Wakapi shows that I have 20,000+ hours of total usage time in AI and dev tools (different tools at the same time). During this period, I would like to share the usage experience I gained with many AI models. The experience is based on coding skills; it doesn't cover daily conversation or topics outside of coding. I will explain with frontend terms for simplicity, but the verdict is general—the same situation applies to the entire dev process.
Key Tips
Unconditionally, always write every prompt in English!!! No discussion.
The fewer choices you give to AI models, the better the result. Especially writing what you want to achieve down to the smallest details is decisive. Example: Don't say "create a button", say "create a button using this component in this color and this size".
Adding the reason to the result significantly increases reasoning quality. (why you want to do it) Example: I want this button to be there because of this-this-this reason.
AI works better with more popular languages and frameworks. Example: Python > JS > Java > Go > Rust | React > Angular > Vue
Preventing spaghetti code is still very difficult, but in general the best strategy is working with microservices instead of working with a full app. Service+API. For AI there will be less confusion and also tokens will be used many times less.
MCP usage loses its rationality day by day, context7 especially creates a lot of confusion. If you're going to add some code somewhere, add the MD version of the repo directly with tools like repomix or let AI use its own web-fetch function.
Skills are very important but try to manually review each skill, don't use ready-made skills without modification. AI generally works much better with existing codebases than creating from 0, but absolutely compare the AI's last update date with the versions of tools you use.
AI models cannot implement advanced solutions very well. Try to use ready-made solutions by importing them. Example: Instead of writing a domlogger from 0 with AI, take information from some domlogger++ repo and integrate it into the project.
If AI cannot solve the problem 3 times, if you look at git history, if you try to solve it by looking at runtime—rewrite the prompt in a different form.
Don't use AI via API for dev!!!
Lifehack: If you constantly experience problems in medium or large projects, absolutely use AI tracing, read the reasoning or load the history and ask AI in the prompt what you did wrong (Langfuse).
Best of List
- Model - Opus-4.7
- Free model - Opencode (Big Pickle)
- Price/Quality - GLM-5.1
- Speed/Quality - Windsurf SWE-1.6 free (300tps) and SWE-1.6 fast (900tps)
- CLI - Opencode
- AI Editor - Windsurf, Cursor
- AI IDE - JetBrains
- VibeCoding - Emergent (especially E2, E3)
- Most affordable subscriptions - Opencode Zen $5, Windsurf Pro $20
- Search - Perplexity
- Realtime info - Grok
If you have additional questions, write them in the comments. :)