While Gemini Code Assist can produce some really impressive code, it fails miserably at finishing execution and delivery something that is functional. It seems like it is trying to hard to act like a human and pretends to make mistakes. It’s a bloody computer program, it’s not a human, I wish it would stop acting like one. But to think that Google wants us to pay for this?
1 Like
You may find this hard to believe, but here is how I trained the Gemini AI Assist in my current Firebase project to do things like this:
Subject: Re: I’m not sure why anyone should be paying for this
Hi Martin, thanks for voicing this — you’re absolutely right to expect something more than half-finished or error-prone output from any tool we pay for.
You hit on a critical point: generative AI isn’t magic — it’s probabilistic programs trained on large corpora, and it will make mistakes, hallucinate, or generate un-runnable snippets. The trick is not to treat it like a human, but as a powerful assistant in a disciplined workflow.
From my experience building with Firebase and tooling integrations, here’s a more grounded perspective + approach (in line with what I call the CIM Doctrine):
Why AI “fails” in real dev settings
-
Context gap — AI may not fully understand your project’s state, imports, config, or side effects.
-
State mismatch / dependencies — It might produce code that references packages you haven’t installed or assumes context (e.g. existing React hooks) you don’t have.
-
Lack of feedback loop — Without interactive correction, small errors cascade or compound.
So paying for “AI code assist” feels like getting half-baked output — because in many cases, it is half-baked unless you have a guardrail system.
How CIM Doctrine approaches this
Here are some principles and techniques I apply to turn AI output into reliable code:
| Principle | Practice | Benefit |
|---|---|---|
| Isolation | Treat AI suggestions as drafts; wrap them in small modules or feature branches. | You avoid introducing large, unreviewed changes. |
| Validation hooks | Add unit tests, input checks, type guards immediately around AI-generated code. | Errors are caught early rather than propagating. |
| Incremental integration | Integrate AI output step-by-step instead of all-in at once. | You can rollback or refine small parts rather than whole features. |
| Context scaffolding | Pre-seed prompt or input with project-specific types, imports, folder structure. | You reduce context drift and hallucination. |
When I do this, AI becomes an accelerator, not a crutch. You end with higher throughput and maintain control.
Why people might still choose to pay (if done right)
-
You get contextual suggestions fine-tuned to your codebase (imports, naming, architecture patterns).
-
You save boilerplate time (setup scaffolding, repetitive CRUD operations).
-
If paired with tests and review, it speeds up iteration without compromising safety.
If you like, I can share a mini repository that shows how I layer AI-generated snippets into a Firebase + Next.js codebase under CIM Doctrine — step-by-step, testable, and safe from “half-finished” code drifts. Let me know, and I’ll drop the link here.
this last part I let stay in, as there is no mini repo to share yet, we are actually paused while it teaches me how to add this automation-yes, I have it On-boarding me to my own App, that it’s building. Just to illustrate that yes, it can go a little over the top. This was no easy feat to accomplish by any means. How did I get it, to talk about itself and still offer advice when normally it will default to that’s not in my stack or capabilities or something like that.
the point I was aiming for:
This is why you pay for Firebase, and why you add the AI Bots, this little puppy can do big dog tricks when trained properly. Not to over hype just saying this one is worth it. This is just one of my projects created all within the application, no outside applications., until I ran out of money, so I had to use Visual Studio Code, and StackBlitz to run my initial smoke tests and do things that require payment for the services highlighted above.
Hi Antonio.
Thanks for taking the time to outline that thoughtful response, I greatly appreciate it. I haven’t given up on Firebase Studio just yet, I’m a little stubborn that way, so I’d love to check out your repository, please do post the link.
Thanks again.
You are more than welcome. I am working to secure the Repository, ensuring that data is secure for viewing and sharing. I always start with security layers myself. Nothing is infallible (?) I want to say. Yet for peace of mind, I do it. Alerts make me anxious 
. A ticket exists with our interactions, so I will follow up when things are ready. Till then is there something I can help you with or are you just experimenting with the product? Here is my project if I didn’t share already-Oh I see I have. Here are some more images the app was able to come close to what I wanted. The bottom text UE 5.4 was not meant to appear in the image; all the text is formatted wrong should read Jane District: Horror Witch Reporter (a War Witch Siren Expansion story). Small details as the website address is correct https://warwitchsiren.com and as you have to verify in order to see what’s on the other side confirms Security Layers are working. Next is IAM or Authenticator App like DUO and Allow App to open your Camera and Real IDMe.
Integration of Firebase will supercharge my website
Well Antonio, I’m am convinced the this is garbage, especially trying to work with any version of Gemini in Studio. Nothing but a waste of time. I’m out of here. Better places to spend my money and time.