8-) Currently working on increasing limits to properly compare Gemini CLI and Claude Code. Will let you know about it!

7-) Final note: After using Claude Code, I can't imagine going back to Cursor. Gemini shows promise but the free tier constantly redirects to Flash despite advertising "pro usage." When using API keys, I hit tier limits immediately.

6-) This approach minimizes double logic and missing functionality problems that typically plague AI-assisted development. My subscriptions are all top-tier and I use Opus as much as possible, but even Opus limits drain quickly with this workflow. Sonnet limits too.

5-) Here's the trick: before deploying to production, export all failed Playwright tests to a JSON file using a custom reporter. Feed this back to Claude Code and ask it to fix all issues until zero tests fail.

4-) Generate Playwright tests for all UI components and buttons. This requires 2-5 iterations. Tests must be as detailed as possible - don't skip edge cases.

3-) Use Claude Code to eliminate mock data and build the actual backend. This takes 2-10 iterations depending on project complexity. Sometimes UI changes are needed. Always update your architecture md file when making changes.

2-) Take all Polymet prompts and feed them to o3 to generate a Technical Specification Document. Add this as architecture md to your project repo. Spend at least 30 minutes here. This document becomes your single source of truth.

1-) Start with Polymet (or similar) to extract user stories and UI specifications. I spend 2 hours here depending on complexity. This is critical - most cursor-based projects fail because the AI doesn't understand what to build. Without clear requirements, you're setting yourself up for failure

Çocuklar demi

@grok ?

I say “trying” because building a solid pipeline is genuinely hard.

Congrats!