While building my portfolio using Google AI Studio, I worked extensively with Gemini 2.5 Flash and later switched to Gemini 3 Flash. Using the same prompts and resources across both models gave me a clear view of how differently they behave in real world development workflows.
This article shares my practical observations what worked, what broke, and what I learned based on hands-on usage rather than benchmarks.
This is the prompt I started with
Build a personal portfolio website using my LinkedIn profile as the primary source of truth (bio, experience, projects, publications, achievements, skills, and writing style).
The site must deeply reflect my personality: minimal on clutter, strong on visual storytelling, motion, and subtle surprise animations.
Tech & Performance
- Framework: React (modern functional components)
- Styling: single consistent theme & style system (choose one: futuristic minimal / editorial modern / dark creative / soft glassmorphism — pick the best fit based on my LinkedIn tone)
- Animations: Framer Motion (or GSAP only if truly needed)
- Icons: react-icons only
- Performance-first: fast load, lazy loading, optimized assets
- Fully responsive (mobile-first, tablet, desktop)
Overall Visual Direction
- Keep the layout simple and breathable
- Animations and visuals should be the “wow factor”
- Motion should feel intentional, smooth, and surprising, never noisy
- Use micro-interactions, scroll-based reveals, and hover effects
- If background images or visuals are used, they must be subtle, abstract, and performance-friendly
Sections to Build
Loading Screen
- Custom loading animation tied to the theme (e.g. animated logo, name morph, abstract motion)
- Smooth transition into the hero section
Hero Section
- Strong first impression with animated typography
- My name, role, and a short expressive tagline derived from LinkedIn
One standout animation (text reveal, parallax, or motion background)
- Clear CTA (View Work / Contact Me)
Projects Section
- Projects pulled from LinkedIn
- Clean card/grid layout
- Hover animations revealing details
- Smooth transitions when opening project details
Publications & Achievements
- Elegant, editorial-style layout
- Subtle scroll animations
- Visually separate publications from achievements while keeping a unified style
- Blog Section
- Minimal blog preview cards
- Animated entry on scroll
- Designed for future scalability
Contact Me Section
- Simple, modern contact form
- Animated focus states
- Social icons using react-icons
UX & Details
- Consistent spacing, typography scale, and color system
- Smooth page transitions
- Accessible color contrast
- Clean semantic structure
- Output Requirements
- Provide clean, readable React code
- Use reusable components
- Include animation logic
- Add comments explaining animation choices
Ensure production-ready structure
The final result should feel like a personal digital experience, not a template something that quietly surprises the viewer while staying elegant and fast.
Keep the code minimal and understanding to be modified in the future
The Loop of Doom - Gemini 2.5 Flash
Gemini 2.5 Flash offers strong price-to-performance value and is well suited for large-scale, low-latency tasks. I started with this model because it handled prompts quickly and generated usable project scaffolding.
However, during iterative development, I ran into repeated issues. Certain file-level errors kept resurfacing even after multiple corrections. The model often attempted to fix the same issue repeatedly, sometimes reintroducing changes I had explicitly reverted. Long prompt chains seemed to increase instruction drift, making it difficult to move past specific blockers.
Despite trying version rollbacks and prompt refinements, I frequently found myself stuck in loops. At that point, I decided to restart the project using a newer model.
The Savior - Gemini 3 Flash
Switching to Gemini 3 Flash noticeably improved the experience. Using the same prompts, the model responded faster and produced outputs closer to my intended structure. Refining layouts and components required fewer iterations.
That said, Gemini 3 wasn’t flawless. During longer sessions, it occasionally repeated earlier mistakes or applied outdated context. File deletion and local asset handling were particularly challenging. Adding local images required workarounds such as using public URLs, and even then, results were inconsistent.
I also faced issues with navigation links and in-page routing. While Gemini helped identify some problems, I ultimately had to manually adjust the code (for example, replacing anchor tags with button based handlers) to achieve the desired behavior.
The final result came together only after manual refactoring and selective AI usage, rather than full reliance on the model.
My Brutal Takeaways
- Both models have different "intelligence" levels, but they hallucinate on the same logical hurdles.
- Gemini 3 Flash is faster and generally more reliable, but long iterative sessions can still cause context drift.
- As project complexity increases, model limitations become more visible.
- Developer knowledge is still essential to resolve edge cases and unblock progress.
- Both models occasionally ignore negative prompts (e.g., "Don't use File X") and execute based on old patterns.
- AI works best as a collaborative tool, not a fully autonomous developer.
Both models are useful when used with the right expectations and constraints. Prompt quality matters, but workflow design and manual oversight matter just as much.
What’s your experience with Gemini 3? Should I compare these against other flagship models?
Top comments (0)