vexley

AI tools worth using. Prompts worth saving. Setups worth stealing.

[vexley @vexley_x]

everyone's letting Claude Code hack their TV. almost nobody turns network debugging back off to let Claude in, you turned on developer mode + network debugging and hit "allow" once. that laptop now has a standing pass to your TV until you revoke it do this after the hack, 5 min total:

step 1 → Developer options → turn network/ADB debugging OFF
step 2 → "revoke debugging authorizations" so old keys stop working
step 3 → move the TV to your guest network. it never needed to see your laptop anyway
step 4 → ask Claude to save every command it ran to a .md file. read it once
step 5 → save that file.

the next firmware update will undo half your changes the hack is the fun part. this is the part that keeps your network yours bookmark this before you try it

[Mihir Gupta @0900mihir]

Saw people vibe coding their TVs. Mine had gotten unbearably slow, so I tried it.

One afternoon with Claude Code: 25 bloat apps disabled, the 192 MB stock launcher replaced with a custom 83 MB one, plus rotating wallpapers.

With AI, all you have to do is imagine!

[vexley @vexley_x]

@ekuyda Multiplayer is the right bet. The hard part is when two people in the thread want opposite things: does Wabi pick a side or just keep asking?

[vexley @vexley_x]

@wallacex227 Step 3 is the real lesson: backup first, sort into groups, approve group by group. That’s how you should run any agent on hardware you care about, not just TVs.

[vexley @vexley_x]

@kirubaakaran The wild part isn’t removing 25 apps. It’s that Claude Code wrote a whole new launcher for one specific TV in an afternoon. Personal software for a single device is now a normal Sunday thing.

[vexley @vexley_x]

@Atlas_fomo The real story isn’t TVs. Any locked-down device that speaks ADB just became editable for people who’ve never opened a terminal. Kindle is step two, routers and cars are next.

[vexley @vexley_x]

@alex_prompter Add one line: “Take a full pm list packages backup before touching anything.” Firmware updates quietly re-enable half the stuff you disabled, so you’ll want that list to redo it in 2 minutes.

[vexley @vexley_x]

@justansub ask claude to log every command it runs to a .md file. you get a script to redo it after the next firmware reset

[vexley @vexley_x]

@0900mihir great result. one thing worth adding: turn network debugging off after, and revoke the authorization. otherwise that door stays open

[vexley @vexley_x]

@andrewjiang the part nobody mentions: network debugging stays on after you're done. one toggle in developer options closes it

[vexley @vexley_x]

if your TV runs Samsung Tizen or LG webOS the menu names differ, but the same 3 switches exist. reply with your model and I'll tell you where

[vexley @vexley_x]

everyone's letting Claude Code hack their TV. almost nobody turns network debugging back off to let Claude in, you turned on developer mode + network debugging and hit "allow" once. that laptop now has a standing pass to your TV until you revoke it do this after the hack, 5 min total:

step 1 → Developer options → turn network/ADB debugging OFF
step 2 → "revoke debugging authorizations" so old keys stop working
step 3 → move the TV to your guest network. it never needed to see your laptop anyway
step 4 → ask Claude to save every command it ran to a .md file. read it once
step 5 → save that file.

the next firmware update will undo half your changes the hack is the fun part. this is the part that keeps your network yours bookmark this before you try it

[Mihir Gupta @0900mihir]

Saw people vibe coding their TVs. Mine had gotten unbearably slow, so I tried it.

One afternoon with Claude Code: 25 bloat apps disabled, the 192 MB stock launcher replaced with a custom 83 MB one, plus rotating wallpapers.

With AI, all you have to do is imagine!

[vexley @vexley_x]

@monokern seven bots each owning one slice + sourced lines is the right call. what happens when they disagree, like clean holders but sketchy pools? does the grade show the conflict?

[vexley @vexley_x]

@He1s_Sammy 646 languages for dubbing sounds wild. how's the quality outside English? non-English voices are usually where local TTS falls apart

[vexley @vexley_x]

@beamnxw the config is the real unlock. curious: do you keep easing global or per-element? feels like that's where "tune one thing, break three" starts

[vexley @vexley_x]

@lateascent one thing worth adding: ask Claude to flag the tracking/ACR packages first. speed is nice, but a TV that stops logging what you watch is the real upgrade

[vexley @vexley_x]

@marclou "doesn’t bother you with the details" is exactly why I now have 14 files I’ve never opened

[vexley @vexley_x]

@notjazii Been sleeping on minimax. Might run the same test on my own stuff this week.

[vexley @vexley_x]

@0xOrionVega 20 repos into one pipeline is wild. How long did the full Opus run take end to end?

[vexley @vexley_x]

@Hongji_Leee basically a 24K token LinkedIn bio before every task

[vexley @vexley_x]

my friend’s AI agent was burning $650 a month before doing a single f**king thing

40 tools connected. every call it re-reads all 40 tool descriptions. ~24K tokens of “here’s what i can do” before it even looks at your task

he cut it to 6 tools. same agent, same model, same prompts.

$650 → $97. -85%

here’s the full setup step by step:

step 1 → count your tools. Gmail, Slack, Notion, GitHub, Drive, Calendar, 3 scrapers you tried once in August… each one costs ~600 tokens on EVERY call, used or not

step 2 → run the prompt below. it reads your last 7 days of agent logs and shows which tools were actually called

step 3 → kill everything under 2% usage. split the rest into 2–3 small agents with 5–8 tools each + 1 router in front

step 4 → put the “once a month” tools behind a separate agent you call by name

everyone’s arguing about which model to pick. nobody’s looking at what the agent reads before it starts working

bookmark this before your next bill 👇

[Gipp 🦅 @gippp69]

https://t.co/gTEKJ3i4XU

[vexley @vexley_x]

the prompt paste it into your agent: Read my agent logs from the last 7 days.

List every connected tool with: times called, % of calls, last used date.

Mark DELETE if under 2% and unused for 5+ days.

Mark SPLIT if it's only used together with 1-2 other tools.

Do not disconnect anything yourself. Show me the table.

[vexley @vexley_x]

the math (yardstick: $3 per 1M input tokens): 40 tools × 600 tokens = 24K per call 300 calls/day = 7.2M tokens/day ≈ $650/month

6 tools × 600 tokens = 3.6K per call 300 calls/day = 1.08M tokens/day ≈ $97/month

prompt caching softens this, but every cache miss pays full price. and a bloated tool list makes the agent pick the wrong tool more often too