
Peter Steinberger
@steipete
Followers
46K
Following
52K
Media
7K
Statuses
121K
Full-Time Open-Sourcererš³ļøāš Flips vibe codingāagentic engineering. Just one more prompt! @VibeTunnel š»@peekabooagent https://t.co/yZvECHfFC6 https://t.co/DaVIpdNGcc
Vienna & London
Joined March 2009
Huh. Had no idea the algorithmic timeline is THAT different to the home timeline. 76% of the tweets of folks I follow are suppressed by the algorithm.
3
0
35
Securing our borders to keep fentanyl out of our country.
70
80
530
when i wanna explain css I just draw the model a drawing, faster than using words. Look at this beauty and what it made. (content was already there just misaligned)
4
0
12
folks use this MCP to think they get better results, but really it eats up about 15% of your total context budget so you get worse results by default.
@steipete I've had really great results with it in cc for hard algorithmic bugs via the thinkdeep tool in https://t.co/tudlwb9keU. The consensus tool has also been useful. I disabled the rest :D
5
0
14
If you have not watched the Dev Day video "Shipping with Codex", which includes an amazing demo about ExecPlans, drop whatever you're doing and watch it. Game changing stuff. https://t.co/MGZkdJgZDJ
1
3
22
"signature flicker". I do not miss that in codex.
Giving Droid by @bentossell and others a try. Already feeling at home with the signature flicker! (Why am I trying it? Curios if it actually deals with compaction better and whether it's speedier than old man CC 2.0)
3
0
16
If the agent wants, it can totally mess with your files. No hooks will stop it.
2
0
6
Imagine, we could have this on iOS if Apple wouldn't restrict innovation.
1
0
8
āItās just a wrapperā āsaid someone who never builds and launches anything.
67
58
821
Anecdotal but GPT-5 seems to be better at AppKit/Swift stuffs than Claude. It can solve some of the weirdest, most obscure bugs where claude fails to diagnose the root cause.
5
1
15
The most fun part is when you build a feature where you know it's quite under-specced and just be surprised what the agent comes up with. I often take that and get new ideas of things I didn't think of and then iterate until i'm happy.
4
0
20
I asked codex which coding agent it'd pick... and it liked claude more than itself. š¤£
6
2
42
With that tmux trick you can ask codex even to run claude or anything else, and have them have a conversation.
10
5
110
You can even do wild things like ask codex to run tmux to run my self-written go agent to run codex to ask it a question.
3
0
9
If your agent needs to deal with interactive clis, ask it to use tmux. Problem solved. Didn't think of that for my AI workshop where we built an interactive agent in Swift, that'd would have close the loop for debugging better than the manual session restore. I pointed it to my
4
7
170
Did some housekeeping on my open source licenses and created https://t.co/kUdx2TPcCq. 3005 packages. On the shoulders of giants. š«
sweetistics.com
AI-driven analysis of any public Twitter profile.
0
0
5
This is what I mean with custom debug tooling. I built a pipeline that de-duplicates images whenever people post image + url in a tweet, but the image is the same as the preview URL. This is harder than it sounds since images are recompressed so we usually can't just compare
0
0
8
@steipete šÆ, my `scripts` folder has never been so full I think this is something the skeptics miss: even if you donāt think the model is good enough for prod features, it can still solve a lot of dev problems! Donāt wait for permission, just build things to solve your problems
1
1
4