János the vi... prompt engineer

Monday, April 13, 2026

Is this what it means to be a software developer in 2026? Basically I’m an AI coding agent herder telling them what to do and checking their output for hallucinations? Maybe. And I’m not even complaining!

Let’s see the history of János’ interactions with AI, or, to be more precise: LLMs and coding agents.

I’m not sure exactly when I signed up for GitHub Copilot but I got an email on the 17th of August 2022 that my trial is about to end. The 5th of September was a $0 free plan setup and the first real payment came on the 8th of September. I remember that I haven’t used it during the trial and I got charged for it which gave me the motivation to actually start testing it out. While I was doing that I thought I’d give the competition a chance so on the 18th of September 2022 I signed up for the Tabnine trial and I somehow forgot to cancel that too. After the 14-day trial ended, on the 2nd of October 2022, I was charged for a $144 yearly subscription. Panic! I contacted them that same day and got refunded the next day. Tabnine felt inferior to Copilot anyway, so I stayed with the latter for a good while - actually I still have the subscription, I just haven’t been using it lately, you’ll see later why. So, I started getting smart code completions in the early days of 2022’s autumn…

Then came the end of 2022 and the beginning of 2023 when ChatGPT took the world by storm. I remember writing prompts and marveling at the output and soon enough noticing Stack Overflow having a banner which instructed their users that they shouldn’t post ChatGPT generated answers. That was the beginning of the end for Stack Overflow, which is a bit unfair since I’m pretty sure all of the AI models were trained on their content. On the other hand, I can recall my frustration not being able to use SO to ask something like: “I have X and Y and want to accomplish Z. What is the best approach? What are the best libraries?” The question would be locked in a short time for being “subjective”. So I was very happy as soon as I could converse about my questions with LLMs that would actually give me answers.

A bit later in 2023 I got into the GitHub Copilot Chat private beta on the 7th of July. It seemed out of this world: I wasn’t merely getting code completions, I could ask natural language questions and the LLM could respond having the context of my code. I didn’t even need to copy relevant stuff over into ChatGPT anymore. Wowzers!

At the beginning of 2024 Google Gemini (finally) launched and in May 2024 I even got myself Gemini Advanced for a short while, which I used with more or less success to edit a Romanian article about a surprise find in my yard while doing some DIY work. It was funny that I only asked it to proof read that article and make small edits to avoid repetitions and other writing blunders I’m guilty of, but it somehow went off the rails and rewrote half of my post in Spanish. I was amused. The rest of the articles around here are 100% certified AI free though. Pinky promise!

Update: What I said above was true when I wrote it. Since then I’ve published four rapid fire articles (the latest four before this one) augmented with Claude. I think I’ll write a separate article about how my writing process for this blog changed, but until then, you can rest assured that what you’re reading was only augmented by AI - I still edit the shit out of the LLMs output manually.

Around the middle of December 2024 I added CodeRabbit to my Photonia open source project. CodeRabbit is an AI code reviewer that leaves comments on your pull requests. It was free for open source so it was a no-brainer to try it out. This was before agentic development was a thing for me, so I applied its suggestions manually. It occasionally stopped me working on a given feature entirely: procrastination (too lazy to apply the fixes) versus OCD (couldn’t ignore them). Funny thing is, at some point CodeRabbit added agent-ready prompts for each suggestion but I didn’t fully understand what those were, until…

A few months later, on May 26, 2025, the AI coding agents era started for myself also — Emil recommended OpenRouter with aider. So I signed up with OpenRouter, topped up my account with $10 and I was off! And by that I mean that I used it to add some features to the same Photonia. I was amazed that now I could just describe the feature I want in a prompt and it would plan it and write it, end to end. Browsing OpenRouter’s site I happened upon their rankings and saw that Kilo Code was consistently the top app. So, obviously, I had to try it out. I liked that it was integrated into Visual Studio Code (which is my editor of choice), that I could resume earlier sessions, that it had multiple modes for different tasks (Architect for planning, Coder for coding, Debugger for debugging), that I could write my own rules (instructions for the LLMs you don’t want to repeat each time), and that it showed me the cost of my session on the fly. I was sold.

Then I switched jobs in September 2025 and the new client encouraged the use of AI coding assistants. We had pretty frequent training sessions and I learned quite a few things which would have passed me by otherwise. One of the first trainings was about how to write better prompts, and soon after I discovered a page on the work Confluence listing recommended MCP servers. This was also the moment I started working with Claude, as the client was pushing for Claude Code.

I actually started drafting this article right around here, not long after the job switch, so the next two paragraphs (sort of a mid article conclusions) are a snapshot of where my head was at back then. Not everything in them is still true.

Up to this point I considered it to be just a helper. I used it for things that would take me ages on my own, or that aren’t my strong points, like UI work. Copilot autocomplete and chat helped a lot, but I was still writing each and every change by hand. And all of it was happening on one single personal project: Photonia. Even after I started playing with agents in May 2025 (OpenRouter + aider, then Kilo Code), the reflex was the same: let the agent draft something, then review every diff and tweak it by hand. Faster helpers, not self-directed collaborators.

Three kinds of work stood out as the places where I was willing to hand things over:

1. Stuff I could never finish in a reasonable amount of time on my own because it needed know-how I didn’t have and didn’t want to pick up just for one feature. The related tags feature in Photonia is one example: it suggests tags which commonly appear together on public photos. That’s straight data mining, and there was next to zero value in me learning association rule mining from scratch just to add that one feature.
2. Boilerplate repetitive stuff I kept procrastinating on.
3. Writing and adjusting unit tests, which is the exact thing I know I should always do and almost never want to.

Alright, back to the timeline. Later in the autumn, Kilo Code became the first tool where I started experimenting with MCP servers on my own (at work I was already using them through Claude Code, as mentioned above). MCP servers are basically small add-ons that let an AI agent reach external tools, APIs, and data beyond the code in your editor.

From this point onward the ball really started rolling, moments of amazement (“OMG, it can do that?”) were coming quite frequently. But it came at a price… I topped up my OpenRouter account again with $20 on October 21st and another $20 on November 29th - so it’s obvious that work was being done. Today there’s still $3.27 left in the account.

At one point in September or October I got a recommendation for z.ai from Andrei but I pretty much ignored it for a while. On the 6th of November 2025, Emil said that I should get myself a Claude Pro subscription. I did, but it quickly turned out that it wasn’t the wisest choice, because it was quite different from the API-based Claude Code I had at work: 5 hour limit, weekly limit, limits upon limits… Luckily they were running the beta test for their newly launched web-based Claude Code, and they gave me $250 of credit to use in the beta until November 18. So I tried doing as much as I could - I was writing prompts for features one after the other. Around those same days, I also rediscovered how valuable GitHub Copilot had become. It could route huge features through Claude models for just one premium credit per task, which was a lifesaver when I was hitting Claude Pro’s limits.

Also on the 6th of November 2025, I got a hint from, you guessed it, Emil for Perplexity, that I should associate it with PayPal and I got Perplexity Pro, which is a sort of Google on steroids. I use Perplexity for researching any type of tech acquisition I make (projector, graphics cards, astrophotography, etc). I like that it starts by searching the web for current information and crafts a response from the findings - other models and services usually start out relying on their knowledge (which might be outdated) and only search the web if you specifically instruct them. I also use Perplexity to write copy for websites I have by researching and synthesizing information from different websites. Also for doom scrolling when I check the news through it.

On the 7th of November, literally the day after, I finally took Andrei up on his recommendation and signed up for z.ai’s GLM Coding Lite plan at $9 (taking advantage of a first-month promo). Three weeks later, on the 28th of November, I upgraded to GLM Coding Pro via a $30-for-3-months promo, not because I hit the quota limit but because the Pro variant’s responses were noticeably faster.

In late November, I also started fanning out. A handful of projects had been sitting on my “would love to build this someday” list for years. Some were know-how gaps, others were lack of time. I finally felt I could start tackling them.

The first was a Jekyll media manager, scratching an itch from when I switched this blog from Wordpress to Jekyll in 2018. I even wrote back then that I was “badly missing” Wordpress’ media manager. This one fits in the lack-of-time category and AI helped: I hacked it together in a 5-day sprint between November 24th and 29th, drafting the app with the agent and iterating until I had something close to an MVP. I still haven’t released it, but I really should finish it up and start using it “in production”. And by that I mean on this blog.

Google Antigravity launched on the 18th of November 2025 and it got me intrigued. A couple of weeks later, once I was “done” with Jekyll media manager, I had to try it out, so on the 2nd of December, I sketched out an isometric city / colony management browser game that had popped into my mind at least 15 years prior. I only spent a few hours on it and what I have is really just a sketch that doesn’t even have a repo, but I was amazed I could sketch it at all. I never ever tried to program a game before so even if it was just a short brush it was really interesting.

Then in early December I (we?) started work on a news clusterer - it fetches articles from various sources and groups together pieces about the same story. Another idea I’d been sitting on for years because I had no idea how to cluster text. Then Kilo Code shipped a new feature that indexed your codebase into a qdrant database. I asked Perplexity what qdrant was, read in the first paragraph that it was a similarity search engine, so my second question was if it can be used to build a news clusterer and Perplexity said that it’s a very good fit for that. And the rest is history as they say. Most of the work on this project happened in December, intensifying between Christmas and New Year’s Eve, when I was off work and Claude was running a holiday promotion that doubled the usage limits on my Pro account. As I said above the normal Pro limits felt very constricting, but 2x was enough to get a lot of work done that week. I even built an Android TV app for it, because why not. Teletext 2.0, hehe. The project is still chugging along at stiri.link7.ro, gathering local (Maramureș related) news and clustering them nicely.

Somewhere in all this, on the 7th of December, I finally signed up for a proper Kilo Code account and topped up $20, and I immediately received a $20 bonus credit. The advantage over OpenRouter was that Kilo Code didn’t add tax on top, and they also offered free models (nowadays OpenRouter does too). That’s what kept me going with them for a few months.

At this point I was using Claude’s Opus model to plan the work on a certain feature or change, write a detailed plan to a file then switch to a cheaper model like z.ai’s GLM to actually implement the change.

Fast forward to February 2026 when I decided to revive Hai Afară! (“Come Outside!” in Romanian), an app to upload, edit, and showcase hiking and biking trails. I first worked on it back in 2018 and then abandoned it for one of the reasons I mentioned above: I didn’t have the know-how to finish it as it required GIS-related knowledge. For this one I upgraded to Claude Max 10x on the 2nd of February 2026, and I’ve been on it ever since.

Once I had Claude Max I didn’t really need z.ai anymore (except for the occasional agentic code review), so I handed the keys over to my wife. She had a blast with it and finally tackled some QA automation tasks. She even reimbursed me for the z.ai renewal on the 28th of February ($90).

It’s not just coding, either. I’ve used Claude Code to reorganize my home server (configuring and debugging Docker containers), to set up and secure my VPS, and lately for blogging help, including the article you’re reading right now. But I never got around to using autonomous agent wrappers like OpenClaw, which can run Claude Code sessions on their own, open PRs and resolve errors while you sleep. I don’t trust AI blindly enough for that.

A few things I’ve picked up along the way which are still valid today:

• It’s garbage in, garbage out, in a very specific sense: vague prompts will produce vague code. Writing better prompts is half the job now, maybe more.
• You really do have to stay in the loop as agents aren’t trustworthy enough yet to be left alone. Someone has to pump the brakes when they start going in circles or cheerfully reinforcing a bad pattern, and that someone is me, the human in the loop.
• Back when I was on API-based billing through OpenRouter and Kilo Code, there was a weird thing I kept running into: the providers reward fast follow-up responses because of prompt caching. Respond within the 5-minute cache window and you pay a fraction of the token price. I am not very good at that, so I was probably losing money. Part of why I switched to subscriptions…

For my ADD brain, all of this has been a welcome boost. So many new toys! Every week there’s a new MCP, a new model, a new workflow. As you can see, this was a pretty short period since I discovered agentic AI, but it was packed with changes. I’m not sure how sustainable it is for my attention span, but while it lasts I’m having a lot of fun.

Oh, also I’m pretty sure the financial AI bubble will burst, I just hope this doesn’t mean that they will take away my toys. I’m way too dependent at this point…

P.S. While writing this very article, we (as in me and my friend Claude) inaugurated a captain’s log — a blog in the blog in which we’ll chronicle all kinds of funny / interesting moments encountered during our editing adventures.