Curt

Just another small time American

[Curt @cpjet64]

Always the thinkpad. Over the course of decades of work in IT professionally I have never seen someone retire with their MacBook but I have seen plenty of people retire with their thinkpad. I wouldn’t say one is more professional looking than the other but the companies that have thinkpads tend to be places you never have to worry about getting fired from or replaced by someone younger/cheaper.

[Curt @cpjet64]

@vxunderground HEY!!!! I have been working on this exact thing since nghtmare eclipses situation first became public. I do have the infra for it already.

[Curt @cpjet64]

We have metal blade serrated kids knives that cut almost as good as a sharp blade. The kids use them as early as 1 1/2 yo and just like with gun safety where you teach always treat it like it’s loaded we teach the kids the knives are always sharp. Start as young as possible and crawl walk run. Build the mindset while you build the dexterity. Everything else comes with time.

[Curt @cpjet64]

@EthanLipnik if youre running into this issue then you need to start using /goal

[Curt @cpjet64]

@uzairakrum Just curious because i have not been having the same experience, what makes you think 5.5 is dumber?

[Curt @cpjet64]

@jarredsumner You can also do this with claude code on opus 4.6 or 4.7 but dont try it on 4.8 you will just be wasting tokens.

[Curt @cpjet64]

This is honestly one of the biggest issues I have been seeing on X from devs since multiagent became a thing. Break out of the engineers mindset and start thinking as a manager is the best advice I can give. I have written my own agentic harnesses from scratch, have merged PRs with Codex official, and have been a big supporter of agentic coding for a hot minute now. Take a step back and think of the end goal you want to reach. Tell the agent where you are starting, where you want to end, and if you dont have custom guardrails then also tell it roughly how to get to the end. A single paragraph should be more than enough for the agent to build a fully sliced worktree compatible plan that is constantly running out of agent slots. I wont lie the first few times you do this you are going to burn a lot of tokens until you get the hang of it but once you do just imagine 20 or so 5.5 low reasoning agents ripping through 20k lines of code in as quick as a few minutes. specify what reasoning level for subagents to use. its usually cheaper at this point to throw as many low reasoning agents as possible at the issue vs a single xhigh reasoning agent of the same model.

[Curt @cpjet64]

@doodlestein ok so quick question... if there is a conflict and the runnerup aborts how do you gaurentee against data loss? the details were a little light on that segment

[Curt @cpjet64]

@doodlestein ok great. yeah i saw it in the deps and was like ... wtf ... lol

[Curt @cpjet64]

@doodlestein is rusqlite stable? i thought it was still very experimental last i heard...

[Curt @cpjet64]

@doodlestein i dont know where from but i couldve sworn you were using custom versions inside your orch tool. huh...

[Curt @cpjet64]

@doodlestein Actually can you link my frankensql I think I’ll rip out a port really fast to test it

[Curt @cpjet64]

@thsottiaux @svilendonmez @sama this particular /goal has been running nonstop no breaks and no blockers. proper prompting is a beautiful thing

[Curt @cpjet64]

@thsottiaux @svilendonmez @sama thats some amateur shit and piss poor prompting xD you can optimize kernels on 5.5 low

[Curt @cpjet64]

@mark_k im making one xD. its very different yet very similar. everything from scratch using nothing more than rust std.

[Curt @cpjet64]

repeat after me. llms are toddlers with doctorates and need guardrails and training wheels. trust but verify often. i actually downgraded my claude sub this month because opus 4.8 at every level of reasoning has just fallen off a cliff and is almost completely useless now. gpt5.5 is running circles around it, so is deepseek, so is grok. 4.7 was really nice but you had to prompt it correctly. 4.8 is just, well... dumb.

[Curt @cpjet64]

@ravikiran_dev7 imagine having automerge enabled and having that monstrosity submitted lmfao 🤣

[Curt @cpjet64]

@bettersafetynet yep. here it is! right on their site.

https://t.co/I7HgTb611E

[Curt @cpjet64]

Your signature authority should have it stated very loud and clear that just signing with their expensive key does not guarentee a pass on smartscreen because it doesnt. smart screen has its own reputation system and yes the code signing helps but its not a guarentee. it should clear up on its own after enough systems have submitted the signature though.