5.5M people saw claude code clean an android TV. a lot of them own a samsung or LG, and for those this exact trick doesn't work. here's what does
samsung, lg, hisense vidaa, vizio and roku TVs don't run android, so there's no adb and claude code can't clean them this way
where the repo is 2 days later:
→ 54 TVs with tested lists, 23 brands
→ 4.7K stars, 580+ forks
→ the showcase has an LG port: the same home screen rebuilt as a webOS app through LG's developer mode
→ fire tv forks that open the new launcher over amazon's home, since fire os won't let you replace it
and yes, android debloaters existed before. the new parts are the per model lists and the launcher
not android? what you can still do today, no hacking:
1. turn off ACR, the thing that reports what's on your screen
samsung: viewing information services
lg: live plus
roku TVs: privacy, smart tv experience
vizio: viewing data
hisense vidaa: check system or privacy, the name changes by model
2. set the TV's DNS to an ad-blocking resolver like adguard. it cuts ads and trackers that load from their own domains. not youtube ads, and the resolver sees every lookup your TV makes
3. check again after updates. people say these toggles come back on
android or google TV? disable, don't uninstall, and switch usb debugging off when you're done
what TV are you stuck with?
Vibe coded my parents' TV this afternoon!
It was showing a sponsored ad before you could even get to Netflix.
Gave Claude Code the model number, it found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher.
Looks like an Apple TV now lol
@Avernyx the interesting part isn't even the TV cleanup, it's that people no longer need to understand adb to do this stuff. but i'd still be nervous letting an AI disable system packages without checking first. curious how well this works across different TV models
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?
agree on the verification part. funny thing is this one barely needed agents, it was mostly claude doing the boring terminal steps. would you let it touch your own devices?
415K people watched claude code jailbreak a kindle. the replies quietly explained what actually happened
the best comments under it:
→ "jailbreaking kindles has been a thing for a while. it's just easier now with AI"
→ "you don't need claude to do this, there is a 1 click setup"
→ "claude didn't refuse?"
so what did claude code actually add?
1. the jailbreak itself: community tools already exist, people in the replies name winterbreak. claude mostly runs the steps and reads the errors for you
2. the part that's actually new is what he built after. his own reply: save an article with readeck, it downloads and syncs to his home server, then lands on the kindle
3. tailscale lets the kindle reach that home server from anywhere, not just home wifi
that's the pattern behind the whole tv, kindle, ps5 wave this week. claude code isn't finding exploits, it's doing the tedious terminal work people used to give up on
before you try it on your own stuff:
→ follow the community guide for your exact model, don't let claude improvise the jailbreak step
→ back up first. it's your device, not a demo
→ ask claude to log every command so you can undo it
what device would you actually let claude code touch?
@vayrenrose "reproduce the one feature customers actually pay for" is the real point here. feels like the moat just moves to data, distribution and trust. which tools do you think get hit first, dev tools or the boring b2b stuff?
@Xylomere the slow facts vs fast facts split is the part people skip. claude code works the same way, CLAUDE.md loads every session so only the slow stuff belongs there. anything that changes daily just bloats it
tried the code-only version of this. no after effects, no prompt-to-video tool. one python script, claude wrote all of it including the beat. about 5 minutes start to finish
what it nails: abstract motion, kinetic type, 3d particles, anything you can describe with math
what it can't do: people, real footage, your actual brand assets
so not a $400/hr designer. more like a very fast intern who's great at particles and has never seen a human face
the prompt:
"make an 18s 1080x1350 motion video fully in python (pillow, numpy, ffmpeg). 120 bpm. generate the soundtrack in code (kick, hats, clap, bass) and cut every scene on a beat. scenes: kinetic text slams, text dissolving into 3d particles, particles morphing into a rotating sphere and then a wave grid, then the script itself next to the render. after the first render, look at a frame from every scene, list what looks off, fix it and render again"
you don't need after effects for this kind of video anymore
the whole thing is code. the 3d particles, the wave, the beat, the sync. claude wrote every line and it took about 5 minutes start to finish
how it actually works:
→ one python script draws every frame. pillow for drawing, numpy for the 3d math
→ 1600 points get projected from 3d to 2d on every frame. that's the sphere and the wave
→ the music comes out of the same script. kick, hats, clap, bass at 120 bpm
→ every cut lands on a beat because the visuals and the audio read the same clock
→ ffmpeg glues the frames and the audio into one mp4
→ change a few words in the script and you get a new video. same beat, same style, new text
→ no stock music, no license to worry about. the beat didn't exist before this script ran
→ the step that mattered: claude looked at its own frames, caught the wave clipping off screen and the code spilling out of its panel, and fixed both before the final render
the honest part: this is abstract stuff. no people, no real footage, no brand assets. for that you still want a human with a real tool
but for a hook video under a post? 5 minutes and zero subscriptions
the prompt is in the reply ↓
@SpXMotivator how do the agents split the work here, one per signal or do they vote before a trade? curious how you keep them from all piling into the same position
@Rorkix@elonmusk he seen / inferred / unknown breakdown is actually really useful. but 13 out of 28 being inferred is a lot lol. curious how much of that holds up when someone who's actually worked on one of these weighs in