As a student who regularly experiments with various AI tools, I decided to test something fascinating: Can I build a complete app using nothing but prompts? I put two AI powerhouses head-to-head – Claude 3.7 and Google Gemini 2.5 Beta – with one simple challenge: turn my idea into a working fitness app prototype. The results? Absolutely eye-opening.
The Starting Point: A Simple Fitness App as Test Project
My goal was simple: Develop a personal fitness app that quickly suggests the right stretching or workout exercises for users. Nothing complicated, but functional enough to test the strengths and weaknesses of both AI systems.
As described in my last blog post about Meta-Prompting, I started with a structured prompt from my Prompt Expert GPT. This optimized prompt then became the foundation for both tests.
Watch me build a complete fitness app using nothing but prompts in two different AI tools:
Test 1: Claude 3.7 in Windsurf – High Expectations, Disappointing Reality
My first choice was actually Claude Opus 4 (also known as Sonnet 4) – one of the most advanced models for programming tasks. But here came the first problem: To use Claude Opus 4 in Windsurf, I would have had to install a separate API and set up a paid credit system. As a student, I wanted to avoid this additional effort and cost, which is why I had to settle for Claude Sonnet 3.7 available in Windsurf.
The result was more than sobering:
- Immediate error messages on the first attempt
- The system was noticeably slow and sluggish
- Multiple iterations led to no functional code
- Debugging became an endless test of patience
- Bugs emerged that I couldn’t fix myself as a non-programmer
- Even simple UI elements caused unexpected problems
The stark limitation of Claude 3.7 quickly became apparent – a painful reminder of how large the quality differences between various AI models can be. What was particularly frustrating: As someone without deep programming knowledge, I was completely helpless once technical bugs occurred that went beyond simple syntax errors.
Test 2: Google Gemini 2.5 Beta – The Surprise Winner
The switch to Google Gemini was like night and day. Within seconds, I had a functional prototype – without error messages, without endless debugging sessions.
Gemini’s advantages in this test:
- Immediate functionality: The first generated code ran without problems
- Easy iteration: Prompt refinements were seamlessly implemented
- Smooth integration: Embedding videos and testing UI flows worked with just a few clicks
- Flexibility: The generated code could be easily exported to other tools
An Important Note: The Future Could Look Different
Crucial to mention: As soon as Claude Opus 4 and Sonnet 4 are also available in Windsurf without additional API installation, this evaluation will change dramatically. Claude with its new Sonnet 4 and Opus 4 models significantly outperforms Google Gemini in capabilities for long tasks and programming. Claude Opus 4 achieves 72.5% on the SWE-bench benchmark, while Gemini 2.5 Pro stands at 63.8% – that’s a significant difference of almost 9 percentage points.
With parallel test-time computation, Claude Opus 4 even rises to an impressive 79.4% on SWE-bench, showing what potential lies in these models. Claude Sonnet 4 even achieves 72.7% on SWE-bench, surpassing both Opus 4 and all competitors.
My test was therefore mainly limited by the technical and financial hurdles in accessing the latest Claude models – not by their actual performance capability.
Sources:
- Anthropic: “Introducing Claude 4” – Official Benchmark Data
- Beebom: “Anthropic’s Claude Opus 4 and Sonnet 4 Set a New Benchmark in AI Coding”
- Entelligence: “Claude 4 vs Gemini 2.5 Pro: Complete AI Model Comparison 2025”
🏆 SWE-bench Benchmark Comparison 2025
Programming & Software Engineering Performance
What This Experiment Reveals About the Future of App Development
This experience shows two important trends:
1. Model Quality Makes the Difference
Not every AI is equally suitable for every task. While Claude Opus 4 might have performed differently, Claude 3.7 was clearly under-equipped for more complex coding tasks.
2. We’re at the Threshold of a New Era
Creating functional apps in minutes? That’s no longer science fiction, but reality. Tools like Gemini, Claude Opus 4, and emerging approaches like Vibe Coding are fundamentally changing how we develop software.
The Practical Takeaway for Your Own Projects
If you want to quickly create prototypes:
- Use optimized prompts: My Meta-Prompt template from the last blog post makes the difference here
- Choose the right tool: For rapid prototyping, Gemini clearly has the edge in this test – but only as long as Claude Opus 4 isn’t easily accessible
- Stay flexible: Transferring code between different platforms expands your possibilities
- Pay attention to accessibility: The best model is useless if API costs or complicated installations get in the way
Conclusion: A Snapshot of the Current AI Landscape
While Google Gemini clearly won in this direct comparison, my test shows above all one thing: Accessibility is at least as important as pure performance. The technical hurdles and API costs for Claude Opus 4 make an objectively superior model practically unusable for quick experiments.
This is an important reminder for the entire AI industry: The best models are worthless if they’re not easy to use. Google scores here with direct availability in AI Studio, while Anthropic, despite superior benchmark values, is slowed down by access restrictions.
The ranking could change quickly: As soon as Claude Opus 4 is integrated into tools like Windsurf, the tables are likely to turn. With 72.5% vs. 63.8% on SWE-bench, Claude is clearly ahead – but only in theory.
What do you think? Have you already had your own experiences with AI-based app development? Which tools worked best for you? And how important is accessibility to you compared to pure performance?
Next Steps: If you want to learn more about advanced prompting techniques, Vibe Coding, and the future of AI development, check out my other blog posts – there’s much more to explore!
- 🚀 From Spark to Flame: How to Use AI to Turn Your Idea into Something Bigger
Learn how to expand and develop your ideas using AI. In this article, you’ll discover techniques for idea amplification, the Socratic method for concept development, and a step-by-step guide to transform initial thoughts into comprehensive projects. Read article → - 🧠 Beyond Basic Prompts – The RISEN Method for Smarter Learning and Faster Research (Recommended)
Discover how to optimize your research process with specialized AI tools. This article introduces my RISEN framework (Role, Instructions, Steps, End-Goal, Narrowing) and shows how to combine Perplexity with NotebookLM for more effective research workflows. Read article → - 💬 Prompt Like a Pro: How to Ask AI the Right Questions (Recommended)
Master the fundamentals of effective prompting with proven frameworks. You’ll learn three core frameworks (RTF, RODES, RISEN), everyday prompt patterns for common tasks, and advanced techniques like Chain of Thought and self-reflection. Read article → - 🔍 Is Perplexity the New Google?
Explore how this AI-powered search tool is transforming research methods. This article details Perplexity’s features, shows how to use Deep Search effectively, and provides a step-by-step guide for researchers, students, and professionals to enhance their information gathering. Read article → - 🛠️ 5 AI Tools You Must Know in 2025
Get familiar with the most essential AI tools for productivity and creativity. This guide reviews the top five AI assistants (ChatGPT, Claude, Perplexity, DeepSeek, and Qwen), explaining their unique strengths and how to combine them in an optimal workflow. Read article →
