Vibecoding
Build for the vibe, debug later
Start new thread

Vibe coding is fun until the AI becomes dumb all of a sudden.

by•

I used a few different tools to vibe code my saas. I used lovable, chatgpt, gemini to code my saas. The thing is, there is a huge problem of these AI's of forgetting when you chat for a longer time.

This has happened to me a lot of times. One solution I found is that you can start a new chat to overcome this problem. But, here the problem arises to explain the AI all the context which you had in the previous conversation.

Curious, has anyone else ever ran into this problem? How did you manage to overcome this?

61 views

Add a comment

Replies

Best

i've had the same issue with bigger feature. now i keep a short summary of the imporatant context for the next chat.

this happen to me when a project gets more complex. i usually save the key requirments before starting a new conversation.

I have noticed the same problem when changing several parts of a SaaS at once. I now ask for a summary before ending each session and carry that summary into the next chat.

for me it get worse when i'm fixing and adding features together. keeping each task separate seem to help.

i usually the AI to summarixe the project before ending a session,

I don’t think starting a fresh chat is really the solution. If the entire product context lives inside one long conversation, you are eventually going to hit this problem again.

What has worked much better for me is treating AI conversations almost like project milestones. Once we finish a meaningful piece of work, I stop and compile what we decided: product rules, architecture choices, edge cases, what is already built, what we rejected, and what comes next. That becomes the source of truth for the next discussion instead of expecting the model to reconstruct everything from 200 messages.

I also would not let the AI decide everything about the product. Some parts should be deterministic: business rules, validation logic, data structures, permissions, acceptance criteria, and things the system should never do. AI can help me build and reason through them, but the rules themselves come from me.

So rather than one endless vibe-coding conversation, I would use smaller working sessions with saved checkpoints between them. You can change models or start a new chat without losing the product every time.

I’ve had this less often lately with tools that auto-compact context, but I still don’t trust one giant chat forever. If a thread starts drifting, I usually make a clean handoff summary and continue in a fresh session with the important project context.