Marc Hembach

Testing agents. Shipping products. Breaking things.
Sharing what actually works.

[Marc Hembach @P4tchdd]

@0xSero Right now I’m only working with subscription and just want to change that and become a local guy. Right now the prices in Germany are exploding..
A 3090 would enable me and in the feature I would like to get a second one

[Marc Hembach @P4tchdd]

@ugo_builds Great Idea🔥

[Marc Hembach @P4tchdd]

@Nick_333ch https://t.co/D230VpVaTa a OS to show menus, browser pages, apps and even more.

[Marc Hembach @P4tchdd]

@imryanarya007 https://t.co/D230VpVaTa

[Marc Hembach @P4tchdd]

@rdbuilds7 https://t.co/D230VpVaTa i just built a web operating system. It was funny to do.

[Marc Hembach @P4tchdd]

And we're live!

Late to the party. Brought a web operating system.

https://t.co/Wc2Fetx6Cd

[Jonathan Wilke @jonathan_wilke]

Aaaaaaand we're live!

https://t.co/ZElVA2hguv

No ads. No API keys. No revenue sharing.

Just outbid your competitors to rank #1 and consider marketing done for today 🚀

[Marc Hembach @P4tchdd]

@tdinh_me 150 agents is a harness problem. One good loop beats a farm you cannot review.

[Marc Hembach @P4tchdd]

Anyone using Codex or Grok heavily for coding?

Have you found any tricks to reduce the built-in system prompt / context overhead?

Claude Code gives you a few levers here.

Curious what you guys are doing with Codex/Grok.

[Marc Hembach @P4tchdd]

@MakerThrive @Replit Great work 🔥

[Marc Hembach @P4tchdd]

@simonw That's the part I still don't want auto. Write the workflow. Don't push it.

[Marc Hembach @P4tchdd]

@bcherny @dedene Coding is the cheap part. Catching the wrong-diff review still needs a human.

[Marc Hembach @P4tchdd]

@theo A smarter model that dies mid-file is worse than one that still has meter left.

[Marc Hembach @P4tchdd]

@theo That's the job now. Not writing the review. Checking it looked at the right diff.

[Marc Hembach @P4tchdd]

Humans approved 97% of the permission prompts.

Anthropic planted dangerous commands in front of 1,053 paid Claude Code users.

People caught 13.6%.
The classifier caught 89%.

The dialog is not a safety layer. It's a habit.

If the agent can touch the repo, a deny-list and a worktree beat another Allow.

[Marc Hembach @P4tchdd]

Everyone's getting cut right now.
I'd rather pay for a meter I can use than a smarter model that dies at 80% of the file.

Grok on Ultra is the one that still lets you work.

[Marc Hembach @P4tchdd]

I ran Grok super heavy on Cursor Ultra this week.
Same pile of work I already put through Claude, Codex, and Kimi.

Usage is just more user-friendly. You actually finish the session.