Everything reinforces my choice of Aider+Openrouter+code. I use models 20x cheaper than Sonnet, completely pay as you go, nobody is playing games with models, their LLM bills and my context to turn a profit. I generate Python scripts with Mongo and Cloudwatch queries instead of getting angry at the mystical sparkly button UX on their apps.
I paid for 5$ of credits from OpenRouter and looked up their hosted models. I installed Aider and run it in watch mode in my project. I then create comments with instructions at various files (like #refer this ai) and add ! to last command to execute it. I have disabled autocommit but I commit frequently manually to recover when it acts dumb and overwrites stuff.
"AI started actively breaking working code" is vibe coding talk, you should monitor what it does and undo at the step that breaks. I've been able to do that just by prompting "undo last step/<specific action>"
I used it before as a free Claude Code to do ad-hoc scripting, pretty useful. Now I have found a bunch of TUI programs (Amazon Q, rovo dev, opencode) that can assist with that kind of workflow. I don't like that it's a GUI app, but I like the fact that it's a terminal app where I can type command directly into, not just prompt. Claude 4 should be fine, was on free, never hit limit so not sure what's the lite experience is.
It feels analogous to what would happen if you put me in front of a broken project without source control that I've never seen before and asked me to fix it without giving me enough time to actually understand it. It starts from errors and bugs, guesses corresponding source code, and tries to narrowly fix just that snippet. Generally it favors deleting, but not specifically deleting new code.
I would have thought it could record a log of its actions and use that log to think about undoing. I would also think it could annotate lines with git blame so it knows undoing wouldn't involve changing anything more than say a day old. Unfortunately that isn't consistent with what I've seen.
I just make a WIP git commit and run git commit -A --amend --no-edit after manually reviewing each unit of work.
Edit: I also wish Claude implemented undo at a higher level instead of relying on the model. Some combination of snapshotting the whole repo and tracking operations that have precise inverses. But I understand that would have drawbacks.
maybe a system prompt to tell it to do checkpoints. Stash is one way, another is jujutsu with git backend. If this was Claude Code then hooks would be the perfect place to put that logic.
I agree - people saying that kind of stuff I just can’t take seriously. Programmers know to use revision control systems, practically from the start of their being programmers. For the same reason: because you can break your own code. Not using a vcs when letting ai make changes just doesn’t make sense at all. And them complaining ai broke their code means they broke their code.
Yes, and I find it mad when this stuff is used without source control.
Claude code, for example, sometimes is incredibly useful in generating repeatable patterns, and sometimes it just walks in circles. But at least I always understand what's it's doing, can revert and tell it to do differently.
Nearly everything about the experience of a terminal is better than iTerm though? I turned off all of the AI stuff since I can just use Claude directly for one off things. But it's very nice having the terminal treated like a text file instead of terminals pretending the mouse doesn't exist and there have been no advancements in text editing GUI in decades. It has issues with ssh ing into one machine but I just use command ssh for that single machine.
That’s literally how I use it too. I like file like editing in the terminal and that’s all. Maybe there’s something I can download for iTerm instead though.
I actually had good results with the free plan but with strong edit limitations (which I explicitly prompt for). Its the same underlying models as cursor (claude 4 is by far the best) and I find it a little easier to use because it takes up the full screen and is less eager than cursor somehow.
I must be very old school then. I am 30 years old, but I still read man pages often, whether I am writing C code, or just use an utility. It is a choice, however. I do not care if people use LLMs for this.
I'm much older than that, and manpages have been a poor format for long time. I read manpages all the time, but I never like it.
They lack ToC, fast search, fuzzy searches, stem searches... really anything we've known to do with searches for the last 30 years. We're largely restricted to linear grepping and scanning.
GNU tried to make `info` happen, but it never caught on.
Well, regexes, even with optimizations, are inherently linear searches, unlike a precomputed index (e.g., tries).
---
But the real problem with regexes is they force you mentally evaluate and skip over irrelevant hits, because it's not the right tool for excluding them.
Since regexes don't understand doc structure, you can't prioritize for a section heading with a word over other instances of the word.
Likewise for fuzzy search for longer phrases or cases where you're not sure of the precise order of multiple words.
They also lack stemming, so you either spend time constructing flexible regexes to handle related variants, or you run multiple searches with similar words.
Etc etc etc.
This is why text search engines are much more sophisticated than plain regexes.
---
Here's an example I ran into regularly when I was building a copy of bash's bind command:
How many irrelevant hits do you have to skip over to look up the the CLI options for bash's bind builtin? There's hits for the letters "bind" all over the manpage. And unfortunately the builtin options are near the end. I finally hit upon using "G" in less to jump to the end of the manpage, and then doing a reverse search for the hyper-specific phrase "BUILTINS", which takes me close to the right spot, but I only realized that after doing it a few times.
I think info probably would have been more popular if the local reader used HTML rather than info directly. Just the reader UI seems to turn people off to it.
Also man has "apropos" for search. It works pretty well IME.
Quick! What's the CLI command to use ffmpeg to cut out the first 20s of a video, drop non-English subtitles, and convert to the AV1 codec while keeping the VBR under 1 MBs?!
While you laboriously hunt for all the correct options, I've gone to lunch.
I doubt anyone's using AI for "ls" unless they're just starting to learn the terminal.
It overwrites files without confirmation, needlessly re-encodes audio, drops all subtitles, and sets the average bitrate instead of the maximum. So with your example, someone else is doing damage control while you're enjoying your lunch.
I'm kind of surprised it got this many things wrong to be honest.
I have no idea what prompt you fed it, but at least personally, I rarely get such poor results for ffmpeg.
But assuming I got this back, instead of running it blind, I wold check the options it suggested, find the flaws, fix them and then run it. Quite possibly still faster than it would take to construct by hand from scratch.
In that sense, LLMs are better search engines than existing search engines in terms of guiding you to the right neighborhood, if not the correct house.
---
When people criticize YOLO vibe coding, I think they're forgetting all the ways LLMs can still be very useful with a human in the loop.
If you think AI is useful for that, then why not stick it everywhere, including the terminal?
My only objection to AI in the terminal is it's one more subscription to track. But otherwise, I have no issue with AI chats in everything.
I mean, Warp is still a terminal at the end of the day. Nothing's forcing me to use AI, it's just handy for the occasional command. I use it sporadically enough that I never even leave the free tier.
I think OP made the mistake of trying to use the Warp model for coding; I wouldn't let anything less than a frontier model touch code.
Everything reinforces my choice of Aider+Openrouter+code. I use models 20x cheaper than Sonnet, completely pay as you go, nobody is playing games with models, their LLM bills and my context to turn a profit. I generate Python scripts with Mongo and Cloudwatch queries instead of getting angry at the mystical sparkly button UX on their apps.
I would really like to learn more about your setup. Can you share details?
I paid for 5$ of credits from OpenRouter and looked up their hosted models. I installed Aider and run it in watch mode in my project. I then create comments with instructions at various files (like #refer this ai) and add ! to last command to execute it. I have disabled autocommit but I commit frequently manually to recover when it acts dumb and overwrites stuff.
https://aider.chat/docs/usage/watch.html I'm trying https://openrouter.ai/qwen/qwen3-235b-a22b-2507
"AI started actively breaking working code" is vibe coding talk, you should monitor what it does and undo at the step that breaks. I've been able to do that just by prompting "undo last step/<specific action>"
I used it before as a free Claude Code to do ad-hoc scripting, pretty useful. Now I have found a bunch of TUI programs (Amazon Q, rovo dev, opencode) that can assist with that kind of workflow. I don't like that it's a GUI app, but I like the fact that it's a terminal app where I can type command directly into, not just prompt. Claude 4 should be fine, was on free, never hit limit so not sure what's the lite experience is.
I've found Claude to be terrible at undoing.
It feels analogous to what would happen if you put me in front of a broken project without source control that I've never seen before and asked me to fix it without giving me enough time to actually understand it. It starts from errors and bugs, guesses corresponding source code, and tries to narrowly fix just that snippet. Generally it favors deleting, but not specifically deleting new code.
I would have thought it could record a log of its actions and use that log to think about undoing. I would also think it could annotate lines with git blame so it knows undoing wouldn't involve changing anything more than say a day old. Unfortunately that isn't consistent with what I've seen.
I just make a WIP git commit and run git commit -A --amend --no-edit after manually reviewing each unit of work.
Edit: I also wish Claude implemented undo at a higher level instead of relying on the model. Some combination of snapshotting the whole repo and tracking operations that have precise inverses. But I understand that would have drawbacks.
maybe a system prompt to tell it to do checkpoints. Stash is one way, another is jujutsu with git backend. If this was Claude Code then hooks would be the perfect place to put that logic.
I agree - people saying that kind of stuff I just can’t take seriously. Programmers know to use revision control systems, practically from the start of their being programmers. For the same reason: because you can break your own code. Not using a vcs when letting ai make changes just doesn’t make sense at all. And them complaining ai broke their code means they broke their code.
I find it mad people use this stuff outside of a vm or virtualisation.
Yes, and I find it mad when this stuff is used without source control.
Claude code, for example, sometimes is incredibly useful in generating repeatable patterns, and sometimes it just walks in circles. But at least I always understand what's it's doing, can revert and tell it to do differently.
I installed this, lasted about 90 seconds on my machine before I went back to iTerm.
Nearly everything about the experience of a terminal is better than iTerm though? I turned off all of the AI stuff since I can just use Claude directly for one off things. But it's very nice having the terminal treated like a text file instead of terminals pretending the mouse doesn't exist and there have been no advancements in text editing GUI in decades. It has issues with ssh ing into one machine but I just use command ssh for that single machine.
That’s literally how I use it too. I like file like editing in the terminal and that’s all. Maybe there’s something I can download for iTerm instead though.
Yep. Exactly as described, pure hype.
I actually had good results with the free plan but with strong edit limitations (which I explicitly prompt for). Its the same underlying models as cursor (claude 4 is by far the best) and I find it a little easier to use because it takes up the full screen and is less eager than cursor somehow.
Use wezterm[1].
[1]: https://wezterm.org/
I switched from Warp to Ghostty and was never happier in my life.
Yeah pretty happy with Ghostty and might even switch back to Terminal in MacOS Tahoe since they finally updated it.
why... would anyone want... AI in a terminal?
For the same reason someone want AI in their text editor and search results.
So you can delete it later?
Ah yes. Like the transparent peel on a new hardware. Like systemd-resolved.
It's faster than poring thru bad manpages for commands you don't remember.
I must be very old school then. I am 30 years old, but I still read man pages often, whether I am writing C code, or just use an utility. It is a choice, however. I do not care if people use LLMs for this.
I'm much older than that, and manpages have been a poor format for long time. I read manpages all the time, but I never like it.
They lack ToC, fast search, fuzzy searches, stem searches... really anything we've known to do with searches for the last 30 years. We're largely restricted to linear grepping and scanning.
GNU tried to make `info` happen, but it never caught on.
Manual pages do have "fast search" (with regexp supported), so I am not quite sure what you mean.
Well, regexes, even with optimizations, are inherently linear searches, unlike a precomputed index (e.g., tries).
---
But the real problem with regexes is they force you mentally evaluate and skip over irrelevant hits, because it's not the right tool for excluding them.
Since regexes don't understand doc structure, you can't prioritize for a section heading with a word over other instances of the word.
Likewise for fuzzy search for longer phrases or cases where you're not sure of the precise order of multiple words.
They also lack stemming, so you either spend time constructing flexible regexes to handle related variants, or you run multiple searches with similar words.
Etc etc etc.
This is why text search engines are much more sophisticated than plain regexes.
---
Here's an example I ran into regularly when I was building a copy of bash's bind command:
How many irrelevant hits do you have to skip over to look up the the CLI options for bash's bind builtin? There's hits for the letters "bind" all over the manpage. And unfortunately the builtin options are near the end. I finally hit upon using "G" in less to jump to the end of the manpage, and then doing a reverse search for the hyper-specific phrase "BUILTINS", which takes me close to the right spot, but I only realized that after doing it a few times.
I think info probably would have been more popular if the local reader used HTML rather than info directly. Just the reader UI seems to turn people off to it.
Also man has "apropos" for search. It works pretty well IME.
Agreed on `info`. Such a missed opportunity.
`apropos` is for listing possibly relevant manpages, but I'm thinking about searching within a manpage. Particularly a long one like bash's.
typing / in the pager doesn't work for you?
Not always. See my sibling comment for details, and an example of when a smarter search strategy would be better.
Sure it's way easier to type "show me the list of files in current directory" and wait for LLM to do it's magic than it is to type "ls".
Quick! What's the CLI command to use ffmpeg to cut out the first 20s of a video, drop non-English subtitles, and convert to the AV1 codec while keeping the VBR under 1 MBs?!
While you laboriously hunt for all the correct options, I've gone to lunch.
I doubt anyone's using AI for "ls" unless they're just starting to learn the terminal.
Then you get this:
It overwrites files without confirmation, needlessly re-encodes audio, drops all subtitles, and sets the average bitrate instead of the maximum. So with your example, someone else is doing damage control while you're enjoying your lunch.I'm kind of surprised it got this many things wrong to be honest.
Kudos for taking this seriously.
I have no idea what prompt you fed it, but at least personally, I rarely get such poor results for ffmpeg.
But assuming I got this back, instead of running it blind, I wold check the options it suggested, find the flaws, fix them and then run it. Quite possibly still faster than it would take to construct by hand from scratch.
In that sense, LLMs are better search engines than existing search engines in terms of guiding you to the right neighborhood, if not the correct house.
---
When people criticize YOLO vibe coding, I think they're forgetting all the ways LLMs can still be very useful with a human in the loop.
I need LLM for that, sure.
But again, my question stands: why would I need an AI in my terminal?
If you think AI is useful for that, then why not stick it everywhere, including the terminal?
My only objection to AI in the terminal is it's one more subscription to track. But otherwise, I have no issue with AI chats in everything.
I mean, Warp is still a terminal at the end of the day. Nothing's forcing me to use AI, it's just handy for the occasional command. I use it sporadically enough that I never even leave the free tier.
I think OP made the mistake of trying to use the Warp model for coding; I wouldn't let anything less than a frontier model touch code.
Search and LLMs for that are truly great indeed.
But that's not exactly what i was asking about lol
the same people open a bare terminal and immediately jump into Claude Code
Use Ghostty