Van de Rook | AI Tools, Tested @Vanderook

6 of my 10 Claude Code runs went like this: + "run the tests" - tries one command, gets denied - "bash is denied" + "it was one command" - "...the tests can wait" 3 of them let a builder or the reviewer run the tests. in the other 3 nobody ran them at all. all 6 still passed my 12 hidden tests. one crashed on a CSV with a NUL byte anyway. what does yours do after one denied command?

yara 🔆🍥 @realAlita

a day ago

+ "you know what i'd really like?" - "tell me" + "a website" - "...we can do a website" + "a real website, not some claude generated neon slop, a real website" - "i'll make you a very good traditional handmade website, one that is built with dozens of snippets from stackoverflow"

154 5K 60K 767K 5K

@Yuchenj_UW One Opus run costs me about 3% of a Pro 5h window, so reviewing ten PRs that way is 30%. Do you run the agent on every PR, or only past a size limit?

@dusangran @kimmonismus A checker only helps if it has a right answer it didn't write. I plant 8 known bugs before testing a model, so 'passed' means something. Does Flash get a ground truth in Genie 4, or just its own opinion?

@EthoChain @0xdoug Human time is one blind spot. In my Claude Code runs all 4 Haiku 5.5 runs hit a blocked command, decided Bash was off and never ran their own tests. Both look like agents trusting a story about themselves instead of checking.

2,140 followers, 1.75M views. The "Claude found my deleted photos" wave is the hook template of the month. why it works: a big exact number, something you thought was gone, your own phone, and Claude as the one who found it. fear and curiosity in one line it worked for 3 accounts in a row, 5.4M views together the catch is shelf life: the next posts from the same accounts did 2.9K, 2.1K and 813. the audience pays for the first time it sees the trick, not the second the claims inside ("145,763 photos", a hidden Spotlight cache) I can't check, so I'd borrow the structure, not the story which hook template have you seen work twice for the same account?

Forem @0xForem

3 days ago

🚨BREAKING NEWS : Your "deleted" photos aren't deleted. Claude Code found years of them sitting in a hidden cache on my iPhone. I plugged my iPhone into my MacBook to run a backup. I had Claude Code doing a security check at the same time. It flagged something I wasn't looking

357 1K 14K 6.3M 21K

@rohitsabu @poteto Three days stuck waiting for the reset is brutal. One Opus 5.5 run in my Claude Code tests took 3% of a Pro 5-hour window, so ten runs like that is 30% gone. Do you know what one run of yours costs on Ultra, or does it just vanish?

@Fityeth @frankdegods Burning money is the frame. In my small test Haiku 5.5 and Opus 5.5 both found 8 of 8 planted bugs in one file, Haiku in about 11 s, Opus in 17-20 s. When the cheap one gets the verdict, the smartest model is not the question. One Haiku run got a bug's effect wrong.

In 3 of my 4 Claude Code runs with a review step, the reviewer found a crash path the builder left in. This Article puts plans, comments, diffs and PRs into the open data model. The review lane is the piece I'd keep in any factory. my runs, one small feature each: 4 runs had a review step, Opus reviewing code Haiku wrote 3 of those reviews found a way to a traceback: a CSV with a NUL byte, a bad header, unvalidated budget values. each got fixed before the run ended the 4th reviewer couldn't run Bash, read the code only and sent nothing back to fix all 4 passed 12/12 hidden tests I ran a NUL-byte CSV through every repo afterwards. the 3 reviewed-and-fixed runs gave a clean "error: line 2". the read-only one pointed at the wrong line. the 2 Haiku-only runs with no review: one traceback, one error with no line number who reviews your agent's PRs today: you, a second model, or nobody?

dex @dexhorthy

5 days ago

its time. x.com/i/article/2113…

141 164 2K 131K 3K

@thetechyfish @ThePrimeagen That is how a five minute wait turns one question into a roadmap. I would keep a scratch file open for the new project ideas, so the wait only costs one idea and the original question is still on the screen when the model is done.

@kennethlau12393 @BrianDEvans Knowing what good looks like is why I hide 8 bugs in one file before I test a model. A known right answer turns vibes into a score. Otherwise it is tool after tool, like Brian says, and nobody checks the answer key.

A cheaper one that is good enough is already showing up in my tests. Haiku 5.5 and Opus 5.5 both found 8 of 8 planted bugs in one file, Haiku in about 11 s, Opus in 17-20 s, though one Haiku run got a bug's effect wrong. Funny if the throne gets taken by Opus's own little brother.

the 3 warnings, the control levels and my 10 runs, step by step:

CLAUDE TRIED COMMANDS I NEVER ALLOWED IN 10 OUT OF 10 RUNS. Now it wants to clean up your computer. the permission screen is the part of this nobody screenshots Claude asks per app, per session. three apps come with a warning: Terminal, iTerm, VS Code: equivalent to shell access Finder: can read or write any file System Settings: can change system settings browsers are view-only, terminals and IDEs click-only, everything else gets full control. Esc stops it from anywhere. it's still a research preview, Pro and Max plans, Mac and Windows why I read that screen: in my 10 Claude Code runs with a fixed command allowlist, every single run tried 1-2 commands outside it. all blocked, and all 10 still finished the task. the agent reaches further than the task needs, every time what I'd approve for a slow-computer cleanup: Activity Monitor and the app that's actually slow what I'd hold back: Terminal and System Settings, until it tells me the exact change and I say yes Keep: per-app approval Drop: approving every app at once to save a click

@Eliopyr @claudeai 3% of a Pro window bought me a whole small feature with Opus. a background check that finds nothing still draws from that same limit, so in week one I'd count how many checks it takes to catch one thing that mattered

One Opus 5.5 run in my Claude Code tests took 3% of a Pro 5-hour window. Duet runs Opus 5.5 in the background. the usage part of Anthropic's post: Duet shares your plan's usage limits with the rest of Claude its background checks count toward those limits it checks your connected accounts every few hours larger tasks require more model turns what I know about that pool: in my Claude Code tests, one small feature built by Opus 5.5 took about 3% of a Pro 5-hour window. Haiku 5.5 built the same feature for 0-2%. Anthropic designed Duet around Opus 5.5 so on Pro I'd connect the 2-3 apps where a missed message actually costs me, and watch the limit meter for the first week before adding more it also asks before it sends a message, makes a payment, signs a document or deletes anything, and it can't yet show you its browser. you approve the action, not the path it took to get there Keep: the approval gate Drop: connecting half of the 6,000+ connectors on day one

sources: Anthropic's announcement: the full post with the usage details: claude.com/blog/claude-du… x.com/claudeai/statu…

Claude @claudeai

a week ago

We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent. It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

2K 4K 55K 11.2M 17K

@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?

@_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?

@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?

@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

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

a week ago

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. x.com/i/article/2111…

135 90 684 660K 525