<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Denisa Ionita on Medium]]></title>
        <description><![CDATA[Stories by Denisa Ionita on Medium]]></description>
        <link>https://medium.com/@denisa.ionita?source=rss-3a4e2f622fb2------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*w6dUdbYj2Zpa_N4b</url>
            <title>Stories by Denisa Ionita on Medium</title>
            <link>https://medium.com/@denisa.ionita?source=rss-3a4e2f622fb2------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 04:17:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@denisa.ionita/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Vibe Coding with AI: A programmer's headache]]></title>
            <link>https://medium.com/@denisa.ionita/vibe-coding-with-ai-a-programmers-headache-b684865d91c4?source=rss-3a4e2f622fb2------2</link>
            <guid isPermaLink="false">https://medium.com/p/b684865d91c4</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[mobile]]></category>
            <category><![CDATA[swift]]></category>
            <dc:creator><![CDATA[Denisa Ionita]]></dc:creator>
            <pubDate>Fri, 03 Apr 2026 09:42:58 GMT</pubDate>
            <atom:updated>2026-04-06T08:53:42.869Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Vibe Coding with AI: A programmer’s headache</strong></h3><p><em>A practical look at AI-assisted prototyping — will it take my job or complicate it?</em></p><p>As a Senior iOS developer for almost 9 years, most of my experience comes from building structured, maintainable apps — thinking in layers, patterns like MVVM, and writing code that’s meant to last.</p><p>Recently, I tried something very different.</p><p>I’ve been experimenting with what the internet calls <strong>“vibe coding”</strong> — building software quickly with AI assistance, focusing more on momentum and exploration than structured, slow paced code.</p><p>In a small prototype project, my team and I leaned heavily on <strong>GitHub Copilot</strong>, especially its agent capabilities, to see how far AI-assisted development can go in practice.</p><p>The goal wasn’t to build production-ready software — it was to move fast, explore ideas, and understand this new workflow.</p><p>The result?</p><p>Useful, surprising… and much more nuanced than your usual <em>“AI will replace developers”</em> narrative.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b795ZPB8dF4ED100chX10g.png" /><figcaption>Creative workspace with travel app prototype</figcaption></figure><h3>What is “vibe coding” — and where did it come from?</h3><p>“Vibe coding” isn’t a formal methodology, and most developers don’t know or approve of it because it was made popular by juniors and non-technical people at first. But it’s been recently gaining more and more momentum (at least on my LinkedIn and Instagram feeds).</p><p>It emerged from the recent wave of AI-assisted development — especially with tools like Copilot and conversational models — where users, both technical and non-technical started asking for:</p><ul><li>speed over structure</li><li>exploration over planning</li><li>iteration over perfection</li></ul><p>Instead of designing everything upfront, you’ll:</p><ul><li>start building immediately</li><li>guide the AI</li><li>adjust as you go</li></ul><p>It’s less about writing code line by line the traditional way, and more about <strong>steering the system towards your goal</strong>.</p><p>Important distinction:</p><p>🚩 This pratice is not “AI replacing development”<br>🏁 It’s actually <strong>AI accelerating exploration</strong></p><p><strong>*</strong>Of course you can use it as an even more advanced tool, with stricter rules, steps (or <a href="https://www.bluebag.ai/blog/what-are-agent-skills">skills</a>) defined to have AI plan and be much more specific and not <em>just execute</em>. This would fit requirements of a more mature project. <br>But that is a discussion for another time, we’re talking <strong>vibe</strong> coding here, so let’s continue.</p><h3>GitHub Copilot vs other AI tools</h3><p>If you’re new to this space, it’s worth clarifying where Copilot sits.</p><p><strong>GitHub Copilot (execution speed)</strong></p><ul><li>deeply integrated in IDEs (Xcode, VS Code, etc.)</li><li>great for inline suggestions and code generation</li><li>can expose a variety of agents depending on your subscription (GPT, Claude, Gemini..)</li><li>now includes Agent mode: with the right permissions, it can generate new files and change your code directly</li><li>also has a Ask mode, which is similar to what I’m describing below:</li></ul><p><strong>Claude / ChatGPT-style tools (brainstorming partner)</strong></p><ul><li>stronger reasoning and explanations</li><li>better for architecture discussions</li><li>not always directly integrated in your codebase (but with Copilot you can link files for context)</li></ul><p>*Of course there is <em>Claude Code</em>, but I have not explored that, I went with the Xcode integration of Github Copilot for ease of use. Especially because it offered me both a <em>chat</em> and an <em>agent</em>, in the same interface — that was a huge benefit.</p><p>If you’d like to read more on the tools:</p><p>GitHub Copilot docs: <a href="https://docs.github.com/en/copilot">https://docs.github.com/en/copilot</a><br>Anthropic Claude: <a href="https://www.anthropic.com/">https://www.anthropic.com</a></p><h3>Agent vs Agentic AI — what’s the difference?</h3><p>It’s important we make this distinction, trust me, I was also a tad confused.</p><h3>Agent</h3><p>A system that performs tasks based on instructions. <br>It needs you to provide it specific context or files to look at, and normally works in isolation and on small, specific tasks.</p><p>Example:</p><blockquote><em>“Generate a SwiftUI view”</em></blockquote><h3>Agentic AI</h3><p>A system that:</p><ul><li>takes initiative</li><li>chains actions</li><li>modifies multiple parts of your codebase</li><li>tries to solve problems end-to-end</li></ul><p>Example:</p><blockquote><em>“Refactor files, update navigation, fix dependencies” — they all happen in one flow</em></blockquote><p>Copilot’s newer capabilities lean into this <strong>agentic behaviour</strong>.<br>And this is exactly where vibe coding becomes powerful — but also <strong>risky</strong>.</p><p>You know what they (Voltaire and Spiderman) say:</p><blockquote><strong>“With great power, comes great responsibility.”</strong></blockquote><h3>My team setup</h3><p>We kept things intentionally simple:</p><ul><li>small team</li><li>prototype-level app</li><li>heavy AI usage (Copilot + conversational prompting)</li></ul><p>The approach we took:</p><ul><li>started building features early, without defining full architecture upfront</li><li>iterated continuously (refactoring and adjusting as needed)</li><li>allowed the architecture to emerge from implementation rather than design</li></ul><figure><img alt="programmer working with computers and an abstract representation of AI as a robot" src="https://cdn-images-1.medium.com/max/1024/1*hn7UEAa2CLyktTrrQNzlYA.png" /><figcaption>ChatGPT’s generated image of a ‘blonde programmer doing vibe coding’</figcaption></figure><h3>What worked surprisingly well</h3><h3>1. Fast scaffolding (foundation)</h3><p>AI was extremely effective at generating the skeletons of the App:</p><ul><li>data models</li><li>mock data</li><li>basic MVVM structures</li><li>screens and views</li></ul><p>It gave me a solid starting point in seconds.</p><h3>2. Exploring ideas quickly</h3><p>Beyond code generation, AI helped with:</p><ul><li>brainstorming features</li><li>iterating on ideas</li><li>deciding what to build next</li><li>refining my work based on stakeholder feedback</li></ul><p>It acted like a <strong>thinking partner + executor </strong>who didn’t have strong opinion. It didn’t once try to convince VIPER is a good architecture.</p><h3>3. SwiftUI prototyping</h3><p>With <strong>SwiftUI</strong>, this was especially noticeable, and helpful in areas I had not yet explored myself.</p><p>Describing a UI often produced:</p><ul><li>decent layout</li><li>working structure</li><li>something I could refine in just a prompt</li></ul><p>Not perfect — but fast.</p><h3>4. Reduced context switching</h3><p>Example prompt:</p><blockquote><em>“Add tab navigation, use consistent styling, update accent color, unify buttons”</em></blockquote><p>And it actually did it.</p><p>No docs. No searching. Just iteration.</p><p>*I had worked with tab navigation on UIKit, because like most apps out there, unless it’s really new or refactored, it will have some UIKit. Well that meant I had not written any SwiftUI code on navigation topics until now. Researching that and doing it the proper way (i like to check various sources, I’m a nitpick) would’ve taken me the whole day.</p><p>Copilot changed the app’s navigation from one screen to multiple tabs, in seconds! I just checked and learned from it</p><h3>Where things started breaking — Woman vs AI</h3><h3>1. Architecture still needs you</h3><p>AI writes code.<br>It does not design systems.</p><p>Deciding:</p><ul><li>boundaries</li><li>data flow</li><li>responsibilities</li></ul><p>🧠 still 100% human work. <br>But at least I don’t have to convince it of my personally favorite flavor of code architecture.</p><h3>2. Losing context</h3><p>Longer workflows exposed a major issue:</p><ul><li>it forgot constraints</li><li>ignored patterns (MVVM, CLEAN)</li><li>generated massive files</li></ul><h3>3. Confident but wrong code</h3><p>Classic AI behaviour:</p><ul><li>deprecated APIs</li><li>invented methods</li><li>“looks right” code that actually breaks</li><li>100 lines that do what could also be done in 5 lines</li></ul><p>❗ Rule: <strong>assume nothing is correct by default<br></strong>Think of it as working with an <em>enthusiastic junior</em>. <br>Yes I said <em>junior</em>, I wish tech companies would hire more of them.</p><p>Because us seniors will retire at some point, who’s gonna tell the AI wrong then??</p><h3>4. Duplication</h3><p>Even when told to reuse code:</p><ul><li>it recreated components</li><li>duplicated logic</li></ul><p>Left unchecked → <strong>tech debt explosion<br></strong>Good luck convincing your Product Manager to prioritize that against some fancy new feature.</p><h3>5. The “agent went rogue” moment</h3><p>At one point, I let the agent handle a refactor + rename.</p><p>Result:</p><p>😵‍💫 corrupted project<br>😵‍💫 broken .xcodeproj<br>😵‍💫 unrecoverable state</p><p>My mistake was not having a git backup. It was the first days of the project and I was still new and ✨dazzled✨ by the vibe coding mirage.</p><p>So I had to rebuild the project, manually recover files. I was able to use what was previously generated by my careful prompts and instructions. <br>But did I waste a day insisting the agent fix the rename? Yes I did.</p><p>I learned my lesson the hard way:</p><p>✨ <strong>Never trust the agent unsupervised on structural changes</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C_wDYinldM6JKLrIV8qXcg.png" /><figcaption>ChatGPT’s generated image of a ‘blonde programmer doing vibe coding and being frustrated about it’</figcaption></figure><h3>Lessons learned</h3><p>After this experiment, a few things became very clear:</p><ul><li>AI is excellent for <strong>momentum and exploration</strong></li><li><strong>Clear prompts</strong> drastically improve results</li><li><em>Architecture</em> cannot be outsourced</li><li><strong>Refactoring</strong> is inevitable</li><li>AI improves with <strong>feedback</strong> — but slowly.</li><li>You need to get past the initial frustration and explain what it need to do <em>differently</em></li><li>Git is not optional (branches = less frustration)</li></ul><h3>Final thoughts</h3><p>AI-assisted development doesn’t remove engineering thinking.</p><p>But it <strong>changes where your effort goes</strong>: less time writing boilerplate, more time guiding, validating, and correcting. And exploring of course, that’s the most fun part!</p><p>Used well, tools like Copilot are not magic — but they are a serious <strong>force multiplier</strong>.</p><p>Of course, this was just an initial exploration into ‘vibe coding’. <br>There is a wider discussion to be had about using tools like GitHub Copilot or similar, for day to day developer work.</p><p>On a mature project, you would need a much stricter set of rules, TDD and architecture guidelines, perhaps even <a href="https://www.bluebag.ai/blog/what-are-agent-skills">skills</a> and clear plans to have a productive collaboration with your 🦆AI rubber ducky.</p><h3>A better question than “Will AI replace developers?”</h3><p>What you need to ask yourself instead:</p><blockquote><em>If AI can already write a large part of your code…<br> </em><strong><em>what becomes your real value as a developer?</em></strong></blockquote><p>And more importantly:</p><blockquote><em>Are you </em>actively<em> learning how to work </em>with<em> it — <br> or are you still trying to outcode it?</em></blockquote><p><em>Listen, I know you were here first, I get it. But it’s worth trying things before dismissing them — just don’t switch off your critical thinking.</em></p><p>If you’ve experimented with AI-assisted workflows, I’m genuinely curious:</p><p>🚀 What worked for you — and what completely broke? <br>Happy to discuss in the comments!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b684865d91c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>