Hey, Today I am sharing a topic related to gemini ai Prototyper. I found some genuine issues in it.
- I found Gemini is not following the instructions properly. It is mostly trying to make assumptions or ignore instructions. I am telling this because There was a bugs in my codebase where I failed to fix that. I thought to use gemini but after using I realised my biggest mistake. I don’t found any Rollback for First message. And the model made my entire codebase broken removing my own logics, converting environment variables from real to placeholders.
- I Given them the correct prompts and clear instructions with “forcely” terms to prevent it but it didn’t do so and made me to code entire system again.
- The model is I think of 2024 a old model I think it requires some improvement about new nectjs 16 version because various rules have been changed or being strict but the model is trying to code for old model and maximum time it is downgrading the versions. I given them the clear instructions but it didn’t followed them as a type of “auto - fix” it is making maximum mistakes.
- I have created GEMINI.md and airules.md in dev.nix but I found it is not applying my rules. I think it is wastage.
- The platform is trying me to build and use only firebase or Google products. Whenever I am telling the model I shown always giving me excuses of “tecnical limitations and programmed limitations” I can’t connect, basically in other providers of object storage or cloud products.
- The model is always trying to use “genkit“ for making any AI features and use gemini only. Whenever I am trying to Build my own connection for other model’s the model changing the Logic to genkit again and again.
- The eagerness and shortcuts is very bad of the ai Prototyper. It mostly try to use shortcuts and use prefixed mocked data. I can’t understand why it is doing so. The user is always trying to build product ready platform but the ai is always trying to make only just a visual illustration and prototype stimulated platform. If any real thing is already implemented then changing it to make best prototype.
- It can only read a single image at a time which is also a bad. I think it may have capacity to read at least 5 to 7 images at a time it will make more friendly to understand users understanding.
- The making of unwanted editing in code. Whenever I have issue like cookies () isn’t function. I understand it is due to new nextjs16 rules it Must be awaited I just commanded the ai to make awaited my Cookies function the error will automatically resolved but in reality it didn’t awaited the cookies it removed all other awaited const from asynchronous (“async”) and then after not solving error as a Auto fix it then downgrade next version to “14.2.2” when I asked then it is saying me sorry but then saying the next version 16 is not available in npm current version is most stable. The codebase is my own property. Then Why it is making such more wrong information it just have to do accordingly to users instructions Instead of making any unwanted assumptions. The Agent my also have some tools like if there are any misunderstanding then instead of assuming according to previous chats it agent will ask directly to user related the assumptions and after approval the agent will update code respective places.
- There are no any ‘copy’ system in terminal. I haven’t know how to copy From terminal. Whenever I required any tokens secret I need to use different idx for getting token the current terminal is not such helpful for me.
- The ai Prototyper is currently supported much more tokens at a time but whenever I am importing source code From my GitHub mostly it is being ranout there Token limits. The most frustrated things is that it is not working until I haven’t remove that souce code again. I can’t understand about reading entire source code at a time.
I don’t know wheather this is only my issues or others also. I think Google may improve it.
Thanks.