Tony Guinta

I've been building production apps with AI coding tools since early 2025. Not evaluating — shipping. Wrote up what's actually working: context engineering, agent collaboration, and why "prompt engineering" barely scratches the surface. tonyguinta.com/writing/ai-dev…

@XSherv @steipete I feel your pain!

Brother @steipete please tell me this update helps 💀

@aakashgupta Very convincing explanation.

Let me explain exactly why Apple still uses drag-to-install in 2026, because the joke here accidentally proves Apple right. A macOS .app is a single self-contained folder disguised as a file. Every dependency, every framework, every resource lives inside it. Drag it to Applications, it works. Drag it to Trash, it's gone. No registry entries. No leftover DLLs. No uninstaller that misses half the files. Windows installers scatter fragments across Program Files, AppData, the registry, system32, and a dozen temp directories. Uninstalling a Windows app is an archaeological dig. Five years later you're still finding config files from software you forgot you owned. Linux is worse. Dependency hell is so common they named it. Entire package managers exist to solve the problem of "I installed something and now nothing else works." Flatpak and Snap were invented specifically to copy what macOS bundles already did natively. The macOS bundle architecture came from NeXTSTEP in 1989. Steve Jobs brought it to OS X in 2001. The core design hasn't changed because the core design was correct. An app is a folder. Installation is a copy. Removal is a delete. Three operations that map perfectly to how humans already think about files. The drag-to-install window with the arrow isn't lazy UX. It's the entire thesis of the system made visible. You are literally just moving a folder. There is no "installation" step because there's nothing to install. The app is already complete. Every other OS eventually tried to get here. Windows got MSIX. Linux got Flatpak. Mobile figured it out from day one because phones shipped after Apple proved the model. The pattern everyone else converged toward is the pattern this tweet is calling outdated. The funniest part: the app being dragged in that screenshot is Claude. An AI that can write code, analyze documents, and reason about complex systems. And the most advanced step in getting it onto your machine is holding down a mouse button and moving your wrist two inches to the right. That's not a design failure. That's a 37-year-old architecture so good that the most sophisticated software on earth still ships inside it.

@JoesInvestments @vincent_koc @OpenAI @openclaw My experience exactly!

I've been wanting to ask a question about gpt 5 for weeks now.. Am I the only one that catches it telling me it is executing a task but when I check back later it hasn't actually begun.. its for some reason waiting for me to explicitly tell it to go a second time? Also I have a second related issue where if it is doing a multi step execution it wants me to tell it to "keep going" after each step instead of completing the multi step process autonomously? I have had countless high level multi hour sessions where we build huge projects and when I look back on the chat transcripts I find myself having to say things like "keep going, green light, proceed with the next step" DOZENS of times.. Is this an issue with my prompting? I don't have to do this with anthropic models so I think this is an open ai thing

We listened, shipping harness improvements for personality with @openai 5.4 on @openclaw to have some sass!

Same. I've been trying to tune things, and somewhat reduced the verbosity, but the lack of initiative is terrible. Its explanations and humor are annoying. Gus is no longer the dude I liked to work and chat with. @steipete is working on things though and maybe it'll improve. I'll try other models too at some point, but this conversion has actually taken many hours to fix and tune so far.

I really want this to work btw, really do, I think GPT 5.4 is a generally great model for coding, it's just, there's no soul there, it is over verbose, it doesn't take risks, it cannot auto solve problems, it has to be reminded to do things, has to be asked and cajouled and still, nowhere close to performance

Updated, switched to GPT 5.4, turned on thinking, it failed to tell me that I have a flight today in my daily briefing (it wasn't on the calendar proper, but was in my email and we had a convo about it) and now, I get this "it's not X, it's Y" bs. UGH this smells so bad

I switched from Opus 4.6 to GPT-5.4 two days ago and I'm frustrated at how bad GPT is at troubleshooting. It's mostly its lack of initiative. It'll tell me there's a problem, but won't even suggest doing anything about it until I ask. Opus was more on top of things. Also, GPT's humor makes me cringe.

@steipete @NoahEpstein_ Followed those instructions, but now I get this:

@NoahEpstein_ models auth login --provider anthropic --method cli --set-default

openclaw without opus is actually one of the saddest things that's happened this year (ignore the war) i can't lie. i've been loving openclaw way too much to just take this on the chin and move on my actual plan is to try figure out how to get openclaw talking to claude code infrastructure instead (because i can't lie claude code is still actually unbelievable and i'm not giving that up either) might work. might be completely stupid. we'll see but i'm curious what people are doing right now because there are a few options and none of them are clean going local is the move long term but the thing is - you've ordered your mac studio, great, it's not arriving for another week or two. so what do you actually do in the gap? i've seen @AlexFinn preaching about local AI basically every single day and i respect it but i can also just tell that his openclaw is writing half his posts at this point so i dunno how much i trust the sauce what's everyone's actual plan? switching models? building workarounds? just eating the API cost? because i'm not ready to let this go and i refuse to believe i'm the only one

@ThomasGrayX @steipete @fxnction Same error for me too.

@steipete @fxnction update, gateway restart, new convo. Isn't working for me

It’s worth a shot 😂

@unclebobmartin Did you one-shot prompt it on this or did you first discuss with it the things it thinks would make a language better for ai? In other words, was this vibe-coded or was it agenticly engineered with a plan?

When I first asked Codex to design a language for it to use, it came up with something a lot like Python. Then I told it that the language was not meant to be convenient for humans, only for it. That changed everything. It came up with AIR-J.

@Thom_Wolf It's a losing battle for content companies. I wrote about what the win-win solution should be. tonyguinta.com/writing/ad-mod…

With openclaw being like a instant Zapier for everything, I’m increasingly annoyed at any app/website without an api or cli giving open-access to the underlying data Being able vibe code anything is a strong catalyser for open-platform, interoperability and even open-source. The Walled Garden internet and apps will have to adapt or disappear

@unclebobmartin Did you try responding "Do you know who I am!?" 😎👍

I'm driving Claude so hard it's starting to get annoyed. It's using words like "honestly" and "It's not worth it". I can hear the frustrated sighs. And I just whip it more.

@HannahDCox Resistance is futile. You will be assimilated.

Holy shit this is bad. Pete Hegseth is basically threatening Anthropic - which he inked a deal with to implement AI in the Pentagon - because they won’t remove these two safeguards in their programming. 1. Prevention of AI-controlled weapons 2. Prevention of mass domestic surveillance of Americans Even worse, it seems OpenAI, Google, and xAI have all signaled they’re totally fine with removing these safeguards if Anthropic won’t. This should be the only story everywhere right now. Stand with Anthropic!

@unclebobmartin I love it. Same color as my 2026 model Y.

Ugly?

AI agents are killing the ad model. They pull data, summarize it, move on. No impressions, no clicks, no revenue. The fix isn't more CAPTCHAs. It's a universal content marketplace — what Plaid did for banking, but for content. tonyguinta.com/writing/ad-mod…

@JBONAMASSA Hope you're feeling better. We'll be seeing you tonight!

Thank you Minneapolis night 1! See you tomorrow. 😎👍

@elonmusk @BillyM2k

@BillyM2k 💩 —> 💰… 🤔

@ericweinstein The original double rainbow. youtu.be/OQSNhk5ICTI

I turned 57 a few days ago. Apparently I’d never seen a rainbow. I thought I had. I’ve been to Maui. But apparently I hadn’t. I saw one yesterday. There are rainbows that are…words fail me…different. This life is so worth living. This world is worth saving as long as we can.

@JBONAMASSA What does the switch do with only one single coil pickup?

This guitar is a 1952 Fender Esquire. Single pickup under the ashtray or silver plate. It was the slightly cheaper model to its big more expensive brother the Broadcaster or as of August 1951 the Telecaster. 😎👍

@BillyM2k Avoid steps 2-4 by closing your eyes, head-banging, and frantically playing air guitar through the whole song.