Six months ago, my entire experience with AI was typing prompts into ChatGPT’s chat window. Today, my agent builds entire blogs for me, configures VPS instances, and handles DNS — while I just say “yes” or “no.” Here’s how I got here.
Stage 1: ChatGPT — “answer my question”
Everyone starts the same way. Open a browser, type a question, get an answer. Magic. For the first few weeks, I used AI mostly as a Google replacement — ask a question, get an answer, close the tab.
I quickly realized ChatGPT wasn’t everything. Others appeared: Perplexity (better search), Google Gemini (free API), Anthropic Claude (better writing). All on free accounts, all through the browser.
Stage 2: Google AI Pro — “this actually costs money”
At some point, free account limits stopped being enough. Google AI Pro offered large packages and steep startup discounts. Over a few months I paid them about 100 PLN total, and thanks to a promotion I’m paying under 50 PLN/month through the end of the year. A fair price for a tool I use daily.
Stage 3: Antigravity IDE — “I don’t have to copy-paste code anymore”
This was the real breakthrough. Until then, my workflow was: ask AI for code → copy → paste into editor → something breaks → go back to chat → explain what’s wrong → get fixed code → copy → paste. On repeat.
Google Antigravity (an IDE with built-in AI) changed all of that. The AI has direct access to project files. You don’t copy-paste anything — the agent edits the code itself, sees compilation errors, understands the context of the entire project. This was my moment of enlightenment: I don’t have to do this manually.
I dove headfirst into projects. I built things that a month earlier were completely out of reach — not because I suddenly became a programmer, but because AI took on the entire technical layer. I was fascinated by the possibility itself. It felt like taking off chains.
Stage 4: Cursor — “one model isn’t enough”
I quickly discovered that Antigravity has limits too. When my tokens ran out in one tool, I’d switch to another. I subscribed to Cursor for $20/month and started working across both.
I also noticed something interesting: one model is not enough. Sometimes a given model (like Gemini in Antigravity) would get stuck in a loop — eating through an entire token limit on a problem it couldn’t solve. I’d switch to Claude in Cursor and the same problem would disappear in 5 minutes. And vice versa.
Today I use several models in parallel and switch between them depending on the task. There is no single “best” — each has its strengths and weaknesses.
Stage 5: Hermes Agent — “I don’t have to explain everything from scratch”
A month ago I discovered Hermes. And this is the next big leap.
Before, every time I returned to a project after a week away, I had to re-explain to AI what we were doing, where we were, what the context was. Hermes remembers. It has skills (my own procedures and knowledge), it has memory (what works, what doesn’t, my preferences), it sees my files and my terminal.
The result: we understand each other better every day. I no longer have to say “remember, my VPS is on OVH and the SSH key is here and there.” It knows. I don’t have to remind it “we use minimalism, don’t add unnecessary dependencies.” It has that in its skill file.
This isn’t a tool anymore. It’s a partner.
My current stack
- Hermes Agent — main partner, coordinates everything, remembers context
- DeepSeek API — paid API, replaced Cursor (cheaper, comparable quality)
- Google AI Pro — cheap Gemini API access (under 50 PLN/month)
- OpenRouter (free tiers) — access to many models at no extra cost
- Ollama (local, on my Legion PC) — offline models, zero cost
- My own VPS — hosting, SearXNG (search engine), dashboards
Along the way I also learned how to work around limitations: where to find free tiers, how to combine models, when to switch to another one to avoid getting stuck. This isn’t knowledge from a course — it’s the result of hundreds of hours of trial and error.
What I’ve learned
-
You don’t need to be a programmer. AI is a translation layer between your idea and the code. You say what you want, the agent does the rest.
-
One model is not enough. Different tasks need different models. Don’t be afraid to switch.
-
An agent with memory beats a chatbot. The difference between “write me a script” and “remember our project? add feature X” is vast.
-
Free tiers exist. You just need to know where to look. OpenRouter, Gemini, local models — you can operate almost for free.
-
Do it on the side. You don’t have to quit your job and “become an AI creator.” You can use AI as an addition to normal life — automate, accelerate, experiment. Like I do.
Six months ago: ChatGPT in a browser. Today: an agent that sets up blogs for me. Six months from now? I don’t know. But I’ll definitely write about it.