By skyforbes Hi everyone,
I’m currently "vibecoding" (coding mostly via natural language prompts) a training platform for medical exams (Quiz/QCM app) using React and Supabase.
The Context: I’ve been using Gemini 3 heavily. Honestly, the start was impressive. We built a solid foundation together:
- Stats: About 3,500 lines of CSS and 1,000+ lines of JS logic.
- Features: Multiple question types (Multiple choice, Zone picking on images, QROC), a gamified dashboard, and a Spaced Repetition System (Anki-style review mode).
The Problem: I feel like I've hit a wall. As the project grows, Gemini is starting to degrade significantly:
- Regressions: I ask for a UI tweak on a specific page (
Ancrages.js), and while it fixes that, it reverts 3 other bug fixes we did 10 minutes ago (e.g., CSS colors for correct/wrong answers). - Context Loss: It seems to lose track of the file structure.
- Looping: It keeps re-proposing code that I explicitly rejected or corrected previously.
My Questions:
- Is this "context saturation"? Is 4.5k lines of code + chat history simply too much for Gemini 3 to maintain consistency?
- Model alternatives: Would Claude or Chat-GPT or other be better at handling this size of project without breaking existing logic?
- Workflow: For those vibecoding apps, how do you manage this? Do you feed the AI only specific files? Do you use tools like Cursor or just copy-paste?
I really like the project, and I really want to have a near perfect result.
Thanks for the help!