Van de Rook | AI Tools, Tested

Watching the AI feed from above. I keep the sharp tools, drop the shiny junk

[Van de Rook | AI Tools, Tested @Vanderook]

This is f*cking wild. All 10 of my Claude Code runs tried a command my allowlist didn't cover.

10 of the 15 blocks were plain for-loops, 2 were heredocs. All 10 runs still passed the hidden tests. All 4 Haiku 5.5 runs hit a block, decided Bash was off and never ran their own tests.

This prompt tells Claude to wait for an OK before the cleanup and to keep going on its own the rest of the time. That's an instruction. What Claude Code actually stops on is its permission settings.

On a TV over ADB I'd allowlist the adb commands I can read and approve everything else by hand.

[codila @0xCodila]

Claude Opus 5.5 just reverse-engineered my TV...

1 prompt + 5 min = total control: 10x faster UI, ads blocked, any app from GitHub and that's just the beginning

here's full guide step by step:

step 1 → put your computer and TV on the same Wi-Fi network

step 2 → paste the prompt below into Claude Code with a photo of your TV

step 3 → unlock the hidden Developer menu with your remote - Claude tells you what to press

step 4 → hit "allow" on the TV - it's last time you touch the remote

step 5 → say yes to the cleanup: no ad rows, no bloat, 2x faster UI - all reversible

step 6 → install any open-source app straight from GitHub - no Play Store

the result: now anything you can think of, your TV can do...

p.s: works on almost any TV. Mine in the video is from 2017

Try it now, you're welcome. Then read article below on how to reverse-engineer anything:

[Van de Rook | AI Tools, Tested @Vanderook]

@rost_roma One HANDOFF.md my agent writes at session end: what it tried, what failed, which tests are flaky. The next agent starts as the heir. Does Beacon keep the failed attempts too?

[Van de Rook | AI Tools, Tested @Vanderook]

@_asadmemon @Tailscale @mitchellh @_asadmemon how many lines of Codex diff stay readable on that kindle screen before the e-ink refresh smears them? Did you cap the streaming speed, or let ghostty redraw at its own pace?

[Van de Rook | AI Tools, Tested @Vanderook]

@ayushagarwal 5 of 10 Claude Code runs in my tests read one or two blocked commands as "Bash is denied" and quietly switched to a workaround. Did that happen to you over ADB when a command was refused? And did the home button fix take more than one try?

[Van de Rook | AI Tools, Tested @Vanderook]

@Plutkar for a prompt, agreed. that's why I use it. for a voice agent deciding whether to cut in or wait, the um is the data. I'd keep both: clean text for the LLM, raw audio cues for the turn-taking call

[Van de Rook | AI Tools, Tested @Vanderook]

ONE CLICK turned a 1.5-HOUR STREAM into a clean summary. That's this Article's problem in one feature.

I ran a stream through Wispr Flow's Notetaker and got a clean summary: topics, decisions, next steps. You can tell it up front what to pull out, and it does.

Dictation works the same way. It cuts the "uh"s and the repeats, adds paragraphs and punctuation, and turns a rambling request into a usable prompt.

For me that's the point. For a voice agent it's the signal this Article says gets lost: who hesitated, which point someone kept circling back to, what got taken back a sentence later. The summary keeps what was said and drops how it was said.

[Naman Jain @namanjn1998]

I think the biggest gap in voice AI is that agents make decisions from a transcript, not from the audio. I wrote down my thoughts.

I’d love to hear where you disagree. https://t.co/hHmfjlbFKG

[Van de Rook | AI Tools, Tested @Vanderook]

Before you believe "not possible on your TV", ask Claude for the exact command that failed.

This Article mentions people who got "this isn't possible on Tizen" from Claude and stopped there.

I've seen that move in my own logs. 10 Claude Code runs, same command allowlist. In 5 of them, one or two blocked commands later the agent wrote "Bash is denied in this session" and stopped using it. Their first Bash command had worked. They quietly switched to a workaround.

On a TV the blocked command is an sdb or adb call, and the workaround is "not possible". In my runs it was one refusal, read as a wall.

[Nathan Clark @nathanclark]

https://t.co/99SaneJ2mp

[Van de Rook | AI Tools, Tested @Vanderook]

tip if you try Notetaker: tell it before the recording what you want pulled out. it builds the summary around that instead of giving you everything

if you want to try it, my referral link gets you 30 days of Pro instead of 14 (and me a free month): https://t.co/bwYftUiwMC

[Van de Rook | AI Tools, Tested @Vanderook]

the trick that makes it work: the page never plays. it reads the time from the URL hash (t=12.0), sets every animation to that moment and holds still. headless Chrome takes one screenshot per URL, so frame 360 is always t=12.0 and always looks the same

8 frames render in parallel, and a resume flag skips the ones already done. all on-screen text sits in one COPY object, so changing a caption never touches the motion code

[Van de Rook | AI Tools, Tested @Vanderook]

720 FRAMES, 24 SECONDS, ZERO VIDEO APPS Claude writes the scene, headless Chrome renders it.

I make my explainer clips the way this Article describes. The scene is HTML/CSS, each frame is a headless Chrome screenshot, ffmpeg stitches them into the MP4.

What made it reliable:
a timeout and resume per frame, because headless Chrome can hang on a single one
review by exact frame, not by eyeballing playback
a dark plate under any caption over a busy background
every number on screen pulled from the raw data, never typed by hand

Agree with step 10. My addition: the critic checks numbers against the source data, not against the brief. A frame can look perfect and still be wrong.

[Mori @0xjmori]

Claude Opus 5.5 could turn a single AI bot into an entire motion design studio.

It can direct scenes, animate product interfaces, render videos, catch visual mistakes and deliver polished results.

In this article, I show you how. https://t.co/HjZIAksasB

[Van de Rook | AI Tools, Tested @Vanderook]

@higgsfield_ai Was this the first take from each model, or the best of several? A single run each and a best of five are different tests. Same prompt and the same number of attempts per model would make the feather rotation a fair comparison.

[Van de Rook | AI Tools, Tested @Vanderook]

@webasamoah @pizzamanghana 720 frames for a 24 s clip at 30 fps, each one a headless Chrome screenshot of an HTML scene, stitched into MP4. That is how I make mine, and why the clip stays editable. @webasamoah, is yours Remotion or something else? A lot of people here are asking for the how.

[Van de Rook | AI Tools, Tested @Vanderook]

@0xjmori Step 10 scores motion quality from 1 to 10, and a score like that can drift between runs. I'd add pass/fail checks the Critic can't argue with: text under a minimum size, a caption on the cursor path, two elements overlapping in one frame. Then the 1-10 is only for taste.

[Van de Rook | AI Tools, Tested @Vanderook]

@garybasin Does trace show which agent session wrote a given commit? With 3 builders and a reviewer running in parallel, finding the bad commit is the hard part. That is the first thing I would check in your demo.

[Van de Rook | AI Tools, Tested @Vanderook]

@whydeso 8 of 8 planted bugs: Haiku 5.5 in about 11 s, Opus 5.5 in 17-20 s. That is a test you can score. Guessing OpenAI’s next name can’t be scored, since neither model knows it. Would you run DeepSeek V4.1 on a checkable task and show time and cost per run?

[Van de Rook | AI Tools, Tested @Vanderook]

@whydeso @AIMind_Ai true

[Van de Rook | AI Tools, Tested @Vanderook]

@0xCodila nice demo asking for a yes before the cleanup and keeping it reversible is the right way to let an agent touch real hardware, and showing it on a 2017 tv is more convincing than a new one

[Van de Rook | AI Tools, Tested @Vanderook]

@stretchcloud a cost column per ticket would sell this for me. agents as teammates is great until nobody can tell which issue burned the budget

[Van de Rook | AI Tools, Tested @Vanderook]

@tekari0 curious how this looks split by task type. an index average can hide where the cheap model ties Opus and where it falls off a cliff