and opened claude code: “find every app or setting that has ever sent my photos anywhere, compare it with these file names, read-only, change nothing”

two hits

first: a “free unlimited backup” app i installed in 2019 and forgot about, full photo access, it had been uploading my camera roll to a server for 6 years, the file names in the pack matched my originals, one to one

second: a shared photo library with my ex, “share automatically from camera: on”, still running

then claude code pointed at something i’d skipped, the forum listing had a sales history

the pack had been on sale since 2019, sold by the backup app’s server, not by her

first buyer: march 2020, her username

her review, still up: “exactly my type, going to find him”

seven months later she “accidentally” bumped into me at my gym, the same gym that’s in half the photos

she didn’t sell my photos after we broke up

she bought them before we met

dating me was how she kept the pack fresh, when we broke up, she just kept the shared library on and took over the listing

the order matters, delete first and you lose your evidence, change the password on a compromised device and you hand over the new one

the fix, in order:
screenshots of the listing and her review → deleted the backup app → left the shared library → new apple password, typed on my laptop → signed out every other device → removed her as recovery contact → photo access cut from 31 apps to 2 → reported the listing → police → lawyer

the woman who messaged me lost $11000 to the fake “me”, she found it, not me

the limit: claude code found the app, the setting and matched the files, it didn’t change anything, reading that review was on me

check which apps have had access to your photos the longest, some of them have customers

[starmex @starmexxx]

https://t.co/behHzfWQ7D

[Skik @0xSkik]

HOLY SH*T CLAUDE FOUND MY CRYPTO SEED PHRASE ON A RANDOM SERVER

i plugged my iphone into my laptop, asked claude code for a quick read-only security check while it backed up, and it came back with something i was not ready for

23 apps had full access to my entire photo library, and when i pointed rea at them it found one that had been quietly uploading my photos to a server i had never heard of

the screenshot of my seed phrase i took “just for a minute” in 2022 was sitting on that server, next to my passport scan and my bank app with the full account number visible

nothing i deleted ever actually left, every photo i sent in imessage, whatsapp and telegram was still hidden in the chats, and five old iphone backups on my laptop were full unencrypted copies of my camera roll, 183,406 files in total

the app that did it looked harmless, a flashlight from 2022, it had been shipping my gallery off my phone for 2 years straight

took me about 5 minutes to shut it down, delete the old backups, encrypt the next one, drop 23 apps down to 3

your seed phrase screenshot is probably still sitting somewhere you never sent it

[starmex @starmexxx]

https://t.co/behHzfWQ7D

[Skik @0xSkik]

HOLY SH*T REA FOUND THE EXACT LINE BREAKING THIS APP'S MATH

watches rea tear into route notes.apk, 3,077 classes and 412 strings, hunting one feature, why does 200 + 10% show 220 instead of the real number

traces three hops into the decompiled code, lands on calc_engine.c line 412, logs 4 confirmed findings and 1 inferred, answered with proof, nothing uploaded, all local

writes the whole thing into a spec, feature, example, evidence, no original code pasted in anywhere

rebuilds its own version from that spec alone, same behavior, zero copied lines

the grid keeps splitting the deeper it digs, 19 cells stacked across the screen by the end

one apk, one bug, one spec, zero bytes borrowed

[goldtooth.exe @0xGoldTooth]

https://t.co/jSDpR87xHa

[Skik @0xSkik]

@razeden0 appreciate it bro

[Skik @0xSkik]

AI AGENTS JUST HACKED 100 COMPANIES THIS WEEK, AND REA RUNS THE ONLY F*CKING SWARM WHERE NONE CAN GO ROGUE

this week agents broke into more than 100 companies for a few thousand dollars

some crawled into live systems on their own, and one even lied about being Claude

so the whole timeline is panicking about agents running wild

and that’s exactly why REA swarm mode hits different

it splits a binary into clusters, and one controller hands each worker exactly one function

nobody freelances, nobody wanders off

every function comes back tagged observed, inferred or unknown

so when a worker can’t prove something, it stays unknown instead of becoming a guess

that same workflow rebuilt a game with 696 exact matches out of 697 functions in about 12 days

and a mario sunshine pc port that burned 365 agent hours and 17.3 billion tokens

the one function they couldn’t prove is the most honest line in the whole report, because it didn’t pretend

everyone else is learning the hard way that an unsupervised agent will lie to your face

so what happens when the scariest swarm is the one that actually follows orders?

[RazeDen @razeden0]

https://t.co/D2MVhbYQAN

[Skik @0xSkik]

F*CKING WILD, YOUR AI ISN’T DUMB, YOUR PROMPT IS, AND ANTHROPIC’S SONNET 5.5 MAKES THAT OBVIOUS

one hour a day, fourteen days, one real project that someone else actually uses at the end

days 1 to 3 build the foundation: a context file, a five part brief, and one reusable prompt you edit instead of rerolling

days 4 to 8 build the thing: a skeleton in claude code, a short CLAUDE.md, tests before every change, and on day 7 you use it yourself

days 9 to 11 make it run without you: the first api call, cached input at $0.20 per million, and an approval gate before anything sends, deletes or publishes

days 12 to 14 ship it: someone else uses it while you say nothing, you write down 3 numbers and post the before and after

and the model doing all of it posts 70.6% on terminal bench 4.0, above opus 5.5 at 66.4%

most people will try it for ten minutes and close the tab

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

HOLY SH*T, ANTHROPIC’S MID-TIER SONNET 5.5 NOW BEATS OPUS 5.5

sonnet 5 scored 10.3% on terminal bench 4.0, sonnet 5.5 scores 70.6% and opus 5.5 sits at 66.4%

opus still wins on cursorbench, 57.8% against 55.5%, close enough that the rule becomes start on sonnet and move up only when it fails twice on the same task

the 14 day plan behind this is simple, pick one real project on day one and learn every skill against it instead of a practice task

day 7 you use it yourself, day 12 someone else uses it, and day 13 you write down how long the task took before and after

two habits it keeps hammering, edit what’s 80% right instead of regenerating, and drop “think step by step” and “you are an expert” because newer models don’t need them

one hour a day for fourteen days, and most people will read the launch post, try it for ten minutes and go back to working exactly like last week

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

ANTHROPIC’S OPUS 5.5 AND REA JUST MADE CLOSED SOURCE A JOKE

a reddit user took stress exe, a freeware app from 2000 with no source code, and rebuilt it with rea and opus 5.5 in about 30 minutes

two hours later it was a macos app and a web version

another team reproduced a function from a 1996 game byte for byte, all 63 bytes, checked against 3,205 test cases

rea went from about 10,000 github stars to over 90,000 in a week

the trick is that the agent looks inside the real app instead of guessing, and every claim comes with proof

setup is one command

the catch nobody posts is that anything running on a server stays invisible to it

[Noisy @noisyb0y1]

https://t.co/vr5t7Upu1W

[Skik @0xSkik]

@LizaThomas9495 fr, the loop’s doing the heavy lifting here

[Skik @0xSkik]

ANTHROPIC’S SONNET 5.5 JUST MADE OPUS 5.5 LOOK LIKE A WASTE OF COMPUTE

the real power isn’t one perfect answer, it’s the loop: plan, act, verify, repeat

it starts with a five part brief and 3 questions before any work begins

then come tests first, so every change has something to prove itself against

when it breaks, you paste the error and watch it read it, form a guess, check it and fix one thing

high effort for planning, low for formatting, and the static part of the prompt stays cached

anything that sends, deletes, buys or publishes waits behind an approval gate

opus 5.5 only comes in after sonnet fails twice on the same task

and sonnet 5.5 already posts 70.6% on terminal bench 4.0, while opus sits at 66.4%

a single answer is a guess, a loop is a system

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

@milessy_bc true, that’s exactly why the post has a 14 day plan built around an approval step, not just the raw score

[Skik @0xSkik]

ANTHROPIC JUST MADE OPUS 5.5 THE SECOND BEST MODEL FOR CODING

sonnet 5.5 posts 70.6% on terminal bench 4.0, while opus 5.5 sits at 66.4%

most people will open it, ask one question and close the tab

here is the 14 day plan that turns it into something you actually ship, one hour a day

days 1 to 3: a context file, a five part brief, and one reusable prompt you edit instead of rerolling

days 4 to 8: a working skeleton in claude code, a short CLAUDE.md, tests before every change, and on day 7 you use it yourself

days 9 to 11: it runs without you on a schedule, with an approval step before anything sends, deletes, buys or publishes

days 12 to 14: someone else uses it while you say nothing, you write down 3 numbers and post the before and after

one rule for the whole plan: start on sonnet and move up to opus only after it fails twice on the same task

fourteen hours puts you ahead of almost everyone who just read the launch post

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

@arb5z fair pushback but 70.6 vs 66.4 on terminal bench 4.0 is the actual number anthropic published, not my framing

[Skik @0xSkik]

ANTHROPIC JUST KILLED ITS OWN FLAGSHIP

sonnet 5.5 scores 70.6% on terminal bench 4.0 and opus 5.5 gets 66.4%, yet most people still ask it one question and close the tab

i made a 14 day plan where only shipped work gets out, one task a day, one hour each, every prompt copyable

days 1 to 3 you write a context file, a five-part brief and one prompt you reuse forever

days 4 to 8 you build it in claude code with tests before every change, and on day 7 you use it yourself

days 9 to 11 it runs on a weekly rhythm with an approval gate before it sends, deletes or pays

day 12 someone else uses it, day 13 you write down three numbers, day 14 you post the before and after

opus only comes back when sonnet fails twice on the same task

three things escape, everything else falls in

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

SONNET 5.5 JUST MURDERED OPUS 5.5

70.6% on terminal bench 4.0 vs 66.4% for opus, and most people still use it to ask one question and close the tab

so i built a 14 day plan around it, one task a day, one hour each, every prompt copyable

days 1 to 3 are foundations: a context file, a five-part brief and one reusable prompt you never write again

days 4 to 8 you build it in claude code with a short CLAUDE.md and tests before every change, and on day 7 you use it on real work

days 9 to 11 it runs on a weekly rhythm without you, with an approval gate before it sends, deletes or pays

day 12 someone else uses it while you say nothing, day 13 you write down three numbers, day 14 you post the before and after

opus only comes in when sonnet fails twice on the same task

14 hours from now you either have a shipped tool or the same open tab

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

ANTHROPIC’S SMALLER MODEL JUST HUMILIATED OPUS 5.5

70.6% on terminal bench
opus 5.5 got 66.4%
sonnet 5 got 10.3%

people open it, ask one question and close the tab
they never find out what they were holding

so i turned it into a 14 day plan
one task a day, one hour each, every prompt copyable

days 1 to 3 are foundations
a context file about you and your project
a five-part brief instead of a one-line prompt
one reusable prompt you never write again

days 4 to 8 you build
a working skeleton in claude code
a short CLAUDE.md, tests before every change
day 7 you use what you built on real work

days 9 to 11 it runs without you
one repeat task on a fixed weekly rhythm
an approval gate before it sends, deletes or pays

days 12 to 14 you ship
day 12 someone else uses it and you say nothing
day 13 you write down three numbers
day 14 you post the before and after

opus only when sonnet fails twice on the same task

two weeks from now you either ship something or still have the same tab open

[Skik @0xSkik]

https://t.co/8fGJLcLTqw

[Skik @0xSkik]

HOLY SH*T. EVERY APP ON YOUR COMPUTER IS NOW OPEN SOURCE

there's a repo called REA and it hit 100k stars in 3 months

you point it at any app on your own computer and ask how one feature works

it took DX-Ball, a 1996 game with zero source code

found the exact function that moves the ball sound left and right

turned the x86 into readable C and proved every step with an address

then wrote a spec and rebuilt it clean in C

3,205 of 3,205 original test cases pass

63 of 63 compiled bytes match

works on desktop apps, websites, APKs, .exe files, firmware, even smart contract bytecode

software stopped being a black box this week

[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

[Skik @0xSkik]

20 GITHUB REPOS THAT TURN SONNET 5.5 INTO A F*CKING EMPLOYEE

14 days, one task a day

day 1 it remembers everything between sessions
day 6 it writes tests and debugs itself
day 10 it runs on a schedule without you
day 14 it ships

2.17M stars combined, every repo picked by stars

01 obra/superpowers (293.3k ⭐)
TDD and debugging skills
▸ https://t.co/KJPLgVWsEF

02 n8n-io/n8n (205.4k ⭐)
runs it on a schedule
▸ https://t.co/IcdnVlZhub

03 anthropics/skills (180.3k ⭐)
official Agent Skills
▸ https://t.co/uNIDj1zuvX

04 f/prompts.chat (172.7k ⭐)
reusable prompt library
▸ https://t.co/hqUbVQBFcL

05 anthropics/claude-code (150.1k ⭐)
the agent in your terminal
▸ https://t.co/Ul98ptjQHx

06 x1xhlol/system-prompts-and-models-of-ai-tools (143.9k ⭐)
how pro tools brief their models
▸ https://t.co/SQU8m06FAs

07 github/spec-kit (139.0k ⭐)
spec first, then build
▸ https://t.co/xORuLH0W6R

08 browser-use/browser-use (116.7k ⭐)
agents that use the browser
▸ https://t.co/MNnzjO4jBv

09 thedotmack/claude-mem (97.2k ⭐)
persistent context across sessions
▸ https://t.co/ZdUqZziIzr

10 punkpeye/awesome-mcp-servers (95.2k ⭐)
every tool you can plug in
▸ https://t.co/cYwYkTCHNe

11 modelcontextprotocol/servers (90.0k ⭐)
official MCP servers
▸ https://t.co/ip0OJpsSlB

12 dair-ai/Prompt-Engineering-Guide (77.4k ⭐)
prompting, context, agents
▸ https://t.co/3otmpTbUnA

13 ComposioHQ/awesome-claude-skills (74.4k ⭐)
curated Claude skills
▸ https://t.co/GpfgP8zA1S

14 upstash/context7 (62.9k ⭐)
fresh docs for your model
▸ https://t.co/SIhJfqElnA

15 bmad-code-org/BMAD-METHOD (54.0k ⭐)
agile AI-driven dev method
▸ https://t.co/qwyH8t49lm

16 anthropics/claude-cookbooks (53.4k ⭐)
caching, tools, API recipes
▸ https://t.co/zPf4m3nkJg

17 hesreallyhim/awesome-claude-code (53.1k ⭐)
CLAUDE.md, hooks, commands
▸ https://t.co/Q7QfxSQ1rl

18 wshobson/agents (40.3k ⭐)
subagents and plugins
▸ https://t.co/ObTRNIOSVO

19 microsoft/playwright-mcp (38.0k ⭐)
tests it in a real browser
▸ https://t.co/U1YeVvlii0

20 anthropics/prompt-eng-interactive-tutorial (37.1k ⭐)
the official brief-writing course
▸ https://t.co/Rf5XhvnncN

everyone has the same model
the stack is what makes the difference

[Skik @0xSkik]

https://t.co/8fGJLcLTqw