How code comes about that can still be maintained a year later. I write about planning, testing, working with AI while developing, and the decisions we forget quickly once something simply works.
Key questions
How do I use AI in development?
Most of my time can be lost before the first line of code. If I tell a tool "improve the user experience", I leave too much room for interpretation: it can change colours, add animations or rearrange the menu. So I first describe what is causing the user a problem and only then ask for a solution. Clear instructions are not just a longer prompt. My thinking is in AI saves me a lot of time in development. I still don't trust it blindly.
A homepage that opens doesn't mean login or image upload work. After a release I'd check the public addresses and the tasks the change touched. A case from building this site: a problem with access to public images only appeared after the transfer, and checking the actual address found it, not the report of completed steps.
What should tests check?
What actually worries me, not what is easy to write. Code review and user experience are also two different views: the code can be correct and the use still awkward.
Is a good prompt enough for a good result?
Not for a project that will outlive one conversation. A prompt solves a moment, a specification solves a project: a short brief, rules I don't repeat in every conversation and acceptance criteria. More in A good prompt is not a good specification.
Speed is not correctness. Code is quicker to get than its consequences are to understand, and a tool's explanation of its own code is not independent proof that it works. When hunting bugs I want evidence, not just plausible explanations, and I treat the release to production as its own step, not as a continuation of programming. Documentation should describe the state, not the intention.
In practice
This site is a server-rendered PHP application with native CSS and a little JavaScript; in production it runs on MySQL, and SQLite is only for local tests. At the last check on 4 October 2026 it has 23 unit checks and 163 checks of routes, login and security that run before a release. The release itself goes through a script that first saves a copy of the files and the database and stops if any file on the server differs from what is expected.
For FitBlizu, a map of gyms in Slovenia and Croatia, the principle on the project's about page is simple: when a fact isn't known, it's better to leave it out than to guess.
Decisions on a gym map that mattered more than the technology: an empty field instead of a wrong one, a source and date on every value, editing without an account, independence.
AI-assisted development talks a lot about prompts. I care about what we're building, by which rules, and how we'll know it's right. Brief, architecture, acceptance criteria, tests and production checks.
What a website has to tell someone who does not know the business yet, and why that is not mainly a question of design. I write about the decisions that come before the first line of code, about reviewing existing sites, and about when a redesign makes sense and when it does not.
Search is no longer just Google. I am interested in how people find information today, what of classic SEO still holds, and what AI answers change for a website. No ranking promises, verified sources.
I use AI every day and I do not trust it blindly. Here I note where it genuinely saves work, where the result has to be checked, and which promises do not survive the first enthusiasm. Experiments, not predictions.
When does a new feature improve an application and when does it just complicate it? Drawing on my own products such as Bizko, I write about the path from an everyday problem, like entering the same data twice, to a tool people actually use.
Technology is the beginning; business gives it meaning. I am interested in what a digital solution really brings a company, which processes are worth automating, and when a simple path beats a big system.