A month ago, I wrote I am a rational pessimist about AI. Yesterday's dispute with my co-worker about unnecessary complexity of AI-infused Korn shell script delivered by our other colleague reminded me there are also radical pessimists among us. However, every day when I dare to go further into the uncertain wilderness of imperfect prompts followed by indeterministic and not-always-right responses, I see progress, new opportunities and new ways of doing better.
Today, I have another success story to share.
It started one evening after a workday recently. I came back from my daily dev job and went running to compensate for hours sitting in front of a screen. After the jogging session I was in a shower and suddenly got a cool idea for my side project.
I thought it would be nice to provide to my future users a convenient way of reporting bugs. But also to make it convenient for me. I would be happiest to have reports appearing as GitHub issues. But teaching BFUs to login to GitHub and creating new issues in a consistent way? Good luck with that. So what if I just provide a simple form on my website and wire it to a backend service that will collect the input and create the issue automatically?
Sounded like a plan. But imagine all the work you need to do. Putting together the form, writing the backend service, figuring out how the GitHub API works exactly to be able to call it... Let's just put the idea into the inner memory deep in the back of my brain and we will surely look at it. One day. Not today... My inner demon of procrastination nearly won again.
But times have changed lately. Now I am armed and dangerous with tireless AI companions. Despite all those legit complaints about coding agents not doing their job as good as you would hope and not nearly as great as AI-driven companies keep promising you, in early 2026 your favorite excuses "I don't know how to start" or "This task is too big to start it today" are no longer valid. You can always just kick off with a simple natural language prompt and see how it goes. Even the last season's truth that the prompt needs to be well designed (and thus you "cannot" start as "I don't know how to write a correct prompt") is fading out.
Before going further, I'd like to make a statement. This doesn't mean I am encouraging you to just fire ambiguous prompt and YOLO-deploy the first iteration right into the production. The AI Manifesto is still valid and you should always verify the results, reason about code and understand everything before moving on. It is easier for me as a senior dev, but if you are a beginner, the more important it is before you get lost and become a hostage of a machine mind. On the other hand, you shouldn't hold yourself back too much. AI is there for you. Use it. Benefit from it. 💪🦾
Before I finished bathing, ate dinner and socialized with my family, it was already getting quite late. But I was feeling lucky and wanted to try.
Here is the prompt for my (paid) Copilot I quickly put together:
I want a new feature in this Nuxt 4 project. I need a new client page
"/report" with a Nuxt UI form entitled "Report problem" that will include
text field for "gameLink", radio switching between "report false positive"
and "report false negative", text field for "issue" and text area
"description" for describing the issue to be reported. All fields will be
mandatory. The filled form will be send to backend API route where it is
validated and then a GitHub issue in <<REPO>> repository is created from
given contents. Issue will be named by "issue" field and contain both
"gameLink" and "description". Selected radio value will result into
corresponding label (that will be prepared in the repositry). The report
form is annonymous, there will be dedicated service account with access
token to create the issues configurable via Nuxt runtime config (server-
side, so NOT public).
Just this. No context engineering, no MCP servers, no skills or whatever. Just my codebase, internet connection and the rough task description in natural language with a few instructions and restrictions. Typos and grammar mistakes included.
Maybe one thing. If you are AI-geek, you may have noticed new model Claude Opus 4.6 was released recently. I did notice and I also decided this was a perfect opportunity to test it in action. And I wasn't disappointed.
I spent two, maybe three minutes writing the prompt (while already have it more or less ready in my mind). Copilot with Claude took about 4 minutes to deliver the result. The surprising part was - the code was nearly perfect!
I was trying, but I was unable to find any significant flaws. It created decent frontend form, it added Zod schema validation, it created backend endpoint, handled and validated the incoming data, put together the GitHub API call and wired the response back into the client. It even noticed there was no <NuxtPage /> yet and made adjustments accordingly to support new /report page along the existing (implicit) /index. Yea, I might do some refactoring and polishing later, because looking back, the code might be more compact and less verbose. But overall it was not glued together somehow until it accidentally worked. It was a decent job. I don't think I would do much better for the first iteration.
My victory wasn't so flawless, because then I spent at least 30 more minutes configuring my new GitHub service account and getting the proper access token (AI-backed again). But overall it was less than an hour from the first keystroke to test issue successfully created in my repository. Wow! 👀
There were a few success points that helped:
- I knew what and how I want to do and I know Nuxt good enough to request general architecture with at least basic level of security (i.e. token must be kept on server-side and not exposed as public config).
- Copilot wasn't building from scratch but inside already established repository with configured Nuxt UI and existing server routes to copy from.
- People from Nuxt ecosystem are working hard to create AI-friendly solutions so it is much easier for agents to get relevant up-to-date information.
I was hoping for good outcome, but this really amazed me. It reminded me that it already me who is being the bottleneck. My hesitation, my reluctance and my incompetence to delegate tasks that should be delegated without losing control. Despite I am still not scared about my very existence as some fellow devs seem to be, I was re-assured I should know much better. And I will keep trying to do so.
What about you? What are your recent experiences with AI? Feel free to share your stories, questions, objections and concerns in the comments below 👇
Top comments (0)