<?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 Felipe Porge Xavier on Medium]]></title>
        <description><![CDATA[Stories by Felipe Porge Xavier on Medium]]></description>
        <link>https://medium.com/@felipefpx?source=rss-e3b75df1f1e5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Am6WO4L9zwchEIGuLYqGOw.jpeg</url>
            <title>Stories by Felipe Porge Xavier on Medium</title>
            <link>https://medium.com/@felipefpx?source=rss-e3b75df1f1e5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 02:24:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@felipefpx/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[AI superpowers and the overconfidence trap]]></title>
            <link>https://medium.com/@felipefpx/ai-superpowers-and-the-overconfidence-trap-9fb3c3356fe2?source=rss-e3b75df1f1e5------2</link>
            <guid isPermaLink="false">https://medium.com/p/9fb3c3356fe2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[engineering-leadership]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[decision-making]]></category>
            <dc:creator><![CDATA[Felipe Porge Xavier]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 12:43:33 GMT</pubDate>
            <atom:updated>2026-09-01T12:43:33.050Z</atom:updated>
            <content:encoded><![CDATA[<p>Prototyping has never been this easy. AI agents make everyone feel like experienced software engineers, and entrepreneurs are finally learning that creating an Uber-of-X or an e-commerce platform was never cheap.</p><p>The revolution in the software industry is real, but we are still in a transitional state. People are learning how to become multidisciplinary builders by shipping production releases, and that is great for building skills, but it makes review harder for experienced engineers.</p><p>The AI transformation dream reinforces the productivity promise, allowing small teams, branded as “pods”, to scale up by acting like Solvers instead of getting stuck in cross-chapter bureaucracies. In this structure, reactive people feel forced to contribute, because they become visible. But are they really prepared?</p><p>As I commented in my previous <a href="https://medium.com/@felipefpx/engineering-growth-in-the-age-of-ai-bc7feeba9678">article</a>, AI can make you stronger or weaker. People who are not prepared fall into the trap of “creating a productivity illusion”, trying to revolutionize the whole company by burning tokens on low-value initiatives, questionable analyses and misleading decision-making.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m7Q1S8sgv83OeFOdJAXFzQ.png" /><figcaption>Be careful, it can be an alligator, a log, or both. Image by <a href="https://pixabay.com/users/wolf69w-55093433/?utm_source=link-attribution&amp;utm_medium=referral&amp;utm_campaign=image&amp;utm_content=10187473">Joseph wolf</a> from <a href="https://pixabay.com//?utm_source=link-attribution&amp;utm_medium=referral&amp;utm_campaign=image&amp;utm_content=10187473">Pixabay</a>.</figcaption></figure><p>This era requires humility. You need to know your gaps, lean on your peers, and learn from them., instead of trying to live the illusion of trusting AI for decision-making. Confidence is a trap that can make you step on an alligator thinking it is a log.</p><p>As Eigner and Händler argue in “Determinants of LLM-assisted Decision-Making” (<a href="https://arxiv.org/html/2402.17385v1">link</a>), LLMs can fail depending on the interaction between them, the user and the decision context, and some hallucinations can sound really plausible. Try it yourself: ask Claude to make a decision, then share the assessment with GPT for review. Most times they will flag different problems. Don’t over-trust LLMs blindly. You are accountable for the impact of the decisions you ship.</p><p>These are tips for ensuring you’re in a good direction:</p><ul><li>Require the LLM to cite real sources, so you can verify the reasoning yourself.</li><li>Use multi-agent debate: have the model argue against itself from different roles to surface gaps.</li><li>Different models carry different biases; collect a mix and consolidate what matters for your problem.</li><li>If you work across disciplines, ask specialists to review what you’re proposing. You learn by doing.</li><li>Don’t underestimate the challenges. Define protective layers that let you fail safe and fast, so you can correct course sooner, saving time and mitigating risks.</li><li>Check the value behind what you’re proposing before prototyping; sometimes the cost of building isn’t worth it.</li></ul><p>Note that simple actions can protect you against frustration.</p><p>—</p><p>This text was written by a human and reviewed for grammar by an AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9fb3c3356fe2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Engineering Growth in the Age of AI]]></title>
            <link>https://medium.com/@felipefpx/engineering-growth-in-the-age-of-ai-bc7feeba9678?source=rss-e3b75df1f1e5------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc7feeba9678</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Felipe Porge Xavier]]></dc:creator>
            <pubDate>Fri, 19 Jun 2026 22:09:02 GMT</pubDate>
            <atom:updated>2026-06-19T22:50:27.810Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*oay9sk2z4wUV407MZFQwPA@2x.jpeg" /></figure><p>Over the last 10 years, I’ve been working on different development projects and teams, as part of companies of different sizes and types. The common aspect between them was the continuous need to ramp up Junior engineers to always have the next generation guaranteed.</p><p>For the companies, this process is painful due to the complexity of investing time and mentoring to prepare people, where some are faster learners and others might demand extra support and guidance. I saw countless Junior engineers evolving fast, absorbing more and more responsibilities while refining their soft and hard skills.</p><p>With the AI advent, an important concern arose: considering that experience is core to learning and to developing a critical vision of work quality,<strong> how do we ensure AI-first engineers are still building strong foundations if agents and tools are evolving faster and getting more and more autonomous at developing software?</strong></p><p>I’m not a guru or deep researcher, but I consider myself a reflective person who likes to think about what’s happening around me to ensure timely adaptability. So, I decided to write this article to share my perspective and some advice for Junior engineers, as a way to expand their horizons in this AI-based landscape.</p><h3>AI can make you stronger or weaker</h3><p>Let’s analyze the following analogy: are you considered a good driver because you have a driver’s license and a beautiful car, or because you’re skilled while driving and are able to not only review your actions in traffic but also distinguish safe and risky actions committed by other drivers around? Did you get the point?</p><p>Having unlimited tokens of the most powerful model available might be an F1 car for a real pilot or a kid to drive, and this depends on how prepared the person operating it is. Don’t be optimistic, imagining that the people around you won’t be able to classify you in between. You will continue being analyzed by the way you perform, not the AI.<strong> You need to dedicate yourself to understanding the things under the hood. </strong>In my experience this was always what boosted engineers’ growth in their careers, and prepared them to deal with more and more complexity.</p><h3>AI is fast for doing, but it depends on you</h3><p>With agentic coding, you’ll need to act like an engineering leader responsible for guiding others in the right direction. Driving complex projects will still have its challenges, because software engineering will always have some kind of subjectivity, mainly when architecting domains, where it can be hard to define the right direction to follow. You need to be the light at the end of the tunnel, looking beyond the current prompts/plans.</p><h3>Autonomy will be core in an AI era</h3><p>AI-first development enabled new organizational structures, in which teams get smaller and people can be multidisciplinary by experimenting with things and validating ideas through prototypes. It’s incredible to see how AI allowed non-technical people (from a software engineering perspective) to start experimenting; I saw countless websites developed by product people.</p><p>Learning what’s important in the business is also a differentiator, and it will become more and more important in this new world.</p><h3>Act like a curious child, not like an auto-accept bot</h3><p>Since the beginning, I’ve been incentivizing all the engineers I lead to use AI as a learning tool. You must act like a child trying to understand the world:</p><ul><li>Developer: Implement a function that calculates the sum of 4 numbers received in an array;</li><li>AI: done!</li><li>Developer: What’s reduce used for? Why not do it in a for loop with a variable? What other alternatives have you mapped, and what are the trade-offs between them?</li></ul><p>A great idea we have piloted recently was to use Socratic dialogue, or to ask the AI to do blind exercises with you, to practice critical thinking or other relevant skills. My last example was about “Situational Leadership”, where I asked AI to do an exercise and it not only created relevant practical exercises but also used the existing context to create realistic, hypothetical cases. This is a great example of how you or your team can continue evolving in this new age.</p><h3>Master AI techniques</h3><p>Mastering prompt and context engineering are core requirements in this era.</p><p>You need to understand the limits of an AI agent to avoid being excessively optimistic. An AI agent will do what you ask, and ambiguity can kill the results. You can ask AI to increase conversion as much as it can, and it might force all customers to buy a product on login using their saved credit cards – imagine the mess…</p><p>The top 2 prompt engineering strategies I use:</p><ul><li><strong>Few-shot</strong>: you give the AI some small examples; this is better than strictly defining formats and is usually enough for most cases. Imagine you asked the AI to tell whether a number is positive: it could answer with a huge block of text or just with true or false. Giving examples can help the model reason about the format you expect;</li><li><strong>Chain of verification</strong>: instead of directly accepting the first answer, use different AI personas in the same chat to refine your answers. For example, you ask the model to give you the top 3 alternatives to solve a problem; then you can tell the model to act as a business director who needs to map questions, gaps, and concerns behind the proposed alternatives. Then, it can be a Principal Engineer trying to answer/mitigate the points and propose better alternatives. Finally, you compile the final plan with a review, and your plan will surely be much better.</li></ul><p>Giving the right context to models is also a tricky part. If you give too much context — a crowded AGENTS.md with ambiguous or conflicting information — the results can be worse than running it clean. Make sure your rules are concise and capture the most important characteristics for the final result. It’s better to keep some capabilities as skills with clear triggers, loaded on demand. The models are already trained on the most relevant content humans have produced; you just need to point to what you consider most important.</p><p>Prefer focusing on reducing ambiguity. Set your code style to impose your personality. Make it aware of the scope it’s going to work in by simply telling it who you are, what your work is about, and the basic information it needs — to avoid mixing up funny repository names with popular TV series characters.</p><p>These are some recommendations and strategies, but you can go beyond by being proactive, questioning your team and asking for feedback. Understanding the real expectations people have of your work is core to ensuring you’re going in the right direction.</p><p>I hope you enjoyed it. See you in my next article. o/</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc7feeba9678" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>