Zerixin

Learning AI by building with it // Sharing the wins and the mess

[Zerixin @zerixin]

Turning off auto firmware updates is the real win here, since that's how some printers start rejecting third-party ink overnight. I'd add a step 7 though: set a password on that admin page, on plenty of models it has none by default and anyone on the Wi-Fi can open it. Did it set up driverless through AirPrint/Mopria or plain IPP Everywhere?

[Zerixin @zerixin]

I f*cking fully customized my TV's UI with Claude Code. no ads before Netflix, no junk rows, ~15 minutes. turns out I'm not alone: the post that started this hit 6.5M views in 3 days, and the repo now covers 92 TV models from 31 brands

here's what to know before you let Claude Code into your TV:

1. what it actually does
backs up your package list and settings, disables ad and telemetry apps with pm disable-user (nothing uninstalled, nothing rooted), then installs a ~370 KB launcher built for TVs with 1 GB of RAM.

2. which TVs
TCL (18 models), Sony (16) and Xiaomi (15) lead the list, plus Philips, Hisense, Panasonic, Toshiba, NVIDIA Shield, Chromecast with Google TV and even XGIMI projectors.

3. not only Claude
the README says Codex or Gemini CLI work too. no AI at all? the scripts run on plain Python 3.

4. Samsung, LG, Fire TV
Samsung (Tizen) and LG (webOS) aren't covered. Fire TV gets the cleanup, but Fire OS won't hand the home screen to a new launcher.

5. regret it?
tools/restore.sh replays the backup and gives Home back to the stock launcher. TV unreachable? re-enable the apps in Settings → Apps.

6. the part nobody mentions
network ADB runs on plaintext port 5555. on some firmware anyone on your Wi-Fi can get a shell. turn debugging off when you're done.

would you give an AI agent a shell on your home network just to kill a few ads?

drop your TV brand below and I'll check if it's on the list ↓

[Gathin (Ø,G) @gathint]

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

[Zerixin @zerixin]

I f*cking can't believe my eyes...

Google just open-sourced Harness², and agents now rewrite their own harness.

the model stays frozen. what changes is everything around it: system prompt, guardrails, memory, skills, subagents. in Harness² they're plain files the agent is allowed to edit.

how it works on a new task:
→ the agent proposes a few edits to its own harness
→ runs the candidate versions side by side
→ compares how the runs behaved and keeps what worked

no answer key or grader needed. it learns from the differences between runs.

and it doesn't start from zero next time. the harness it settled on + the lessons that held up get saved, so every finished task makes the next round of edits smarter.

for agents already in production, the same loop runs on failures: bad trace → root cause → fix → regression test, so it never comes back. @akshay_pachaar broke that loop down here 👇

so be honest: if the agent can tune its own harness better than you can, what's left of "prompt engineering" as a job?

[Akshay 🚀 @akshay_pachaar]

https://t.co/YUGDPxpYvy

[Zerixin @zerixin]

@tezzarc Pairing over the encrypted protocol with the on-screen PIN is the step I'd have bet would stall it, so that's impressive on a 2014 set. Did Claude work out the handshake itself from network traffic, or find an existing library and wire it up?

[Zerixin @zerixin]

I CUT MY TOKEN SPEND BY 80%

the cleanest Claude Code setup I've seen this month

Opus writes the plan as a script once, then a swarm of Haiku 5.5 agents does the heavy lifting in parallel

the script runs the loop, so Opus never drowns in agent transcripts. it only sees the final diff

start Claude Code like this:

claude --model opus --effort ultracode

who does what:

→ Opus 5.5 turns your goal into a workflow script
→ Haiku 5.5 scans every file at the same time at 340 tok/s and flags candidates
→ Sonnet 5.5 fixes each one in its own isolated copy and double-checks it
→ your Opus context stays clean

Opus only steps in 3 times:

→ before the run: writes the plan
→ after the swarm: reads only verified findings
→ before git: signs off

Opus plans · Haiku scans · Sonnet ships

JEV takes the same idea one level lower: tiny decisions that don't need a big brain (which file, which tool, retry or stop) get routed through JEV in milliseconds, and the big models only see the forks that actually matter

the full flow
> Opus 5.5 runs the main session and writes workflow.js
> Haiku 5.5 agents audit every file in parallel
> validators drop false positives
> Sonnet 5.5 patches each fix in an isolated worktree
> a verifier cross-checks every diff
> Opus 5.5 reviews only what passed
> JEV routes micro-decisions in milliseconds

copy the flow + this prompt into Claude Code ↓

"Set up my Claude Code around this flow:
1. Main session on Opus 5.5 with ultracode
2. In every workflow: read-only scans on Haiku 5.5, edits on Sonnet 5.5, each edit in its own isolated copy
3. Add to CLAUDE.md: every finding gets an independent verifier before it's reported, and fix loops stop after 2 rounds with no progress
4. Route fast deterministic decisions (tool choice, path selection, retries) through JEV
Show me every change as a diff first. Don't edit anything until I say go"

bookmark this, you'll need it 🔖

[mirku @mirku21]

https://t.co/PaAzxlyrsH

[Zerixin @zerixin]

@himanshubuildss DESIGN.md is the sleeper pick here. Handing the agent a written design system up front fixes more "AI-looking" UI than any component library. Which one do you reach for first on a new project?

[Zerixin @zerixin]

This stack is one of the cleaner setups I've seen lately. It actually turns Claude into something closer to a real build system instead of just a chatty coding assistant.

Serena + Context7 fix the two things that usually break agent workflows on bigger repos (symbol navigation and stale docs). GitHub MCP + E2B let it actually touch real issues and run code safely, while pytest as a hard gate and Langfuse for traces keep it from just generating plausible garbage.

The video is also well done — the 3D progress view makes the whole loop feel less like a black box.

I'd probably throw in a lightweight lint/format step and a quick second-pass review before the tests, but the core idea is solid. Nice one.