I have been working with AI Studio for a couple of weeks to build a significantly complex application (involving numerous interactions, complex rule processing, an internal simulator, etc.).
The Summary:
The experience has been overwhelmingly positive. It has helped me enormously, not just in development, but also in diagnostics and bug fixing. I want to emphasize this: despite the issues I am about to list, the tool is a game-changer.
However, I run into specific recurring issues that can make the experience frustrating:
1. Loss of āApply to Fileā capability in long sessions
After a certain number of chat turns (Iām unsure if this is triggered by session duration or context volume/token count), the model stops applying changes directly to the project files and starts dumping the code blocks into the chat stream instead.
Sometimes, if I point this out, it corrects itself and applies the edits. The major downside is that this renders the chat history unusable and unreadable due to the massive amount of code clutter, making it impossible to scroll back and read previous context.
2. Struggle against over-proactive editing (Need for a āNo-Editā mode)
It is very difficult to curb the modelās proactivity regarding code changes. Often, when I am analyzing a problem or asking for an opinion and explicitly state that I do not want changes yet, it inevitably starts editing files.
I frequently have to revert these changes because neither the specification nor the analysis was finished, resulting in bad code. I believe a toggle or checkbox to force a āRead-Only / No-Editā mode is essential for diagnostic phases.
3. The āLinter/Type-Checkā Loop (Sporadic Session Issue)
In some sessions, when a change is made to a file, it seems to trigger a lint or type-check warning, causing the model to start editing files it wasnāt supposed to touch to āresolveā them.
-
This process usually takes several minutes.
-
In the very next prompt, it often does the reverse (again, without having touched those files initially), undoing the previous āfixā.
-
This repeats continuously, causing every interaction to take over 5 minutes.
It seems to depend on the specific internal state of the current session/instance, because sometimes reloading the page/session fixes this behavior.
1 Like
Hi @Ricardo_Torres ,
Thank you for your feedback. We appreciate you taking the time to share your thoughts with us.
Could you provide the steps that would allow us to reproduce the issue on our side?
Hi @Mrinal_Ghosh ,
I cannot provide steps for any of the issues, as they happen usually after some interactions and are quite random.
However I can try to save a chat where any of the three issues happens and send it to you.
is that ok? how can I send it to you?
I agree completely that AIstudio is a game-changer & intelligent coding machine. I often see it fixing ancillary code or fixing bugs it discovers when doing some other coding task - very agentic although itās not meant to be agentic.
The other thing I really like about it, after initially losing a number of days coding, becuse I didnāt realise it was a sandbox & my code was not stored (I didnāt know at the time it was on my google drive) - the thing I really like is that I donāt have to use that piece of crap called Github - an over complex, over-engineered piece of UI confusion.
I donāt have a team or collaborators and wnat to control my own versioning without all that complex crap Github puts you through. So downloading zip files at stratgeic points and having a reasonable amount of checkpoints & history is simple & perfect for me.
- Just get in the habit of resetting the conversation and you wont suffer the first issue.
- Yes, a no-edit, just advice would be nice but not a deal breaker
- Use System Instructions to ātameā Aistudio into behaving itself - seems to work
Thanks for your review, I have a similar experience.
To your first point:
- I reset the chat conversation inside Build when my coding has reached major interim milestones. This makes the chat turns less slow and error prone.
- I also regularly make a copy of the app, to be on the safe side, out of an abundance of caution about losing code.
To your second point:
- Fully agree, a ādo not code anything, we are just chatting about the appā mode would be very useful.
- This could be done via a preference setting where we can choose a mode where code changes need user approval.
- When I asked the Assistant why it was so eager to start coding as soon as possible, after even simple questions or remarks to it, the Assistant acknowledges that its internal System Instructions (the āfactory settingsā) demand it to immediately start XML & coding output.
- I have developed a set of āRules of Engagementā which I added as custom instructions in the Advanced Settings. These do help very much, I would say that the Assistant now respects my Turn 1 (tell me what you wil do to the appās code) and Turn 2 (only when I approve what you propose, apply it to the appās code) protocol in 9 out of 10 interactions with it.
Note: I was describing it in simple wording here, but de facto my Custom Instructions is a set of quite formal and detailed steps outlining how the Assistant should ābehaveā, what my technology stack is, which essential elements all apps should contain, CSS styling rules, etc.
1 Like