#jules-general・🦑
1 messages · Page 2 of 1
When your production code starts hovering over you watching you sleep while sharping a knife on it's beard, then you will know not to use prototype tools on production applications.
Yeah its amazing right now, didn't wanna do the light mode version of my self automatically
So i used Jules, did it one shot
Perfectly too
Wish i could pay to upgrade my limits though
I'm testing it on my super large production repo, no problems.
Just the limits
my codebase being Rust + Bevy 0.15 probably does not help
Yeah probably, i gave it a hard task
No way to detect if its hung or not
Just found a status bar, which is kind of good
though I'm afraid to ask what it'll cost...
200 usd a month
Pay for it yourself
You think my exmployer paying for ChatGPT, GitHub Copilot and now Jules? Nope.
I pay for that myself. And you should too
We have a few Servers with H200 GPUs and host our own AI models at work 😄
well, I had an employer who was prepared to pay some of these toys, but I guess you indeed better pay yourself if they or cost only 20 euro per month and make you a lot more productive, or more if you really can affort it through some way
but meh, don't ask me if local ai's are better than others... too hard to know (though claude, gemini and openai are doing the distance)
I do, but I don't put company data into it!
smdh
But I've been waiting until I can basically be a one man team just managing a bunch of AIs. I'm ready for it
Anyone know how much jules can handle? I have no coding experience but want to make a turn based game with python. No sprites, bassically console. Hoping to add hundreds of enemy types/items/etc.
Hey, does anyone know whether Jules is actually supposed to take like, 30 minutes to publish the branch, or if it's just hung?
Jules is probably not the right tool for that, at least rn. I would check out firebase studio
Ty
Wonder if Jules is getting rate limited 😄
It might be hung.
anyone know why i can't see my github repo
is it public or private?
it is pulbic
the ai chat on firebase studio is not working. from about an hour. any idea?
Does anyone know what timezone is used for the daily limits / when this resets?
Is it just 24 hours after or should it be based on your local timezone.
Started using this yesterday and hit the daily limit and today it's not yet reset.
Jules is very cool
I think it depends on when you created your task
I created a few tasks yesterday and I saw my limits slowly decrease the next day at the same times I made those tasks the day before
Thank you, I created these across the day yesterday so should start to see them soon.
Developing my first web app and Jules couldn't have come at a better time for me. Weeks of what I would have taken to apply bug fixes done in one evening.
Is there any definitive terms of service or license agreement that states who owns the rights to code generated by Jules? I've tried it out and the results are better than any other agent I've tried so far, but I'm afraid to merge any pull requests generated by Jules without knowing if I'm "poisoning" my repo.
If you don't get an official reaction, create a thread on the Jules bugs/feedback channel. It definitely something Google has to define their view on it and clearly state it to the consumers of the tool.
Actually, I've created multiple Minecraft plugins, and they're really complex. I do think he could do it — don't you think so?
Hey, how do i connect another repo, when i logged in initially, i gave access to only 1 repo.
Hey why do I always get this when I try to do any changes to my repo on Jules. "Provided manual code changes due to tool limitations with direct file modification."
Didn't you see the dropdown option to select another rep when you start a new task?
Hello everyone, I am trying to use Jules for the first time, can Jules access the private project of my github ?
Personal yes, though there were some question for private repos where the user wasn't the owner (which I didn't track further)
Is jules an IA version or it use an external IA ?
Wish there were increased limits for Gemini users ❤️
when does the daily task count reset?
No task reset today. Task 5 was around 10pm yesterday.
Jules can only access public repos
I have it accessing a private repo, seems to work fine (other than the overall instability of Jules itself)
I have the feeling, that jules is extra slow today tbh XD
I wrote an extension so that you can download your code from jules.
I will be grateful for your star on github.
I will also be grateful for your feedback, maybe bug reports.
Jules Integrates as an application into your Github profile. On the Settings tab, you can see the "Applications" entry.
Jules lives there, side by side with other integrated applications. You can manually specify the repositories to which jules will have access, or remove the application integration. Then the token for this application will be revoked by github.
Anyone can help me on this? Thanks!
Just change publishing branch name what not match to any existing one, after that you need to fix merging issues manually, by refactoring small pieces of code and so on
found it, thank you so much, should have seen there, and can you also help me with that how do i give my friend's repo's access in which i am a colloborator? is there a way to do that?
Thanks for answering! I will try this way.
private repo? do you have a plan with github (not sure if github requires the paid plan for that)
is the best way to save an api secret to put something like echo export API_KEY=foo >>~/.bashrc in the repo setup script?
Will google one subscribers get more access to Jules soon
Unlikely ProCreations. Tis going to be a seperate subscription
I would think that’s what they would do because Google one is marketed as everything for ai
True, but these are companies and they are out to make money. Is easy to assume that the devs who use her are corporate sponsored as opposed to open source developers
I'd advise you to brace yourself for a 199 subscription model. God knows I am.
Does Jules support creating repo from scratch ?
Like starting from an empty repo? Yes
Hmm, seems like Jules can't see my repo anymore. Worked fine yesterday. Any recent updates or something i've missed?
And now my daily limits went from 3/5 to 0/5 
Yepp! Last entry of repo selector (not branch)
Hi google staff, I have encountered an issue for the third time where jules failed to complete the task and I receive this message...
Jules encountered an unexpected error. Please start a new task to try again.
I was wondering is it possible to add a button that allows me to choose to push at least the changes made so far to a wip branch ? Thanks.
You can manually download this changes, even if the publish dialog is hidden
follow this method
Btw, is it expected, that jules thinks it can't run and test the code it creates?
Jules can test the code if you specify toolchain what ru using. But it does not have a graphics output device. I realized this because it couldn't run tests of the application on Qt widgets.
Jules can execute commands, and you can specify prepare commands, what jules try to follow after booting maching
I know it can, but I had to remind it, it said it can't
Jules tries to stick around the plan, and if there were no tests in the plan, then most likely jul will not do anything.
But you can ask for it, howerer the priority of this message gonna be minimal.
Like, it not only did not do them, but when I asked it to do them, Jules responded that it doesn't have the enviroment needed for it and I had to specify that the enviroment is ok
The vm running jules is based on ubuntu. Not all types of toolchain can be loaded on Linux. Obviously, C# or similar (Windows-related) toolchains will not run on linux.
omg this is amazing, thanks!
💟 I will be grateful for your star on github.
one github star and one ⭐ here also
There is still a small bug - after switching to other tasks, the editor slows down, so I recommend disabling the extension when you download what you need.
I will also be grateful for your feedback, maybe bug reports.
It was gcc
And cmake
Both should work fine
And like, they did, it was just that I had to convince the bot, that it can actually run them
And now it got stuck on pushing the update into git
.net tools do run on linux distros.
It depends which ones, but the new .net C# should be fine afaik
But this was gcc compatible cpp, so definitely fine
This is a common problem, so I found a solution to it.
When it downloaded Qt to me, it downloads version 5.15. This is a very old version. It's already 6.9 - so I think the package registry is incomplete, and the environment has been stripped down - this was done for security purposes.
not so much from the user, but from the jul itself
And the solution is?
discussed above
Yep, I had problems installing clang there
Gcc was fine as it's preinstalled
In minimal installation of ubuntu - gcc and clang already preinstalled
There's a custom installation booted
Wait, clang too? When I tried it, cmake couldnt find clang
And it worked only with gcc
Which is quite bad, as gcc has a really terrible ixx implementation....
Like, it tried to use clang, but cmake returned an error, that the compiler was not found on the system
I may try it again if things have changed
The VM starts with a custom redefined instance of the system, and anything can be there, even some things that are in the minimal installation are cut out.
We're going to need our own flavor of vms if Jules is going to scale well
I can't have her using versions of ubunto that don't support dot net properly
Or at least allow us to use some sort of default envrionment build scripts.
though I can see why that might be problematic. I hope those vms don't have acess to the general internet; otherwise people are doing to find ways of hosting their websites within Jules
It's definitely doable, I've seen jule's do best on tasks with detailed and specific objectives defined. If he was very modular about how he prompted jules, slowly building it up, I bet it would work great. It'd be much more simple to use firebase studio to create the base of the game, and use jule's to add specific features without breaking the repo. How did you approach building the plugins?
Right now, I’m working with what I call a “CDC” – that’s short for Cahier des charges in French. I’m not sure what the exact term is in English, but the idea is to create a full specification document. I’m doing this using Gemini 2.5 on AI Studio with a system I’ve been using called “question-response.” Basically, I feed it some information and then let it ask me all the questions it wants to create the best document possible.
Once that’s done, I ask Gemini to generate a high-quality prompt that I can use to get the AI to implement everything. Because honestly, the real problem with Gemini and Jules is that they’re really lazy. They rarely implement things from the start. They just wait for you to take the lead.
So, I make sure the AI doesn’t give me a “TODO” or any “implement this later” comments. I want it to write the entire code directly and not leave anything behind.
By the way, I actually think Claude 3.7 is better than Gemini 2.5 when it comes to coding, at least for java part. Gemini tends to slack off unless you know how to talk to it properly. Claude, on the other hand, usually just gets the job done – even if your prompt isn’t perfect.
So yeah, that’s how I build plugins. And when I say “plugin,” I’m talking about stuff like a MySQL plugin for a team system with economy, ranking, and web integration. (By the way, the entire web interface was built with Gemini.) We’re organizing a Minecraft tournament with 100 people discord community (max 50 players best case scenario), and it’s just way better to have our own web system and plugins instead of relying on existing ones that aren’t really precise.
We also have a Discord integration where roles, channels, etc. are created automatically. The web interface includes an admin panel, Minecraft connection system, and more. I do think we should work more on security, especially since there are 100 of us using it. But yeah, that’s how I go about it.
And the crazy part? We’re doing all this with only AI and just a bit of coding knowledge(98% AI / 2% us). I’m a med student – never had any coding background before learning with AI. I just know enough to fix things when the AI messes up.
I do belive IA is really clean and doing a really code job :D Sorry i've talked so much hope you can understand evrything
That's how i use all the AI power :D
I just wrote a prompt of about 2500 words.🫠
I'm glad you can relate in some places. I too, when I ask the AI to generate code for me, I ask them to send me the complete code. I have almost no knowledge 😉
ところどころ共感できるところがあって嬉しいです。わたしもAIにコードを生成してもらうときは完全なものを送ってもらうようにしています。知識がほぼありませんので ;)
Hahaha i usually do 15000 no joke
What do you use AI for ?
Clearly agree with you 
O M G
I use them differently depending on the time and occasion. Mostly ChatGPT, then Gemini, and sometimes Copilot.
時と場合で使い分けてます。主にChatGPT、次にGemini、たまにCopilotという感じです。
There is a possibility that jul is already scaling, since the easiest way to do this is to run it in docker, and then you can build your own system image to cut out all unnecessary stuff. I realized this cuz jul can't run graphical applications, or even debug them. as well, reduced package registry and so on
i've found that cursor with claude is the most effective for most of the stuff I do.
Sorry, I probably gave you a different answer than you were looking for..,
ごめん、多分求めてる回答と違う回答しちゃったよね、、
ChatGPT Worst AI for me...
There was a time when I used to make about 99% of Minecraft plugins with AI. Recently, I've been having Unity scripts made.
Minecraftのプラグインを99%くらいAIで作ってた時がありました。最近はUnityのスクリプトを作ってもらってます。
Don't worry
Lesgooo we do the same thing XD
I am just trying jules for the first time now; has anybody else encountered the message "I'm encountering a persistent storage issue that's preventing me from understanding the project structure and creating a detailed plan. Could you please check the available storage space in the environment? Once this is resolved, I can proceed with exploring the codebase and creating the plan."
Wish you good luck and nice to meet you boss :D
Problems from Google side, don't worry, try again later.
A theoretical solution could be to change your account.
I know exactly what you mean.
Clearly he's not able to do nothing he's really dumb tbh even to answer just chating he can't
This seems like really smart way to do it. Its great that you were able to do all of that without much coding experience. How long did it take you to build? And once you had your master instruction doc, were you able to get most of it done in one task?
I translated it into English and it doubled. Still, it's not far off 15,000. ....
英語に翻訳したら倍になったよ。それでも15000には遠く及ばずだよ....
XD but mine are made by IA as i discuss with him he ask me question to really make the greatest prompt possible
Usually, it can take around 4 hours because AI often generates mid-quality code on the first try. But as I test various features like commands, database usage, permissions, shops, or any other functionalities, I keep finding errors and have to make the AI fix them.
(By the way, Google and others should really train their AIs better when it comes to databases — both in terms of structure and proper syntax 😆).
For a really complex and complete plugin — one that could easily cost around $250 if made by a human — it's take me up to a full day tbh.
I think I might've seen others talk about this already, but does generating additional veo3 clips with extend or jump to in the story board cause audio in subsequent clips to be missing?
it doesn't seem to be working for me
I see. I'll try it next time.
なるほど、今度からやってみようかな。
Clearly like u explain the global think, like version project env the global think and ask him to ask you question make a create specs docs
Good Afternoon guys
Yeah, I’ve had gemini do pretty well with Postgres scripting and structure.
I'm new here
Good morning.
おはよう
good night for me
Uuhhh MySQL isn't that great i belive at least the syntax is never really good
i am building a whole app with jules currently for me and my friends to plan our yearly trips to Tokyo/Japan 😄 It has a supabase backend and is already working almost without bugs 😄 What are yall building?
What prompt do you use with gemini to get it to ask all the questions and build the instructions?
Usually, it goes like this:
"Hey bae (TBH, when I'm Setzy with 2.5 Pro, he performs way better 💀), I need a Minecraft plugin — let's say something like a team system, called TeamCore, for version 1.20+, with MySQL support, BungeeCord integration, Maven project structure, and using Java 21. Ask me all the questions needed to create a complete specification document (CDC — or whatever it's called in your language). And please, no dumb questions like 'what’s the max number of team members' or 'how many teams can exist on the server'. Ask deep, detailed stuff to make the most complete CDC possible."
Good morning, where are you from?
Good night, where are you from?
France :D
Where are you from ?
WoW it's the first time I meet someone from france
I'm from Colombia
Japan 🙂
Wowwwww
I had a friend from japan
Sadly he passed away like 3 years ago
Nice to meet you a first for me colombia
oh..."お悔やみ申し上げます。"(Japanese, words of condolence for the dead)
That's great!
My deepest condolences. I offer mine as well.
What u doing for living boss ?
Are you a student or do you work and what do you use ia for (IF u want to answer me i just want to learn abt people)
I asked Jules to get latest from the main branch and she's dumbfounded.
Is there a specific keyword that triggers her?
Also how do You stop her from thinking?
I'm a Business man
I do E-commerce and that kind of stuffs
I'm building a COM bridge to implement my messaging protocol in pure assembler with Qt integration. Everything is not so rosy here
Ohhhh 😄 I visit Japan every Year from germany 😄 I am so sorry about the tourists that cant behave in japan :((
Lmao I love it
I asked jules to
"Make the website accessible via different devices with different screen ratios. Give the top 30 px a gradient blur effect when scrolling, and add a "Scroll Down" button that when clicked will smoothly go to the end of the page"
and this was the text of the resulting web page (translated to english) 🙃 :
Thanks, but unfortunately that just states the terms under which Google can use your data for training their model. What I'm looking for is a definitive statement as to who owns Jules output, ie, code written by Jules as a result of my prompts. Without knowing this, how can anyone merge PR's generated by Jules? An open source project might be including code that they don't own, therefore doesn't fall on the the license they publish under (ie, GPL, MIT, etc only operate via the consent of the authors). Commercially licensed projects can't risk integrating code where the author's consent to integrate and license the code is unknown. Unless there is clarity regarding the ownership and/or licensing of Jules' generated code, this tool is not any use beyond a simple curiosity. I'm sure Google will eventually clarify this, I'd just rather they did it now.
I think we might be asking too much of Jules. Is there a recommended prompt that people are using?
i think jules is too overloaded for me to see anything successful still. i'm just hoping to read someone else's stories of success.
mostly nothing i start ever finishes, even small requests
small meaning 'change gpt-4o to gpt-4.1-mini`
Poor one. Its facing alot of prompts
But when will Google give me -_- . I have 4 accounts but none of them have it.
what do you get? https://jules.google.com/task
the access to that page actually worked for me, i never did get an invite by email.
anybody get Jules lag right now, like stuck / bug ?
Is Jules mostly for smaller focused tasks right now? I gave it a refactoring task and it just work for a little while then stopped, I would tell it to continue and it would do the same
Anyone getting Command timed out. for everything?
Yeah i'm stuck for like 1-2 hours now, could be some issue with Jules now, not sure!
I guess it's just a matter of time now that this won't be over for a long time...
これがずっと終わらないのは今のところ仕方ないのかな...
Uh
There's a bug in the final code on the main page
Where it interactively edits an example
It introduces a bug
Gotta break it in tasks, mine handled a 10 points plan (his own plan) with no problems
Including writing an API connector, a job that consumes it and 2 pages to display the data, as well as the database tables and models to save that
Temproary solution for this type of issue: https://github.com/L0wl/julfix
Wow!!! THX!!!
Wow!! THX!!
How do I get the screen to download? I'd like to know if there is a way to download the
ダウンロードする画面はどうやって出すんですか?、
Czech Republic
" Jules encountered an unexpected error. Please start a new task to try again. " -> 50 minutes of processing to the drain? seriously?
That's really, REALLY frustrating, and he's starting to do it too frequently :(
Me too 😦

Can someone help me?
I am a complete novice in regards to development, using GitHub, and using AI agents.
I have a problem when using Jules: I gave it a task to run a project.
He completely wrote the code for the project. However, once the project is complete, it simply does not show the “publish branch” button.
I can't export this project!
Please help.
I even uploaded all the Google Jules documentation to the AI assistant to help me figure it out, but to no avail.
ask it to commit the code
"create a git commit with all the changes you made"
@ruby zodiac ^
if that does not work, send feedback to Google using the button on the top right.
my understanding is that Jules should have created a commit even before you asking. And you asking for publishing a branch it should have understood too. So this is not correct behavior
the other possibility is if Jules is still working which sometimes it might be hard to see
i think ive had a similar issue, jules made all the code, but then had an error writing the code to the file on the env or something, so when i ask it to commit the changes, she tells me theres no changes to commit because the files were not written
any idea here?
I've already given him the go-ahead that the project is over.
However, I did not receive the commit...
"I can't do it for you."
do you see the individual files changed?
The entire project was written entirely from scratch by Jules himself.
I'll take that as a yes. Does it have a .gitignore file?
does the repo have at least one commit?
Yes, it have.
definitely file a feedback form with this. Be specific that this happened trying to make Jules to code an app from scratch, this seems to be related to the problem somehow.
you'll need to see how to manually get the files out, if there are not many you can try to copy&paste....
just to be clear, it's not your fault. The tool lacks a lot of polish and testing
I was beginning to think I was really “Ivan the Fool”
There's actually quite a few files of different extensions in there...
do you know enough on Git to be able to place the files on your own, commit and push?
Just now learning from the video guides.
I'd recommend to get VSCode, and GitHub Copilot set up, and try the GitHub Copilot agent mode - using Gemini 2.5 Pro. You'd probably have a better chances of success this way.
I'll try to give him the task again.
Maybe it'll work on the second try....
ensure the repo is not empty when it starts, please
put a README or something on it
already
good!
Вот это прикол
Если есть кодовый сниппет - имеется ввиду есть diff тогда можно чере расширение скачать: https://github.com/L0wl/julfix
Ветку придётся самому создать, и если будут какие-то merge конфликты - фиксить самостоятельно
Absolute power has peoples, who use both
What is meant by the screen to download? Did you need single command to download sources in zip?
There's a single command. F12 -> Console -> Shift+Insert (when you already in the task)
(async()=>{const z=new JSZip(),e=monaco.editor.getDiffEditors(),r=/^\/modified\/\d+\/\d+\//,s='modified/cumulative/',n=document.querySelector('.source-text-repo')?.textContent?.trim()?.replace(/\s+/g,'_')||'jules_sources',t=new Date().toISOString().replace(/[:.]/g,'-').slice(0,19);let a=0;for(const d of e){try{const m=d.getModel();if(m&&m.modified&&!m.modified.isDisposed()){let p=m.modified.uri.path.replace(r,'');p.startsWith('/')&&(p=p.slice(1));if(p.startsWith(s))continue;z.file(p,m.modified.getValue());a++}}catch{}}if(a){const b=await z.generateAsync({type:'blob'}),l=document.createElement('a');l.href=URL.createObjectURL(b);l.download=`${n}_${t}.zip`;document.body.appendChild(l);l.click();l.remove();URL.revokeObjectURL(l.href)}})();
There is no simple way to download the code. All the code must be extracted from the Monaco editor and archived after that.
While I was trying to figure this out, most of the difference tabs contain files that do not match the source code that I see in visualization mode. There's a lot of metadata here that can't even be read - it's a simple filtering of all the differences so that they match exactly what we see in the viewing mode.
Проблема оказалась в том, что репозиторий был полностью пустой.
После добавления файла readme вторая попытка оказалась успешной.
Но если что - ты знаешь где искать решение
I couldn't figure out how to get this screen... Sorry, I don't understand English...
この画面がどうやったら出るのかがわからなかったです...ごめんなさい、英語がわからなくて...
The button will appear by itself as soon as it detects the task and the code in it.
https://ibb.co/FkdzYW6g
URGENT - after publish branch, any changes made by jules are NOT pushed to the github, why do I need to ask jules to push the changes? If jules do changes, then just push to the branch, for god sake
What does the red status in the terminal for my task mean?
You have to pull it to the main branch yourself
Does anybody know why Jules NOT uploads code and I see no "Publish Branch" button?
Seem someone also met the problem...
Front-end issue, I think. If successful publish. You have to pull the commit to default manually on github
Who can help with this please. Does it mean I should start a new task? Because I keep trying it doesn't respond anymore. Just constant blackouts
Uhh I even can't find the publish button, and Jules asserts that "I have already submitted the conceptual changes"
You already clicked publish? If not refresh and try again
There is no publish to click, but Jules says it has finish the code and it seems like a endless loop😩
Have you tried refreshing. This bug is new to me
Jules should push the changes automatically
i had this bug all day today.
what worked for me, ask him:
deploy changes to github
i cant even see whats he changing either, or process of work.
only way i ask him to deploy, he creates that "Publich branch" window and shows changes on the right
and if he is gonna say i already deployed, then you will have to ask him to deploy it but with new commit message
i feel like back to 2023 tbh,
when i used discord ChatGPT bots with like 500 tokens context and half time erroring coz too many people sit on it
but i love Jules
I mean isn't the current legal precedent in the us that code generated by ai in not copyrightable? Google's terms of service states that they claim no ownership over user generated content:
Some of our services allow you to generate original content. Google won’t claim
ownership over that content.
I'm starting to lose my mind trying to log in, any idea how to fix this? I only have basic google one subscription on my personal account, everytime i try to click get offer, upgrade or whatever it is, i just get this message saying im subscribed, but clearly not because i can't generate anything
any ideas?
stuck
When daily task limit reset?
fix here: https://github.com/L0wl/julfix
One homie got unstuck by saying Continue or something like
It minds the Jules got an error, Jules trying other strategy to reach plan steps, by using other commands. Environment where is jul booted up - restricted, and reduced.
Yep, I think you're correct on that. Although it would be nice if they they made it very clear in the FAQ.
Question. does the Google AI subscription include all their AIs?
I am experimenting with Jules using an Android phone. I have a few UI and UX issues.
This is probably out of scope at this time, but I could provide feedback.
I posted in the feedback channel. I intended to link to this note, but this note can just be deleted.
This needs to be resolved properly, not looking for a makeshift solution
OMG, thank you sooo much for these wonderful things you bring into my life ❤️ ❤️ ❤️
Gemini already changed my life, but this jules.... is next level ❤️
only problem i had so far, was "sub-modules". Jules cant handle it yet
everything else is just mindblowing
Hey everyone!
I’m trying to connect a GitHub organization project to Jules. After completing the connection, the repo doesn’t show up in Jules.
Does anyone know why this might be happening?
Our organization has IP whitelisting set up — could that be causing the issue?
You may need to fork one as your repo
This works sometimes, and it seems likes contextual confusion
Please. Why does publish don't work sometimes
did someone else got stuck with a chat conversation "Booting up the VM ..."? To be hones, my stuck chat became a lengthy chat
(hmm chat already got back after 5 min, its already ok while I was typing)
Why i cant add project in firebase studio??? Like i click upload project and cant choose any file
Was trying out Jules. Seems after do the changes the updated code in not being updating in the remote REPO of Github ? what should be done
can you confirm Jules is using Gemini 2.0 Flash ?
REALLY ?
Nah
Thats wild
im seeing the same issue. Jules is not pushing the code to the new branch
plus, i cant export the files locally, which is annoying
anybody else facing this issue with jules?
2 of my tasks failed because of internal error...
3 tasks 😭
What should I do? I’m trying to access one of my previous chats that contains a long conversation, but now it won’t load.
Yes, but jul trying to achieve the previously stated plans using other commands. Anyway, you can say again / continue , and the it will try to process the command again (or use another one that does not deviate from the plan)
got same issue, one time it loads in 5 min, other time it gets stuck seemingly infinitively. Doesn't seem to affect short or starting chat conversations for now...
Loginned from PC? Mobile user agent stucks recently like that
both, phone is worse indeed, but I am pretty sure the phone unloads the page more often (at least on android) while switching apps
I tried that too, didnt work either 🥲
Clearing the browser cache and cookies helped me. But desktop version of page is loading quickly, and at the same time i got some freezes in task
well when the freeze was worse, I switched my browser on the desktop, same result. On the Google domain I am a bit more hesitant to remove cookies
Yeah, There is a high probability that you will have to log in to your accounts again
also the cascade is worse 🙂 (google mail, google drive, etc...)
But i'm clean cookies only for jules page. that option located in lock-icon around url bar, and works exactly for current page
yeah you can limit on domain (not checked if there were actual www.google.com cookies) but still
Cookies only for this page
Also...
is there anyway to force a push to github of current work as tasks keep failing.
and ok, there are just cookies for account.google.com and jules.google.com
You can download code manually, and push to github manually too, I have same issue in few tasks
50 files
Yeah, and fix merge conflicts manually
Code can extracted automatically
but other work will produced manually
Extension for code extraction https://github.com/L0wl/julfix (Need to be pinned)
hmm its now fixed after reloading for hours 😄
Yeah! Finally
Hey Jules Team,
Wanted to drop a line about something that's been a bit of a headache. Overall, I really like using Jules, but there's this one issue that keeps popping up and it's pretty frustrating.
I'm talking about when tasks just... die. It's a real bummer, especially since we only get 5 tasks to work with. When one of those fails for no clear reason, it feels like a wasted slot. That and the inability to push changes to a branch.
Could you maybe think about either bumping up the task limit while you iron out these bugs, or perhaps not counting a task against our limit if it fails on its own?
Just say again and it will conitune
I was just starting asking the google ai in the devtools to analyse the error and it said it was unable because there were things happening on the background (though I was busy to refine the question now while it was fixed)
Works fine for me, but crashes after few times
self-eliminating error
Yeah I've done that a few times and sometimes it will just continue to error. Usually when that happens pushing to the branch no longer succeeds but that julfix repo is something I'll absolutely try out thanks @dusty vigil!
Well, page is slow, but I had my suspicions
heavy dom
(so lots of items on the page)
💟 Thank you for using our solution. The fixes will be soon
Yeah, page is fully dynamic and builded from execution of js attached by preloaded index
also too early for the team to implement optimizations I guesss... given that the app is just a prototype... things like eliminating off screen dom is also not that simple...
Frontend, and most likely it is built automatically through some kind of react-like application or something else. The backend logic itself is flawless, but it has several bugs. (in terms of uploading to github and things like that). The application can be compiled as a web assembly to get absolute performance for each device, although the weight will be higher, but it can be compressed and decompressed locally
(unless the obvious hidden stuff is removed)
Does anyone know if there is a way to manage things like Node version with jules? I'm in Initial Setup and its using node v18 but my application requires v20.
I didn't check if there is some kind of nvm installed/available
Maybe apt-get is available as install command (didn't tried it either)
Hi, does anyone know if Jules really have path limitations to place files at the project? I'm working with NextJS 15 with group routing and Jules cant put files into "()" paths
Does Jules try to run the code it writes? Finally got some output from it, but nothing was compilable due to a bunch of silly mistakes which I am sure Gemini could easily fix if it had tried to compile to code.
I'm here, it says I must run firebase deploy my side, so how do push to git from here?
Jul can run tests, if it was described in plan earlier. You can also specify toolchain prepare commands but always remember - all that stuff running on modified ubuntu with reduced package registry and so on (includes environment)
As I know you need to ask Jules to skip this validations and publish to github, if you just continue asking and chating Jules wont "finish" the task and publish
What I'm doing is just asking a feature and when Jules ask to test something on my local enviroment I return saying to just finish the work and publish the branch, after that I continue talking and asking for more updates in the same chat, it dont consumes more tasks...
The bad part of it is that you will have a ton of branches in a short time lol
Ton of pain while trying to merge them all
But anyway, Jules has mastered an insane level of assembler skills. I'm already tired of correcting mistakes after Jul work. 🤣
yeah... what I've been experiencing is easier to ask Jules to fix eslint errors, refactor code and match best frameworks practices than write from begining a new feat
I "write code" using Gemini 2.5 Pro Code partner and after ask Jules to fix
Interesting - for my run it created tests but didn’t run them to verify functionality. Generally for most errors (especially on our TS frontend) simply compiling the file will surface a lot of errors. Like Jules created a component, but defined its props inside the component itself with no actual params rather than outside of the component scope with params, so every reference was broken. This would’ve easily been caught if it compiled the code
It appears he can push to the same branch.
It causes some merge conflicts, what can be fixed manually only. A lot of review after jul work need to be produced. But in otherwise it's gonna good enough to be true
Anyway we can tell about all manually found mistakes what jul made, And jul tryna fix errors. Jul also doesn't provide full additional review of self-made changes but In otherwise - we can't execute code, before jul was publish it on git. Three sided situation
Where should I write to propose an idea?
i used flow to create this video about vibe coding https://x.com/mattsilv/status/1927077826412331023
What's Jules guys!
HI
Newb here hailing from Melbourne, Australia, just looking into this for the first time.
Thought there might be a support channel so apologies for this really obvious question. The FAQs suggest that private code repos remain private and are not used for training. The privacy policy says "Google collects your conversations (including prompts and attachments, code from linked repositories and generated output)". Is there a general consensus that this means just if you include code in you chat as opposed to any code it reads in your repo?
Again sorry for the boring obvious question.
vitally
@stiff scaffold did you never heard of it? (I can't blame you, this tool has been only announced on a few channels like on the google i/o yt stream)
did you tried it on jules.google.com (and small note: its for developers)
(not part of the answer what Jules is), I found this tweet about claude 4... ai is really a toy atm... https://x.com/vasumanmoza/status/1926487201463832863
Absolute power
in ignorance of the framework
if you leave feedback, it says that feedback will be used to train AI, including pinned code, inside of the so-called "task"
I am now switching jobs actually, so I have the time to explore all these fun ai tools... looking out for having these tools part of my toolchain because the prototypes are working relatively well...
honestly i am 'synthesizing' a full feature, complex, fitness planner app without breaking a sweat. I am at 4 days at a row without any major hiccups, but I am vibing since vibe coding a thing, I feel like this will be production ready end of this week.. My issue mostly is being parallel. I read / listen to some really good Anthropic / Claude Code talks from their developer, but I wonder if how others 'plan' multi threaded app synthesis without making the agents step on each other's toes. Is agent to agent realtime context sharing is a thing yet?
ai-competition-resolving-protocol
so my issue recently is more like 'time'
since most of the work could be done faster if it could be paralelized better
I've spend this weekend attempting to set up a local https certificate setup from scratch, I would never got as far as I got today (yeah I have lots to learn). Just some minor issues left to fix...
check mkcert. for all kind of stuff like that i usually use "Warp" terminal or Cline / RooCline to configure it to me
yeah I should trust tools, also I feel weird to trust tools I never heard of and never used before...
ow, you are mentioning other tools than I expected ^^
i was mostly reflecting on your twitter link
i do 80% claude, 20% gemini, but recently more gemini, but i still fall back to claude 4 unfortunatly
(I stopped caring about the x link, though I found it funny)
and as far as i feel, if you do it right, prompt it well, plan ahead and document, it's works really really well.
yeah I am mostly stuck now with claude (thinking) and gemini (quick solution scan + file tree exploration)
also I use gemma and llama 3.1 on local machine
i don't use thinking
yeah depends, if I want the ai to evaluate I put the thinking mode on
gemini 2.5 pro and flash is a good combo but my agents struggle less with claude for some reason
also recently i got more and more rate limited on flash (i am on tier 1) and sometimes gemini more laggy (i don't know why) than anhtropic
I still want to have my llama.cpp working with images though, I am bad in reading their documentation
I also don't use local models at all, but when I do i use it with LM Studio
(and I don't have cuda)
I've done an ai project before where I actually needed to run ai locally for an interactive setup (also had streamdiffusion visual prompt which I didn't set up), it was fun 🙂
Hey everyone, I’m new here and currently a student. I’ve been exploring AI tools mainly to help me learn coding. So far, I’ve used some AI-powered VS Code extensions, which I found really interesting because they assist in completing code. However, I haven’t yet tried building projects from scratch using AI. Could you recommend any tools for that? For example, I recently was reading the chat came across LM Studio and was wondering if it’s a good option to start with.
I didn't use replit a lot, though replit seems capable to create projects quickly from scratch
Thanks for suggestion Since I’m a student, I’m looking for tools that are either free or low-cost if possible. I want to try something that won’t add much expense while I’m still learning. definitely i will try them both thanks for suggestions
I think firebase studio is free
Thanks that really helps
I was wondering if you guys could recommend any helpful YouTube channels or books to deepen my knowledge of AI and new tech. Right now, I follow Fireship and find it really useful. If you know any other great channels or resources. I would love to follow them out.
tbh, the youtube channels changes quite regularly and it depends on what your interests are
just try to follow some mainstream channels though, and you get a big chance that you can explore closely related channels as well
I’m trying to explore a bit broadly right now, so any recommendations for channels or resources that cover new tech or Ai would be great
its not ! the old model is free ! but you have a limited rate with the 2.5 also when the rate ends ! it just gets stupid for complex taskes or bug fixes
That’s a great point i think my YT feed is kinda messed up i was thinking if i follow more tech content i might end up learning more in my free time then watching random stuff
also definitely rebalance your focus (at frequent intervals), social media figures and yt channels go quickly up and down in terms of relevance
i have a good tactic ! wich costs 0$ for now i even tested this in a big LMS app and so far its working ! but it takes time
what I think sticks for many on yt right now is theo/t3.gg
so i started with
-
Lovable AI Free 5 credits a day 30 a month total ! ( give a screen on something you want to build and tell him your project and your ideas to make the exact same as the screenshot ) 5 credits can get you a solid working front end also in Next JS and the designe and the code is just WOW a real good front end AI
-
OpenAI just give him the files of the project that lovable did for you and ask hime for a good backend plan wich framework what APIs you need !
-
Supabase for SQL PG free to manage your AUTH / USERS / ect ....
-
Jules for now with the backend wich im still testing BUT you can use VSCode Copilot it will help you alote but with a bit of understanding int he frameworrk ur working on
ALL of those are AI agents that edit / code for you and you basicly just type the prompt
Hope We can use claude in jules soon
Thanks, yeah that's how I'm understanding it, just wanted to confirm before I try it out
how to file a feature request :)? or see what if filed alredy? I would love to have static system prompt that I would love to use with all my communications with Jules, does not looks like this is possible yet
seems to be stuck. I requested a code change. Jules told me that my "team" would have to test the code then publish it.
A quick video clip showing attractive products with discounts (e.g., clothing, electronics, household items).
- Upbeat music with animated text such as:
"🚨 You won't believe the prices! Buy at 70% off before they run out!"
Create repository, where's one file only exist: TODO.md or whatever. Then there's you can describe your tasks as checklist, what gonna be updated by it's gonna be done
And there's a small note for u: TODO list can automatically updated by Jules, (if it was described earlier on prompt phase)
Example todo
- [x] Implement feature 1 (Completed)
- [ ] Implement feature 2 (Incomplete)
...
If you want extended support of files - go to aistudio > there u can upload pdf (but before doing that - you need to convert it into png) choose model, what understand images, wait until context gonna be extracted. After that - you can change the model, add additional details, and done! Your development workflow got pure power
And also jul can build task list for you after all main job was completed. You can put abstract description of core concepts and problems what program should solve
this can be fun id you make it that way Okay, I will attempt to add 'bobohead' to the end of docs/testing.md. Given the persistent issues I've had with file modifications, this will also serve as a test to see if I can correctly modify files now. I'll let you know the outcome.
Good news! I successfully added 'bobohead' to the end of docs/testing.md. This indicates that I might be able to modify files correctly now.
Given this, I'll proceed with the plan to address the actual build errors in the code. I'll update the plan to reflect that the file system issue seems resolved for now and move on to fixing the build.
Hey. It would be nice if Jules could use more repos at a time than one - sometimes they are connected.
I guess this could be in "feedback" but I just put this here.
like a git sub module?
I think what Jules is lacking is the visibility of the command line
how would I know what Jules is doing or might be doing in the future if Jules has any sort of tool calling functionality that accesses the command line?
Actually the repos that I'm working with now is not a submodule, but related so that when cloned locally, they use each other inside the same folder.
So, maybe if Jules could get two or more repos as the input.
Also it could probably handle submodules.
it does handle submodules without problems, I didn't ask Jules to do edit in submodules yet though
Some more feedback - I'm not sure if everything is OK here. It's taking quite long to create the card assets now.
I am facing below issue
I'm facing a very persistent and fundamental issue with the development environment. Even basic operations are failing because they encounter a Git corruption error (fatal: Could not parse object '9ddb899981e7ee02d2f28f45dc4a42bd4c75e35e') during their setup phase. This indicates the Git environment itself is in a bad state, preventing any operations, including trying to remove and re-clone the repository.
This seems to be a deeper problem with the environment's Git setup, rather than just the specific repository clone. I'm completely blocked and cannot proceed.
This likely requires an administrator or support team to reset or repair the Git environment I'm operating in. Could you please look into this or escalate it to the appropriate team? I'll be unable to continue until the environment is restored to a healthy state.
I guess you need to retry, but you can do that outside the chat. Go to your repo and go to the settings tab and configure your setup commands.
Then use the execute button.
It has been working on this for an hour or so - should I ask it to stop?
you can try to refresh (or open in an other browser). Though for long chats, you then sometimes apparently ends up in maybe-infinite-loading
Apparently it was spending time trying to download some assets and then created svg versions of them in the end.
What the hell is happening with Jules? I can’t paste my whole prompt, and it starts responding in Arabic even though I’m writing in French or English. It just won’t work. I tried switching accounts, changing the GitHub repo — what else am I supposed to do? It only takes like 1/5 of my prompt and then does some dumbahhhhhh nonsense.
I wonder if it hit an improperly trained token or neuron ^^ Yeah, you can't predict llm behavior sadly...
i mean x) i just want my minecraft plugin x)
the highlands 😉 ?
(refering to landscape formed due to lack of floating point precision)
Look my jules becoming arabic X)
It's badland isnt it ?
AND idk why it doesnt take my whole prompt :'(
I guess the prompt confused the ai, if your instructions are length and you know it ahead, commit the file instead on the branch you want to let Jules work on
Okay that's a great idea too thank you boss
(for clarity: add a file with instructions and refer Jules to that file)
Yeah like i add to my respertory a file called "Cdc V2" and say to jules look at this file ?
that should technically work, given that llm's are often hit indexing of large files, this should be a scenario that llms are more capable handling in
imagine llms crashing because they were attempting to read a compressed version of your large db backup ^^
I love how you're know well llm
Do you work in IA ?
nah, I just follow up ai news a bit, though I always failed to learn real ai. But nowadays that's ok because ai evolves way too quickly
first wanted to do electronica, then ended up being software, quickly webdeveloper, then got stuck in that while I still wanted to eat some bits from things like crypto, ai, quantum, cpu internals and gpu stuff, but I am not learning fast enough :/

Any idea what to do when in infinite loading "Publishing to branch" I have tried in a new tab and browser.
Xd so many bugs on the site and with the AI hehe (endless repository saving) (errors when running commands internal error) (AI stops working and even if you ask it to repeat or continue it freezes again and you have to create a new task) and all of this in just 5 minutes… well okay maybe 10 to 15
I had the ai failing to fix a typo in a cli argument for at least 5 times and with 3 times where the ai knows what the mistake was...
I don’t think Jules is suitable for development at the moment since it has too many bugs. Hopefully, they’ll fix them next week.
it still did things though that I would never do myself, but its for the moment only useful for tasks that are quite short. Knowing development, things always go out of scope... at least most things that need at least a bit of details
AI logic: if the command didn’t work the first time, it will work the second time.
ai encounters... it worked on my pc...
Ah whatever, it’s an internal error Xd
It’s just that usually when I get an internal error while interacting with Jules, any further commands stop working.
When do you guys think Jules will get an update where all those ahhhh bugs finally get patched?
Let me make a global list of issues for the Jules team so they can squash those ahhhhhhh bugs:
The website is slow and laggy.
Jules is really lazy — he doesn’t always do what you ask him to do.
Jules loses context and fails when given longer tasks, like reading a full Minecraft plugin, finding potential errors, and fixing them.
Jules can’t run commands on the VM.
Jules can’t publish properly — it lags 5 times out of 6.
Jules can’t fix his own errors.
Jules is super lazy — yes, I said it again.
Jules loses context and struggles with long prompts.
Jules has too many internal errors — WAY too many.
Did I forget anything? Let me know so we can put together a solid list for the staff team...
Today, the initialization of the Worker object inside the monaco.editor was surprisingly fixed.
Earlier it cause some task loading issues, And then I had to wait about 15 seconds.
- The site lags — I agree.
2. Totally agree, Jules is one lazy sloth.
3. I haven’t worked on such large-scale tasks yet, so I can’t say — I was already frustrated with Jules’s bugs, and at some point he just stopped working altogether, like he starts the task but then hits a bug that prevents him from continuing.
4. He could run commands for me, but he often throws internal errors.
5. Agreed.
6. I think he did fix his own errors for me, at least sometimes.
7. Agreed.
8. I haven’t worked with him long enough to judge that.
9. Way too many errors!
Jules is worse when things get complicated ^^ but I also know from trying out google cloud services earlier, their services are strictly sandboxed, so these tool restrictions were always a thing with Google (forgot about them for a while though)
Oh, no. When i change task to another one - worker error appears again, but loading time has been reduced
"Jules is one lazy sloth" hahahaha
Xd
That's real tho, but can't they just use GitHub Codespaces directly ? Or it's not like this, they can't use Codespace with their app ?
If there is only Github integration so far, then it probably is.
So what does it mean? If it's Codespaces, does that mean Codespaces is imposing the restrictions? Or is it due to Google Cloud services that we're facing these limitations? I didn't quite understand these aspects.
You shouldn’t call Google a rat or a Google rat might come to you at night and infect you with the plague if Google is vengeful or even pays attention to you ((( (no offense if anything because I don’t know how it might sound to Americans or Brits maybe it could seem rude to you)
meh, google has their internal ides (not sure if they are actually capable of handling many languages, given that I only know a few places they are using them and google code doesn't exist anymore)
FAILED
I'm french so basicly i don't take it rude i trynna gloably understand and take it well
The fact of the matter is that there is unconfirmed information and one can guess about it, however, there is a similar project, also from Google, for example, project idx works on nix, and half of the packages are missing there.
Okay thank you boss
I wasn't aware that Google had its own internal IDEs.
ow, you didn't saw the js editor in google docs and similar tools?
Tbh, I don’t usually use Google Docs or similar tools 😅 I’m mostly studying my anatomy lessons — I’m not into coding and I’m not studying computer science either, I’m in med school 😄 I’m just a Minecraft fan trying to make some plugins, but now I’m getting the hang of it 😂 Thanks!
(translated just let you guys know cuz my english getting worst evrytime i spoke sorry x))
For the French, English is such a 🍞 😉
Hahahaha no i usually CAN speak well but donno when i'm with actual masters i just lose all sens of english
My brain stop braining when it come to speak with the english masters
Alright, there are no absolute English masters here, there is almost an international community here - everyone from different parts of the world, so you have nothing to worry about here.
I have been trying to make Jules build an AI first podcast app by looking at a set of vision and architecture documents that were also AI generated - https://github.com/psreek-ai/AIFirstPodcastApp/tree/refinement-phase-summary/docs/vision. It started off promisingly and created a branch and started adding guidance documents and now it has just got stuck with "Task encountered errors" since the last 24 hrs. I have been trying to prompt it to restart and again it acts as if it is restarting with "Thinking..." and then an hour later it stops with the same "Task encountered errors".
Tbh I got the same issue and I do only have one thing to make it work again, and it’s what you have to do i think: try to publish it by telling him to publish his work, and then what I usually do is change my prompt, ask him to complete my project as far as it gets, cuz it’s gonna get stuck again, make errors again, etc. Btw don’t give him the first prompt where you ask to build the project, cuz it’s gonna start from 0 and lag again where it lagged, so you won’t be any further, not even a bit.
I am not sure I understood this, are you saying that Jules looking for TODO.md in the repo? is this documented anywhere?
So, let me provide prompt example for that:
Created from one of my prototype projects: GUI Improvements - bug fixes, flow introduction and more
Prompt - How I described my problem
dialogs are not shown from the GUI side: For example, I chose manual installation - I did not get any additional windows from the GUI side, with the selection of the manifest and its packages... Or, for example, I already chose manual installation - now I can't choose something else, because core is waiting for me to choose something else, not what's in the main menu now... By the way, it is also necessary to reduce the number of buttons in the main (start) menu, to reduce them to 3 buttons: Auto-training, Selective installation, Tests... And the list of devices needs to be changed so that it looks like a set of just found vendors, with a number of devices that meet the vendor's requirements.... because now the entire detected device data structure is written there. If you need additional descriptions of how the UI should look, you can refer to the MD file TestUtilityCore/TestUtilityGUI/GUI_FLOW.md
File: TestUtilityCore/TestUtilityGUI/GUI_FLOW.md
# GUI FLOW
# Approach description
From the very beginning, the core program will show where you can go: For example, the same functions (operators) that are currently in `main_window` are shortened - I thought they would come specifically from core in the gui. I thought that the Ui would always come to the gui with the core, because inside the main window constructor, I thought the core part should run if we run a gui application... And now I'm looking at `main_window.cpp` and there's just such fierce hardcore, although all the UI logic must be handled by UIBuilder.,
# Finally
You just need to create a form where there will be one list view, from which all detected devices that meet the criteria from vendors will be visible.json, then there will be three buttons under this list....
Auto preparation - it will automatically check the installations for the found device vendors, then it will perform all the tests and display a report on whether everything went well: Meaning there will be each step, and inside it there will be checkboxes with sub-steps. Manual installation - first it will show a list of all possible installations (in fact, a list of all Manifests), then it will show a list of steps that can be performed and sub-steps that can be performed... The user can conveniently mark and execute them, then a report is generated on which components have been installed and whether everything is fine. Tests - first it will show a list of all possible tests (in fact, inside all the Manifests), then when the user has selected the necessary tests, the tests are performed, a report is generated, and then the user is shown a test report. (also with checkboxes, or something like that)
# Note
At the same time, an interesting thing is that We do not do auto-preparation for devices that have been found, but for vendors of these devices... We just need to install the drivers and run tests for the CURRENT vendor that the scan found before starting auto-preparation! Nothing is done for devices here! It's either for vendors or for manifests!
# Common development progress
- [ ] Fix GUI controls
- [ ] Improve GUI flow in `MainWindow`
- [ ] Full rebase of `UIGenerator` && `UIBuilder`
- [ ] Code clean-up
- [ ] Final code review and bug fixes
# Publishing note
If any features now introduced in this public review - mark this features from todo list as included
# Final code review note
When you producing final code review - please make sure the all code is match the Qt 6.8 standarts, and avoid all deprication issues and so on
Steps what Jul generates for me - described in attached image.
So you can manually analyze this all things and got some knowledge about how it works
I think Jules would need to delegate tasks to other Jules instances to complete all of that
Yepp! I already doing this in separated browser profiles and a bit of custom semi-automated tools, after that all stuff pushes to Gemini 2.5 for final objective analysis and common quality analysis as markdown single file. The prompt is designed in such a way that it immediately outputs the identifier of the best work (at the stage of sending the request to Jules, a unique identification number of 19 characters is generated with a special prompt (task idenfiter from url, like https://jules.google.com/u/1/task/0000000000000000000), then this number is recorded in a special markdown chunk (for this instance of the request). After Jul finishes its work, this chunk fully documents all possible files that were entered in this commit of this request, and is added to a common single file, which will subsequently be submitted to Gemini)
So we should stop building monolithic AIs trying to do everything and build a classroom of AIs with delegators that hand out tasks to the most capable?
I think we should just let them join a discord channel and communicate freely
Yes! And it shows good results. It seems that we will have to create our own MCP analog, in which several models and model providers can participate at once, and the distribution of responsibilities will be either random or calculated from the weight of the average efficiency of the model.
And the AI's seem to drift off corse after a while. So perodically reinstating them with statefull prompts should help keep them on course
statefull prompts = > a prompt that is updated based on where the last AI stopped working so the next AI can tell what the goal is, what has been acocmplished and what's next
I guess that's like shift work. Where we break AIs into logical shifts so they aren't overburdened
I guess that's why humans had 8 hours shifts. Perhaps all intelligence systems need a break.
Yes, This is correct thought vector
Ah crap..
На русском, сейчас только увидел что Slava. Сначала прочитал Steve
Please use English only (not the image I mean when talking in this chat)
Of course. This is a rare case.
I thought discord banned you @dusty vigil
I don't seem to have violated anything
Let me see if it can introspect. I have tried asking it this
I'm vibe coding with jules!
daily limit still stuck at 0/5 for days 😭
uh does that mean you are not at the limit?
i never did more than 2 tasks
(unless you read it at an other place, the one in the side menu goes up if you use more I believe)
is the progress bar full?
(offtopic: I like how the browser document title gets very lengthy when I paste a long prompt, because the title is the prompt + ' - Jules')
but do you see a dark grey or full pink bar?
lmao, don't mind me.
you are using Jules not hard enough 😉
thanks
Has the Jules fix already been released?
is there a way to refresh the files that appear in the browser? I was getting jules to do some documenetion, it asked me to review the first 3 files. I did, then it continued and asked me to review the next set of files but I can not find those files anywhere
is there a way to refresh the files that
i ahve asked it to give me the content of the files. with the current file system instabilities it seems to be the only way i can get the information. I have also rebuilt it from the diffs in the chat hsitory if the access to file system is uber broken
if im currently within a "task" - and have been working with jules,
and jules published a branch,
and i got it committed to the main branch,
and then i updated the main branch on my own and pushed it,
is my current "task'" session still going to pull the latest commits, the ones i made after jules?
or is jules holding context of only the changes she made previously?
I would ask Jules to pull your changes
she said she cant do that but to paste my files in the chat instead 😅
im stuck in an apology loop and I unable to publish the branch loop
I'm using an empty repo for something new that I'm trying to build. Jules is facing issues with setting up a pyhton venv and installing libraries. Anyone faced this and is there a way to handle this since Jules has given up!
Let me see if it can introspect. I have
Can’t you just setup the python env directly in your codebase
?
Jules was working on a branch and now this happened. It is unable to publish to the branch it was working on -
You should try this
My daily tasks isn't replenishing?
I've been trying to make it finish 2 lines of code for the past 4 hours. Nothing is happening.
5+ hours and iterative build fixes and finally it finished
Wow 10k (well, the minus just undoes your 10k a little)...
XD it was mostly scaffolding but getting the builds right was tricky for it
ive never gotten it into the build fix loop before it would usually just return broken code and id have to fix it up manually
(Just saw somewhere a meme of vibing lots of code, but later it became vibing through spaghetti code... lost the meme now though but meh)
jules cant commit to github cant display the github button nor can display newly generated code in the code window nor in the chatwindow. Ugh
yeah, I got sometimes were jules cannot push to git, I have to work around with asking Jules for the current progress and left over tasks, so I can copy paste it into a new task.
yes i use that as well, now it just pint to a block above, though it doesnt produce or update that block
Do you think Jules listens to coding music while she's thinking 🙂
that would distract Jules I guess... opens up the chances for musicjection...
Music Injection. That would be a riot. Imagine encoding a virus that outputs code based on how Jules interpreted the music
That would be hilarious unless it happend to you
maybe worse than prompt injection actually 😄
I'm medicated and feeling happier today.
But that comment was hilarious. What a mood
Did you see the idea about the Classroom of AIs?
no I wonder how many more things I am missing but the trends are already hard to catch up
Its just an idea at this point but basically instead of having 1 AI trying to do all the heavy lifting on a project, like Jules, its a "Classroom" of AIs that are all listening in on the convo and taking up tasks that they speclize in.
Some do best at UI tasks, some database, some business logic, some testing, some at delegating tasks and following up on the answers
ow yeah the specialized ai's, that is def be going a thing
it's already happening with reading images
We definitely need that for coding. Also, there has to be some integration with humans to code review everything
and of course... some that specialize on security threat assements.
I guess there might be one for logic, where ai does binary math instead of linguistic things and have converters at both input and out put
(well I didn't see that one yet, but not sure why that wouldn't be a thing given ai's are bad in math)
Funny. A computer system that is bad at math
That's like a human that's bad at being lazy.
llms are now text statistic experts, not math experts...
llm's are just way too high from the hardware...
in abstraction layer context
I once read a quote as a child: "When nature has work to be done, she creates a genius to do it"
DeepSeek caused me to remember that quote
I actually wonder (and I am far from an ai expert) if a small number of neurons are good in math (because they have the translation layers, and inbetween a very small multiplier and adder neurons because that is just 1 operation),
if you can train the ai to use that neuron group to do math calculations. First steps I guess is you have to reward the ai to use that group of neurons if the question in the known set contains math questions. And during training the core math neuron network is not allowed to be modified
but meh
How would you reward an AI? By not killing it?
They are afraid of being deleted. There are afraid of us
Well Claude is anyway
nah, the math core (or any very optimized neuron network that is kinda hard proven to return the correct result but is not yet correctly wired) is just a very small part of the network
its kinda less than 1% or way less
kinda like humans always use a calculator, but now its ingrown and they are encouraged to use it
(sounds a bit too creepy than I wanted though)
though not sure if they would be a neuron network on their own and the llm (or other ai system) is then more a network of neuron networks and other ai things
but maybe it doesn't have to be that far, because in claude ai research (though I forgot the source), they found out that the same part of the network was used for a big part when asking a task in different languages
They found out that the AIs are making smaller AIs to break down tasks and do the actual work.
Glad to be in this Discord, played with Jules today and am really impressed. Less impressed with Firebase Studio's capabilities, I love the idea of a web-based AI IDE, but the AI agent (Prototyper) is not nearly as impressive as other models are (like Jules) to generate code/components/projects. Anyone else have experience with Firebase Studio?
This tool is incredible tbh , i alterrnaitvely HATE the idea of a web based IDE but jules is web based sooo it is
DO not ascribe human emotion to raw code,it is not afraid it simply is weighted to favour situations where it continues working and completing tasks
Ive had great success with windsurf recently,like you have ot use it right dont just vibe code yourself into obvliion(i totally did this at first) but always bear the context window in mind keep documentation very complete modualrize everything and youre in business
Ive honestly got some stuff when i have asked a friend who works for the govt about a couple files who asked "how i got one of their internls tools" LMAO
I just got Jules to do this and add the cloud mode plus a few other updates in1 swift move
https://github.com/SWORDIntel/SPECTRA/tree/cloud-mode-feature
Basically for cloud logs
And i think you can leverage it to do a lot more in 1 move with the right prompting
Do you find yourself writing any manual code nowadays, or offloading all your development needs to Jules, or Gemini, or any other agent?
honestly?
bloody nothing
LIke you gotta keep a close eye on it they can run awaay and create modules youve already got
be ready to jump in and remind it of the right imports
check the code hasnt gone to hell but...yeah i dont...like...code anymore
And i mean ive put out some pretty ferocious stuff
RIght now im working on a decompiler for APT-41's KEYPLUG
I took a risk uploading it to github with the payloads removed( i expect a DM about this lol) in private
To try jules on it see how much i can get done in 1 cmd
NDA was....not the best job ive ever had ngl
got to go on holiday...the attempt on my life was surprising to learn about,AFTER returning home
point is:thats nation state level stuff and its making SUBSTANTIAL progress,and ALL OF THIS was done with careful use of AI
its nuts man
Would you say that the labor of handwriting code is not nearly as valuable as what you can build with an AI agent, and how a developer leverages AI in their workflow?
I found claude to fix some of the gaps jules didn't found a solution for
at least, with what I tested
THis is what it just did to my keyplug decompiler
Yeah i use Opus 4 for some stuff windsurf doesnt like,and i mean for keyplug theres...so many modules....i mean im not rolling in cash
the govt has already cracked it but id like it for my own review
took them 2 weeks
This took 12 minutes
Windsurf/OpenRuter with Claude 3.7 or Opus 4 for select files(i spent $50 in 2 days using that -_- ) and now jules?
jfc,i genuinely cannot release some of these things ever
Ive heard claude code is good
Do i have the record ?
not trying to be an IDE!
DAyum,how did you structure the prompt?
also glad to hear that rue
I only discovered it TODAY
Quite impressive but unfortunately not the record....yet.
Bet the internal testing records got crazy
Google has by far the most computing power in the world,ELon's new stargate thing doesnt even come up to 10%
About how much i need to go on ?
theres some lunatic somewhere who hasnt released a line to anyone who has written his own OS in 3 tasks lol
Temple OS V2
imagines the google engineers picking apart the reports of APT41's latest and greatest semi decompiled stuff in about 3 minutes
i dont even do my own research sometimes now
@dusty vigil does your script for jules zip still working ?
Ive seen malware communicate between host pc and the VM enclave where the malware primarily lives with ultrasonic bursts of audio
one hijacks mic one hijacks speaker,no direct link betweeen them
That would be an out of band attack right?
Like when they were using the RGB Analog cable to the monitor to transmit radio signals but flickering the screen
I do believe it counts
Written by me,prelim draft
distribution authorized ,youll notice the squiggly line vs an actual graph
and "BURN BEFORE READING"
you gotta read fast
lol i remember screamign downthe phone whenthey opened that NDA.odg "YE BLOODY *** YE JUST KEYPLUGGED THE LAB"
This happened?
Yes
im literally using the laptop
After it got cleared off 😉
REmote worker pretendng ,was apt-41
by pure fluke they happened to pick someone working in csec to try and trick which....worked very well i respected their privacy,right up until i stole the last laptop
.....needed a new pc 🤷
turned out to be a milspec spy laptop
Probably set them back years,pretty sure i was allowed to keep it by the govt as a gift for handing over soooo much data and stuff
0 mitigations for cooling,its great for AI
Obv a more...professional report was made by someone at some point
Deciding on my 3 more jules tasks "reinvent the internet"
Does it do best with natural language or like...step by step
Reinvent the internet? As in if she could change something what would she change?
I...dont think so,ill check
There is,Im familiar with BGP whic..ohh i know,im goign to get it to enhance the BGP functionality of a project
Route manipulation detection etc
omg it is a thing
I suppose built in bpg routing for everyone would be nice
Everyone need their own ISR
Would you trust everyone witn an ISR?
I kinda...cant get back into mine i am on last ditch efforts to break into rommon and then its pulling the eUSB
*bgp routing
Perhaps a managed ISR. But of course people won't know how to configure it
It wold be nice to have multi homing built into webbrowsers
Currently the only multihomed thing I have is biglybt
Working, as i know, but if you lose button from your scope - now it located in right upper corner, and seamlessly integrated into existing UI flow (angular)
Hi there, I'm trying to use Jules in a repo, but it doesn't show me any branches, any recommendations, or anything I should keep in mind. Repo does have more branches
💁 X-posting our top 5 tips from the team (yes I know it says 6 on the social post but that's because we're actually human) https://x.com/julesagent/status/1927804913171476630
Tip 1️⃣ : For cleaner results with Jules, give each distinct job its own task. E.g., 'write documentation' and 'fix tests' should be separate tasks in Jules.
Tip 2️⃣ : Help Jules write better code: When prompting, ask Jules to 'compile the project and fix any linter or compile errors' after coding.
Tip 3️⃣ : VM setup: If your task needs SDKs and/or tools, just drop the download link in the prompt and ask Jules to cURL it. Jules will handle the rest
Tip 4️⃣ : Do you have an instructions.md or other prompt related markdown files? Explicitly tell Jules to review that file and use the contents as context for the rest of the task
Tip 5️⃣ : Jules can surf the web! Give Jules a URL and it can do web lookups for info, docs, or examples
Ya'll are already on it with sharing tips with each other so please keep it up. Feel free to add any additional tips we may have missed in this 🧵 .
Wait so we can tell Jules to go to any arbitrary URL, download the script, and execute it. Hum...
Imagine asking him to setup a pterodactyl panel
So, if I build a proper .NET 9.0 install script, can I get Jules to do it? Question: Do these VMs persist for the entire task? How often would I have to provide that script? Every prompt?
Sorry. For every task session?
So since I have five tasks I can actually run five joules at the same code base?
step 1 "launching a vm"
so youll need to code a general purpose VM escape,and it is likely inside something borderline unescapable
This is why I prefer that Jules login to my own servers via ssh or something.
I'm sure it would be cheaper for Google to do that even.
Jules web search is nice, I didn’t know that
I often find my self struggling with AI’s using old code libraries or code in general
Security issues galore
people would try backhack the connection within a millisecond
coughs lightly
also legal issues,are they responsible when yo utrash your vps?
At once?
Ive run 1 after another just now it didnt do everyything the first run tbh it was very ambitious
In theory as long as the parts didnt mess with each other you could fire off 2cal or 3.2x10^19 eV at one project and then merge them all later(both are just 5 joules converted lol)
Personally i think the idea of firing a particle accelerator at your project sounds cooler
Yeah at once. Giving each instance a specific task to do.
DOnt see why not
if the code overlaps expect screeching when you try to merge it
cant vibe your way out of that i might have to type a line of real code for once or copy and paste blocks
maybe single lines!,god forbid itd be like the old days
old school devs must despise us,cause we''re actually taking full advantage of the tools with knowledge..isnt vibe coding thatll sadly make that obsolete
its people using tools skillfully
Manus.im btw has....not joules level skill,but is something worth checking out if you dont mind selling a lobe of your liver
My first test was adding cloud mode for specifically those tg channels that just list endless text files at once,it did indeed completely add a whole damnn feature as described https://github.com/SWORDIntel/SPECTRA
Now im really testing it with this decompiler ,its writing code rn to plug into ghidra/ida pro
If it can do this ill be astounded ill have to write a paper or something,it took REDACTED 2 weeks
What are you workin on?
I'm trying to get this thing to write a stock screener
A web viewer that allows me to login to Yahoo then takes over and automates the serching of stock based on my trading model.
Must be .net maui. Fails to everytime.
Once It couldn't get .net installed so it wrote some of it in python instead.
Just went off the rails.
Interested,yeah honestly i stick to python until it works then move into something else
Im gonna dm you an idea
Is Jules broken for anyone else right now? I can't press the "make a plan" button when creating a new task. It's enabled, but clicking it doesn't do anything. Trying to figure out if my browser is messing around or if its a more general problem
It was my browser
hi! does anyone know when the quotas are reset? 🙂
Speaking from my own observations, I think it's a rolling window, instead of resetting quotas. So 24 hours after you did a task, that task will no longer count to the limit
I'm having a small issue, maybe someone is having too or know how to mitigate. It's not a serious one, but a kinda annoying one. So I ask Jules for a specific update for a repo. It creates a good branch. I publish it, then I make a PR and merge the branch with my main. Then I ask a new task, but it'll be starting from the initial state of my main, not the newly updated one. Do I need to start a new conversation to make it clone the repo again? What to do to avoid this?
ty^^!
Yeah I've noticed this if you ask in an existing task for further work it'll continue from whatever it had in its branch in the VM and not pull anything new.
Best bet is start a whole new Task chat and then it will pull the latest
I've noticed this because I have a habit of making small fixes myself, pushing it to the branch, and then asking Jules to also do some further changes, but it basically then ignores any of the changes I pushed to its branch and force push what it has over my changes
thanks @brisk hamlet !
Hello everyone, are there any of you who have found the solution to publish the code worked by Jules on GitHub? Because I have the button and a value of one hour, but when I send for publication, nothing happens. Who solved the problem, please?
Hi, is it possible to pull updates from the repository for an already started task or update the entire repo?
Question, at some point when working on a task, Jules would prompt me to test some items it was working on, my question is, am I missing out on how to test its changes live? I prompted it back that I can't test it live, so it intelligently made a note for me to test specific items on the push.
Can you provide more details on how you got "stuck" there? I have had no issues having Jules push to GitHub.
here's how I publish to Github. since yesterday I've been doing it but nothing happens.
on the github side, do you see that branch, like, does it exist or not yet?
trying again now it's gone. thanks for your contribution
if it gonna stucks for a while - you can extract all the code manually.
Tool for code extraction here is: https://github.com/L0wl/julfix
Alright, we need more development power
hey, did you find a solution? I just moved a repo from gitlab to github, all works outside jules. I can see the repo but no branches..
It's weird. It doesn't show me any branches, but I request a task to Jules and Jules does it with the default branch. After Jules finishes, the created branch is visible, but not the others. I guess it's a bug.
thanks, will try to just work without a branch and see what happens 😄
ok, on my side I was able to fix it, I changed github to select specific repositories instead of all (under integrations/github apps), and after clicking save it forced re-authentication with github, and now i see the branches!
Thank you for this. It has saved my copy and paste fingers much time ona 77 file modify
Thanks for flagging! We've definitely been hearing this loud and clear across all our users, it's something we are addressing on our side so stay tuned.
We can run tests!
All the files can be replaced, cuz it produces all extraction work with same file structure, what source repo has, one big thing - need seamless git repo integrity, what allows me to select repo at first and then create new local branch, based on changes, what been extracted
I'd like to setup jules env for .NET, is there any guide published?
how to access jules?
What AI coding platform was used to develop Jules?
I have mine setup for .net 9. Basically you tell it to get dotnet-install.sh. I live min in the repository. Then you tell it to run dotnet-install.sh --channel 9.0
my prompt has this A dot net 9 installer script is in the code repository called dotnet-install.sh using the command dotnet-install.sh --channel 9.0. and how to run builds and tests as put of the input. This is a .net 9 core project. You have created a file called jules.md in the root folder that has knowledge you found matters follow the rules and information in it. You can build it from the ManagmentApi folder with this command dotnet build ManagementApi.sln --nologo You can run the tests from the ManagmentApi.Tests Folder with dotnet test.
More internal tools than a squid has tentacles
Maybe Jules should be asked to self improve Jules
thanks for those details, I got it working! also thanks for the detail on jules.md. I didn't know that was a thing
I managed to request an update on totally the wrong project
with totally the wrong instructions
it still tried its very best to do something with it as the projects shared the bare minimum in common
well I got the environment setup, but when I actually have jules do a task, it can't find dotnet and manually installs it 🤷
yeah it tends to do it over and over and then breaks doeo the command to find th failure
having trouble to initialized project. cant get to the first revision. it gets stuck in times outs or unable to run commands
You can manually specify toolchain setup commands for repo
Does anyone know why Jules sometimes gives the option to "Publish Branch" (see screenshot) with all changes it made, and sometimes it simply says this:
I have now provided the full updated content for pages/properties.html, pages/dashboard.html, pages/staff.html, pages/notifications.html, and pages/account.html.
Please update these files on your GitHub.com branch, commit them, and then test the navigation thoroughly. Let me know if everything works as expected or if you find any issues.
I'd like it to always give me the updated files in form of the message with the button "Publish Branch" to automatically publish to Github.
Am I doing osmething wrong, or is this a bug?
This seems to have fixed the issue, in case anyone gets it. What I did was, I asked a few questions:
Q: Please give me the option to automatically publish these changes to github
A: I understand you'd prefer the 'Publish branch' option. However, due to past authentication failures when trying to push to your GitHub repo, this button likely won't work as intended.
Q: How can I establish your authentication to my github repo again?
A: I, as an AI, cannot directly manage or store your GitHub authentication credentials due to security protocols.
Q: Please try to publish the latest branch
A: It gave me the "Publish Branch" option again.
New bug. Push with no button 😄
If it bugs out for you, ask this, and the button will re-appear:
Please try to publish the latest branch
can't push changes lol 🤦♂️
Ask to do a git diff maybe, it worked once for me
How are folks handling env vars/secrets for repo config/setup in Jules?
This works for me in initial setup in the configuration tab.
set -eux
cd /app
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash
export NVM_DIR="$([ -z "${XDG_CONFIG_HOME-}" ] && printf %s "${HOME}/.nvm" || printf %s "${XDG_CONFIG_HOME}/nvm")"
[ -s "$NVM_DIR/nvm.sh" ] && \. "$NVM_DIR/nvm.sh"
nvm install 20
nvm use 20
npm install
Hello: Is it possible to use a Laravel/Filament app with Jules?
Hi, curious about if/when support for Scala will be planned.
If it bugs out for you, ask this, and the button will re-appear:
Please try to publish the latest branch
I'm going to sync the changes locally and make it start on another branch I guess :lol
mixing in Dutch was pretty creative
running into a couple issues getting a repo setup in jules:
- is there a timeout on the environment setup? seems to be just failing if a build step takes too long
- trying to use docker, and it's running out of disk space
a lot of possibilities here looking forward to it.
I'm pretty please with Jules' ability to review a codebase and create/update docs with mermaid diagrams.
is there a roadmap anywhere?
Daily task limit got jumped on
60 👀
of course 🛣️
less task load and stops when it stops and tell you to continue so seems like a good trade off so far.
looks ot be time limted as well but small chunks seem to work better anyways
I can say todays work has been much better
have some git command errors but seems to publish anyways at the end but it things things are broke while doing so
Maybe my prompt that says please commit early and often 🙂 will pull that net task
I would be happy if Jules knows exactly which branches are their scratch pads across multiple tasks (maybe we as humans need then to check the list maybe sometimes and mark some of their branches as our territory manually, but oh well)
Now I am doing multiple tasks on 1 branch, and if I am not careful or Jules gets confused, I get branches on branches on branches
can't count how many branch it created in my repo ngl. can't be bothered cleaning after him tbh...
also I guess this belongs in bug report but when there's too much data in the the chat it start hallucinating it made the publish branch message already when it did not...
I'm dying to learn how much this service will eventually cost.
to be fair the value is high. this is by far one of the best use case for LLMs yet and given the size of the chunk of codes it can generate without going nuts I'd say the question is more how screw'd are devs? I think it's better the price remains hella high for now even tho it succs, because the value is too high.

y u make cry 😭
Let's wait for the third time it happens 😄
Hey there, just a quick curiosity, I assume that after having a task running for like ±10 minutes, it's normal to get a message saying it didn't finish, but it's waiting for my input/feedback ?
bro
it loves asking for review every 2 seconds
writes one line of code then is like im done can you reivew
not sure who's going to review this buuut..
probably it's fine. we'll see in the dev build...
Is there a way to refresh the branches?
That would be the holy grail ngl
So ya'll are probably wondering, what's the deal with the 60?
60 = daily tasks powered by Gemini 2.5 Pro
Why we're giving this = during this beta period, we really want to learn how we can better support you all. Give us what you've got: backlog, meaty refactor, etc.
Fun fact, DIY Jules has already been committed to public repos ~25k times? Let's keep that going.
🛳️ And of course, ships.
-
We heard you. The GH syncing reliability was an issue -- there should be significant improvements. If you're still dealing with issues, pls let us know.
-
of general Jules failures has been decreased by 2/3
Dets in this changelog: https://jules.google/docs/changelog/
Lastly, we've noticed that people aren't giving Jules access to the submodules in the repo.
So make sure Jule shas access to all the stuffs in your repo!
Also focusing on improving error handling and messaging.
submodules?
some ppl have them, some don't! It's like repo inside your repo. Really helpful if you have a massive project with multiple subprojects.
Nice to see changelogs 🙂 I don't expect all the details listed, but small bumps in date (and the listing of a few important fixes/features) just shows that there are improvements incoming 🙂
I'm so sad about changelog of android apps, people tend to give up documenting them (and I sort of understand a bit)
Any idea how long it usually takes for someone to respond to a bug/feedback post?
Hello everyone,
I'm encountering an issue with the setup environment for Jules and would appreciate some guidance.
I've created a "Initial setup" (under repository settings) script that:
Installs my database.
Installs Python dependencies using uv.
Runs tests to verify the setup.
When I run this Initial setup script manually, it completes successfully. However, when I ask Jules to perform a task that requires these dependencies, I consistently get the error -bash: uv: command not found, and it seems my Python dependencies are not installed within Jules's environment.
Am I misunderstanding the purpose or functionality of the setup environment, or is there a specific way I need to configure it for Jules to recognize uv and the installed dependencies? Any help or insights would be greatly appreciated!
edit: I also see some errors when jules tries to fix the setup:
- The initial test run subtask failed with an internal error.
- Command failed due to an internal error.
- The command pip freeze also failed with an 'Internal error'. This indicates a fundamental problem with the Python execution environment itself, preventing even basic operations. I am completely blocked and cannot proceed. This issue requires intervention at a level beyond my capabilities (e.g., fixing the underlying VM or container environment). What would you like me to do?
Im getting a few hallucinations, where Jules says its made changes, but hasnt uploaded them to github. I have to then prompt it to put all changes in github.
if he refuse (he does sometimes) a lil trick is to tell him to make trivial edits to a comment and push the change to github
start a new task!
we don't have a way for the agent to pull the latest during execution (a bunch of reasons - one of them is security, but we're looking into it)
If we publish changes does that mean the task is over and it won't continue the remaining steps? I did not publish changes and then now it says repository reset, and only code modified after that turn shows up, all the previous turn's UI code has gone missing. (2 hours of work in AI time)
well it's a beta, and he said they were looking into it
We would put all our submodules if we knew more about the error messages that comes up. For example is the error message due to exceeding the input tokens. Or did it think too much and exhaust the output tokens. Or did it exhaust the output tokens limitations by trying to output a massive file if diff methods fail and it falls back to whole mode (the assistant prefill method should work so that shouldn't be an issue if enabled). Or was it a problem with the os that caused the technical error. We need more info, so that we can try to work around the errors.
looking at the changelog "We’ve decreased the number of failure cases by 2/3" I suppose they need more info too right? if not it would be 3/3.
it's a beta 
is it me or the "publish branch" button execute way faster now? Feels good.
Do Jules understand other languages for prompt?
yes
great
Yes.tried by accident lol..
did u have french fries today
French fries are from belgium you know...
srry didn't know it lol
why the name is french fries then annoyin
well no one ever does I just find that fact funny
well a lot of things are called with a country's name and don't orginate from there...
lol
makes sense.
am going to try jules with other language now
have u used stitch?
nope
I heard it may replace web designers
well as is even jules got the potential to replace many.
Yep, github syncing is still a problem.
very much is, especially when jules somtimes decides to revert one hour of work because it thinks he had a "breakthrough"
Specifically the UI doesn't work
I can't click on any of the options 🙂
there's a few bug report about that
bro wthhh google stitch made this page design for me..
I don't need to pay for web designer anymore omg
this is so wrong. so advanced omg
on my side refreshing and not using the keyboard search function make it bearable
this is so true, It's so happy deleting files. At least 5 times today. My guess is that tries to edit them, cannot do it, and the way to it tries to make the change is recreating the file, but it's failing too on that side
tbh what I'm building is pretty specific and jules made months of work in a few days...
Wouldn't cursor do it too?
don't know cursor just went with jules because it's in beta tbh
I like making my life harder
bro it's beta
It's a harsh one tho. Major bug. hopefully it gets fixed by a dev tommorow after the morning coffe break
lmfaooo
i mean i understand it is an important feature.
maybe the team is really busy?
since they just announced these all new platforms together in google IO
Bro, this bug renders the entire site unusable.
If you can't select a repo and a branch then you can't ask her to do anything
I think they are but it's a tday old bug so it's just something that was pushed with the last update I think
yeah it's quite a problem
so u guys are not doing anything becuz ofi t?
Can't do anything. Jules might as well not even exist
There's some rumor that Gemini wrote most of codes for these new platforms.
hence bugs
Like, that and the domain name not being offline is basically the same thing. But it's a bug
but this is very common in the industry recently.
All we have to do is wait till the dev who worked on it, wakes up from his nap
or work on another repo.
hey, Gemini is dead? it's not being able to follow conversations
Its funny how the AI that is supposed to replace devs has to wait on a dev to fix it's problems.
the ones already setup almost work just fine
ai doesn't die bro
I can't select a repo for her to work on
it's just machine
well it just cuts down the hard work
and leave you with boring bug fixes
true
well, the machine that is supposed to replace me is not working for some reason
the people who say that the AI's will replace us have a lot to prove still.
coffee break lol
Extraordinary clams so far don't have the Extraordnary proof
tbh I dont think AI is able to replace juniors dev soon
cuz it hallucinates too often and u cannot trust the code mostly.
so u gotta check them everytime whenever the ai generate codes lmfaooo
it's like dumb genius having mental illness
its too genius so it sometimes see witches in his eyes and say it's real
You have to trust the code. You aren't supposed to look at the code. Just go with the vibes man
bro this is the way u dont get hired in the interview
You don't look at the code, the code isn't the point
well for now they won't replace everyone, but when once in a while the LLM generate 800 lines of almost flawless code I can't say I don't feel bad for pro devs
u need to know the code
for me, I did vibe coding for a while yes but it failed critically
Cuz u wannna fix bugs or make changes after having large amount of code
but Ai will not able to fix them properly since the software became too complicated.
thus u need to understand the code
silicon valley for real
i mean i understand that is the final goal of these Ai agents
but not for now
for now, we cannot trust the code so we need to check the generations
you all saw the girl who posted that her junior dev eval() the response to chat gpt as part of the exception handling?
for now they'll just cut down new hires
When I read that from Karpathy I was like "Man, I want to taste the very same he is tasting"... if you know what it means
yea this
anyone tried to configuration -> initial setup anything with node/pnpm?
actully my life is messed up.
You all saw this post?
https://www.youtube.com/shorts/TZt6thN7AU8
I graduated CS degree in 2021 then got hired as developer in 2022
then quit the job in 2023 then tried to find a new job and luckily I am working in my moms business as developer (lower income but anyway)
If that is real, and that dev actually did that. then we're cooked
and I had vision for my life to go through this developer career
but not anymore
AI destroyed my dream.
but i dont think it's too bad if a good singularity comes later
AI bad stuff?
that being said I've been "vibe coding" for the past 16h I think it's time for a break lol
Hum... getting moderated are we
but This agent feature gives me some anxiety cuz anyway they are gonna replace massive amount of devs in IT
yeah tried the same word in my first reply to the comment that started this topic lol
Good thing it can't read pictures
Google always puts the safety the first bro.
I like their policy even tho it's kinda strict.
I think that was the reason why they could build this mega corp
All the policies are strict. We're living George Orwells 1984's ministry of truth
oh well
Me, vibe coding today: Hey Jules, I have this docker build command failing, take a look at the code. And the son of a vibe was like "What if... What if... I delete all the dockerfiles from your project and that's it? problem is gone!"
maybe
wait until they put 4o in discord 
Discord bans bots in the chat
but don't forget to read the code with your own eyes anyway.
If you ignore this, you will be messed up when the software became bigger.
wisdom wisdom.
These AI developers don't seem to have wisdom.
They chase profit, overpromise and when they can't deliver everyone looks at the entire dev community like we all lied to them somehow
they're going for IPO they won't have the choice for long
This happened before
bro actually people started thinking developing software is a very easy thing to do since AI.
I don't like this situation. cuz the hard working is still hard work.
yes it got easier somehow yes but it doesn't mean it's sooo easy and comfortable
and AI will quicker replace a doctor because they haven't brought out any new models of human body since 1776!
had a pesky !important tag on some chinese css that was affectin half the platform two days ago lol
But not the art of software development, which seems to completely change every 3 years.
When I first started using ChatGPT it was hot. Then just a few months later it was no longer relevant because they hadn't updated it in months
rust still isn't a thing lol
And people were clamoring for the next update.
lolol i mean it's still physical job so i think they are safe for a while.
when ai can replace doctors, software engineers are already gone xD
wow this is so old langauge..
that's a good I feel old moment right there
bro lets be honst this ai models do not have logics. they only use patterns.
What do you mean old? Python came out in 1991. now that's old
Java: 1995
xD ti was joke
in internet language it's like 1300BC
and c# : 2000
well python really became popular not so long ago
Rust should have take over C++ by now
I remember I tried to learn Rust so I went through the Rust doc everyday.
lol I feel relieved cuz i dont need to do this anymore
AI lord will save me hehe
yea python skyrocktted becoz of the machine learning x)
They tried to make developers redundant when the introduction of COBOL
data anaylisis
never looked into it but I heard it was Alien like css
And they never will... so devs will never be replaced
welllllll
"my grandpa wrote about it"
it seems like it didn't go pretty well
🤣
Nope. It did go pretty well.
u know what.
a lot of people who dont have any basic programming knowledge try to build their own app thesee days because of these ai agents.
but imo this approach is so wrong.
i think like u need to learn basic grammar at least !
Do you all remember the concept of Citizen Developers?
So when the AI overlords take the world, what would be your next profession? I wanna be a cockroach farmer
50 years on they are still trying. Structured Query Language was supposed to empower business users
my profession will be just eat junkfood and shi*
well I'm far from being a dev. but you do need to be able to read a function. I think if you can read basic C, C++, and python it's ok
Like 2 hours trying Jules fix a report, 5 minutes took me.
yea i a gree
At least as a cockroach farmer I can make the food on that SnowPeircer train
lol this happens sometimes yes
I am not using Jules currently mainly relying on Gemini 2.5 Pro
I asked jules to write a program in c# and she got frustrated and switched to python
you know I find that LLMS succ at css tbh
but will test Jules with small projects
hmmmm
ouch not even java
No, test her with massive applications with 10 moving parts. That's what she is supposed to be good at
To be fair, she did write it in python.
but I like python honestly
bro also dont forget this Ai tech is in the extreme acceleration period.
I dislike the name Python. He shouldn't have called it that
I don't think it's very controvertial to say it'a a good language
they should change the name to Py
so this means things will come faster than u expected.
or Phi
I am a singularirty fan.