How I Actually Use ChatGPT, Claude, and Gemini for Development Work

Every few months someone asks me which AI tool is "best." The answer depends entirely on what you're doing. I use all three, for different things, and none of them replaces thinking.

Here's my honest breakdown after a year of using them daily for development work.

What Each One Is Good For

ChatGPT: Best for open-ended coding tasks where I need a quick script or boilerplate. GPT-5's code output is solid for one-shot generation: "write me a Python script that does X." I reach for it when I know exactly what I want and just need it typed out.

Claude: Better at reasoning through messy problems. When I have a half-baked idea and need to talk through the design before writing code, Claude's longer context window means I can paste in the relevant files and get a coherent analysis. I use it for architecture discussions, code review, and refactoring.

Gemini: The Google ecosystem integration is genuinely useful. I use it when I need to pull real-time information: documentation lookups, API reference checks, current pricing. It reads web pages natively in a way the others still struggle with.

Where They All Fall Short

The same limitations across all three:

• They hallucinate APIs. Every model confidently invents function signatures. I've had ChatGPT write me Stripe code using methods that don't exist. Always check the docs.
• They flatten your thinking. If you ask for a solution, you get the most statistically common answer, not the best one for your specific situation. You have to prompt for alternatives explicitly.
• Context is still too short for real codebases. Even with 200K tokens, you can't fit a meaningful project. You're always cherry-picking files to include, which means the AI is missing half the picture.

How to Actually Get Good Results

The single most important skill is writing a good prompt. Not "prompt engineering" in the ceremonial sense: just being specific.

Instead of: "Write a login form" Try: "Write a React login form with email and password fields, using react-hook-form for validation, shadcn/ui components, and a POST to /api/auth/login. Handle loading, error, and success states."

The difference is night and day. The first prompt gets you a generic form. The second gets you something you can use.

What I Don't Use AI For

• Debugging race conditions. The models don't have a mental model of time or async execution. They guess, and they're often wrong.
• Security-sensitive code. I've seen AI suggest eval()-adjacent patterns too many times.
• Novel problems. If I'm doing something I haven't seen done before, the AI's training data won't help. It'll give me confident-sounding wrong answers.

Pick the tool that fits the task, keep your expectations realistic, and always verify the output. Same as working with a junior developer, except this one never gets tired of code review.

You Might Also Like

• Cursor IDE Advanced Features for Pragmatic Engineers: MCP, Rules, and Agents in 2026
• Fine-Tuning Llama 3 with LoRA and Unsloth: Developer Guide
• Running Local ai Models with Ollama: My Journey from Server Bills to Privacy-First Development
• vLLM vs Ollama: Local LLM Throughput & GPU Benchmarks

Frequently Asked Questions