A New Era of Coding

It’s funny looking back at the times when people got hyped over Copilot and hitting Tab. Honestly, I’m not even sure when AI stormed into the industry. I didn’t even blink before my job changed 180 degrees. About a year ago, after hours, I started testing the first editors with agent mode. Thanks to some awesome folks at work, I quickly got a license and the green light to play around. I went with Cursor, because it was the best at the time. Is it still? Hard to say. I’m starting to lose track of what’s going on with all the new tools and tech. 

The beginnings were as exciting as they were frustrating. Total hallucinations one day, then a bomb of a solution the next. Gradually, I started generating more and more tasks through AI until one day I made the bold call to stop coding completely. It’s been half a year since that experiment started. I had to learn how to work with it efficiently. At first, it was rough because most of what we now take for granted didn’t exist yet. You had to manually generate action plans, load key info into context, paste documentation, approve command execution. But things evolved week by week. The pace of updates and improvements was insane. It felt like a textbook agile team over at Cursor. 

Looking back, I think you really need to burn a lot of hours experimenting to get a grip on how AI behaves in different scenarios. The number of variables is massive. Agent modes, project complexity and structure, language model, well-written rules, prompt precision, task complexity. There’s no golden recipe, but here are a few takeaways. 

What Works

✅ Small, specialized prompts. They’re easier to verify and leave less room for garbage changes. 

✅ Clean, well-structured projects. AI tends to adapt to the house rules on its own in this kind of environment. 

✅ Solid rules and documentation properly loaded into context. That limits AI’s wiggle room and keeps stupid stuff to a minimum. Keep all that in the repo. 

✅ Iterative work. Throw a prompt, check changes, tweak, repeat until you hit the mark. When you want to move on to the next task, open a new chat. That goes without saying. 

What Doesn’t

❌ No .NET debugging support. Total embarrassment. 

❌ Auto mode. It feels like it’s running some model from a decade ago. 

❌ Trying to generate big tasks in one prompt, even with great docs and a top model. That’s a mess and hallucinations galore. 

❌ Git commands in Auto-Run Mode. Sometimes it makes a huge mess through commits, cherry-picks, or whatever else. 

❌ Project starters. Even with popular JS libraries, I’ve never managed to generate one faster than I could do it myself. 

Killer Features

🚀 MCP for Playwright and Figma. You just say what view to open or click through and boom, it generates E2E tests like a bomb. For Figma, you run the app locally, copy the frame ID, paste it into the prompt, and the view is done. Literally. Few things feel as magical as these two tools. 

🔍 Analysis and discovery. New repo, new task? A few questions and explanations later and you have a solid grasp of what’s where and how. It’s best to ask for references to specific files. 

🧩 Multi-repo workflow. Need to change something across several layers? Add all repos to one workspace and take care of the whole task. 

🧠 Critique. Asking for reviews, refactors, or vulnerability checks is where AI really shines. 

📄 Documentation generation. Nobody likes doing that. Just generate it and commit it, then thank yourself later. 

How Much Does It Cost?

Using the latest standard model (currently Sonnet 4.5) and skipping deep thinking or max mode keeps me within about $10 extra above the base plan. Using max mode and the best models costs a few dozen cents per request. 

What Do I Think?

I can’t imagine working without a tool like this anymore. In the right hands, AI supports every stage of development: analyzing, discovering, suggesting, planning, implementing, testing, and evaluating. You don’t have to waste your brain on pointless research and repetitive changes. You can stay on a higher level of abstraction and make the key decisions. 

Also, you can go beyond your usual skills quickly. Lately, I’ve been writing fully functional pipelines, even though my skills in that area are basically zero. 

The jokes about “vibe coding” and the “how long did that take you?” questions, I hear those all the time. But honestly, if someone codes faster manually than AI, it’s usually either because they don’t know how to work with an agent or they are a pure genius. In my experience, work moves faster and is more qualitative. Still, there are traps. Sometimes laziness makes you miss some dumb mistakes or get stuck without the right strategy. But those aren’t tool problems, they’re about how we work and learn to use it. 

In Cursor I trust 🚀