<?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 Rossella Ferrandino on Medium]]></title>
        <description><![CDATA[Stories by Rossella Ferrandino on Medium]]></description>
        <link>https://medium.com/@rferrandino.tokyo?source=rss-1317749d29b1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*gSrlgyFYzIbg_RkLuTnuiA.jpeg</url>
            <title>Stories by Rossella Ferrandino on Medium</title>
            <link>https://medium.com/@rferrandino.tokyo?source=rss-1317749d29b1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 17:27:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rferrandino.tokyo/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[Ask for a guess, not an overview]]></title>
            <link>https://medium.com/@rferrandino.tokyo/ask-for-a-guess-not-an-overview-da54fc56daa7?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/da54fc56daa7</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[best-practices]]></category>
            <category><![CDATA[ai-prompts]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 03:06:00 GMT</pubDate>
            <atom:updated>2026-09-05T03:06:00.965Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xR4n548x8JgzeYnp" /><figcaption>Photo by <a href="https://unsplash.com/@seagrave?utm_source=medium&amp;utm_medium=referral">Nick Seagrave</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Last week ago I started on a new project. Normally the first day or two goes the same way: get access, open the repo, start scrolling through code trying to piece together what actually connects to what. This time I decided to let AI take the first pass.</p><p>My prompt was simple:</p><blockquote><em>“I am onboarding this new codebase and I want to have an overview of what external services (outside of Shopify) the theme connects to.”</em></blockquote><p>I already knew there were custom apps and external systems involved, I just wanted the map laid out before I went digging manually.</p><p>I also got lucky before I even got to that step. Someone who’d worked on this project before me had left some documentation in the GitHub wiki. It was an amazing starting point (albeit in some places outdated), and I’m genuinely grateful they took the time to write it down.</p><p>Claude scanned the code and put together a first-pass overview, complete with a few mermaid diagrams showing how the pieces connected. I adjusted those diagrams as I went, and used Claude’s Slack integration to cross-reference what the code suggested against what the team had actually said and decided about those integrations.</p><p>Then the corrections started.</p><p>A relationship between two systems that Claude was confident about turned out to be wrong, I corrected it, and the next answer flipped it the other way, still wrong. Ownership of one integration sat with a different company than the code implied. Most of this wasn’t visible to Claude simply because it couldn’t be: the code for those other services lives in separate repositories, so all it had to go on was this one repo’s side of the story. Round after round, I kept supplying the piece of context that only existed in a Slack thread, an old decision, or documentation someone else had written, since this was my first time on this project and I had no memory of my own to draw on. Everything I corrected with, I’d had to go find first.</p><p>By the time the document actually matched reality, I’d lost count of how many corrections it had taken. Somewhere around round thirty, I stopped and thought: <strong>there has to be a better way to do this.</strong></p><p>Not “a better way for AI to guess right the first time,” that’s not realistic when the answer depends on institutional knowledge that lives in people’s heads and old Slack threads, not in a git repo. The better way was in how I’d framed the ask to begin with.</p><p>I’d asked for “an overview,” which is a request for a finished, confident answer. What I actually needed was a request for a <em>starting point</em>, one that was upfront about its own limits. Something closer to:</p><blockquote><em>“Give me a first-pass map of this codebase’s external integrations from the code alone, clearly flagged as inferred, so I can correct and fill it in from what I actually know.”</em></blockquote><p>The difference matters more than it looks. “An overview” produces something that reads as finished, and something that reads as finished is easy to under-scrutinize, especially by whoever onboards after me and doesn’t have access to the same context I did. “A first-pass map, flagged as inferred” produces something that visibly still needs verifying, which is closer to the truth of what a codebase-only analysis can actually deliver.</p><p><strong>The cost side of this, that’s easy to skip past</strong></p><p>It’s worth reflecting on what those thirty-plus rounds actually were: full inference passes, plus my own time reading, checking, and correcting each one. Multiply that across everyone doing this kind of onboarding, and “just ask AI for an overview” stops looking free. The speed only shows up if you already hold the ground truth needed to correct it efficiently. Without that, you either ship an unverified document that looks confident, or burn even more rounds, and more compute, more energy discovering what you don’t know by trial and error.</p><p>The fix isn’t “prompt until it’s right.” It’s asking for the inferred first pass explicitly, up front, so the model is wrong once, clearly labeled as a guess, instead of confidently wrong dozens of times before landing on something true.</p><p><strong>A wrinkle I didn’t expect: verification isn’t equally available to everyone</strong></p><p>My team is bilingual, some documentation and Slack threads are in Japanese, some in English. I’m still learning the language, so for the Japanese sources, I couldn’t always verify what Claude summarized the way I could with the English ones. I had to trust the translation and the summary more than I trusted anything sourced in English, where I could read the original myself and catch a wrong inference immediately.</p><p>That’s a gap worth naming honestly. My verification loop only worked as well as it did because I could personally check most of the sources. Where I couldn’t, I was closer to the position a less experienced teammate is in all the time: taking the output on faith because checking it isn’t realistically possible for them either, whether that’s a language barrier, a lack of tenure on the team, or simply not knowing which Slack channel to search. My approach was “verify everything,” but that assumes verification is actually available to you. It isn’t always.</p><p><strong>A few things I’m taking into the next onboarding:</strong></p><ul><li>Ask for a first-pass map explicitly labeled as inferred, not a finished overview</li><li>Treat AI-generated architecture docs as a draft that needs my ground truth, not a source of truth on its own, and notice when I actually can’t verify a source myself</li><li>The real accelerant wasn’t the AI reading the code, it was being able to feed it the context that only exists because we’re a distributed team who writes things down</li></ul><p>None of this would have worked without that written context, in the wiki, in Slack, in old decisions someone bothered to document. The AI made pulling it together faster. It didn’t make the knowledge itself unnecessary, and it didn’t make the rounds it took to get there free.</p><p><strong>A question for whoever reads this</strong></p><p>I could verify most of what Claude gave me because I had the language, the Slack access, and enough tenure to know who to ask. What do you do when you don’t have all of that? Do you go in with a prepared list of questions, the way I would before an alignment meeting, and use the AI’s first pass to figure out what to ask? Or is there a different approach entirely for onboarding when the gaps aren’t just “unknown,” but structurally out of reach, whether that’s a language you’re still learning, a team you just joined, or documentation that simply doesn’t exist yet?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da54fc56daa7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What the JLPT registration rush taught me about high-traffic systems]]></title>
            <link>https://medium.com/@rferrandino.tokyo/what-the-jlpt-registration-rush-taught-me-about-high-traffic-systems-b392503a660e?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/b392503a660e</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[user-experience]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 13:19:35 GMT</pubDate>
            <atom:updated>2026-08-30T13:19:35.946Z</atom:updated>
            <content:encoded><![CDATA[<p>Last Sunday I stayed up until midnight to register for the JLPT (Japanese Language Proficiency Test).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eQIr0OqGG3TB4wGosyuiyQ.jpeg" /><figcaption>My prep materials for the exam</figcaption></figure><p>I had already created my account, so I expected the process to be straightforward: choose the level, confirm my details, upload a photo, pay the fee, done.</p><p>Instead, the website kept failing.</p><p>Most of the time I made it as far as the photo upload before seeing an error page saying the site was congested. The particularly painful part was that an error would also kick me out of my account, so I had to log in and start the process again.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/699/1*uBPjyTHTA7QW2n-aRImxqQ@2x.png" /><figcaption>Error message from the JLPT website</figcaption></figure><p>I eventually gave up for the night and tried again the next morning. At around 11am, after what must have been close to 50 attempts in total, I finally completed the registration.</p><p><a href="https://info.jees-jlpt.jp/">JEES</a> confirmed that the application site was experiencing connectivity problems because of concentrated traffic. Demand was clearly high: applications opened on August 24, and N3, N4 and N5 registrations closed on August 27 (expected closing date was supposed to be September 7).</p><p>I don’t know how the JLPT website is built, so I can’t say exactly what was failing.</p><p>But I have recently started working more with full-stack applications, and the experience made me curious: <strong>what actually happens when too many people try to use the same application at once, and how should that state be communicated back to the user?</strong></p><h4>What does “network congestion” actually mean?</h4><p>The error message described the problem as “network congestion.” From a user perspective, that tells me traffic was high, but it doesn’t tell me where the bottleneck actually was.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b9Nzn0NAXdLDDuUp" /><figcaption>Photo by <a href="https://unsplash.com/@awan_project?utm_source=medium&amp;utm_medium=referral">Awan</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Before looking into this, my mental model was basically: too many requests reach the server and eventually it can’t keep up.</p><p>That’s not completely wrong, but it leaves out a lot.</p><p>Out of curiosity, I ran the MyJLPT application through <a href="https://www.wappalyzer.com/">Wappalyzer</a> to see what could be detected from the outside. It identified AWS infrastructure including Amazon CloudFront and an Application Load Balancer.</p><p>Wappalyzer can only infer technologies that are publicly visible, so this isn’t a complete architecture diagram. But it already made my original idea of “too many people hitting one server” look much too simple.</p><p>CloudFront can cache and serve some content closer to users, while a load balancer can distribute incoming requests across multiple application targets.</p><p>But distributing traffic doesn’t mean every part of the application suddenly has unlimited capacity.</p><p>Those application instances may still depend on shared resources further downstream: a database, session storage, file storage, or other services. A registration flow is also doing much more than serving a static page. It needs to keep track of the logged-in user, read and write registration data, process an uploaded photo and eventually hand part of the payment process over to another provider.</p><p>Even though I kept failing during the photo upload, that doesn’t necessarily mean the upload itself was the bottleneck.</p><p>What interested me more was what happened after each failure.</p><p>Every time I saw the error, I did what most users would probably do: I tried again.</p><p>If thousands of people are doing the same thing, those retries create more work for a system that is already struggling. Users who were halfway through the process are logging in again, loading the same pages again and resubmitting the same information.</p><p>So even the way a system handles failure can affect the amount of traffic it has to deal with.</p><h4>A better version of the same waiting problem</h4><p>A few days later, I applied for a lottery for a new Pokémon card release.</p><p>There was high demand there too, but the experience was completely different.</p><p>Instead of immediately entering the website, I was placed into a virtual queue.</p><p>It told me that I had around nine minutes to wait. Once it was my turn, I entered the site and completed the application without any interruption.</p><p>I had technically spent nine minutes being prevented from using a website, yet the experience felt significantly better.</p><p>I knew what was happening.</p><p>There was no reason to refresh the page. I wasn’t wondering whether the next attempt would work. And, most importantly, thousands of us weren’t repeatedly hitting the application while it was already under pressure.</p><p>That made me want to understand what the queue was actually doing.</p><h4>A queue doesn’t make the website faster</h4><p>One system I’ve encountered several times is <a href="https://queue-it.com/">Queue-it</a>, which provides what it calls a virtual waiting room.</p><p>The basic idea is simple.</p><p>Instead of allowing everyone to reach the application at once, excess visitors wait somewhere else. The website owner controls how quickly people are allowed through based on what the underlying application can handle.</p><pre>A lot of users arrive<br>        ↓<br>Virtual waiting room<br>        ↓<br>Controlled number of users enter<br>        ↓<br>Registration site</pre><p>So the queue isn’t adding capacity to the registration system. It’s controlling demand.</p><p>This connects to a broader engineering idea called <a href="https://medium.com/@jayphelps/backpressure-explained-the-flow-of-data-through-software-2350b3e77ce7">backpressure</a>. When one part of a system is receiving work faster than it can process it, something upstream needs to slow things down.</p><p>Otherwise, the waiting still happens, just somewhere less useful. Requests pile up, take longer to complete, or eventually time out.</p><p>A virtual waiting room makes that waiting explicit and moves it outside the part of the application you are trying to protect.</p><p>For something like an exam registration or a product launch, the traffic spike is often predictable. Everyone has been told that registration opens at midnight, so everyone tries to arrive at midnight.</p><p>Trying to make the whole system capable of serving all of those people simultaneously is one approach. Another is to control how many people are allowed into the expensive part of the application at once.</p><p>Queue-it also supports a pre-queue for scheduled launches. People who arrive before the opening time can wait outside the application and receive a position when the event begins.</p><p>That also changes user behaviour. There is less incentive to sit there refreshing at 23:59 hoping your request reaches the application a fraction of a second before everyone else’s.</p><p>But a virtual queue is not automatically the solution to every high-traffic problem.</p><p>If the actual bottleneck is an inefficient database query, a fragile upload flow, or another dependency with its own limits, a queue doesn’t fix that. It can reduce the pressure reaching the system, but the underlying weakness is still there. That distinction was useful for me.</p><p>A load balancer, a CDN, more application instances and a virtual queue all solve different parts of the problem. Reliable infrastructure is not one component you add and forget about. The pieces have to work together.</p><p>And then the system still has to communicate what is happening to the user.</p><h3>What happens when one step fails?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/522/1*RrzCD7z9-x0IXHHwFRS_fw@2x.png" /></figure><p>That was another part of the JLPT experience I kept thinking about.</p><p>In my case, a failed photo upload meant losing my progress and starting again. If I had already selected the JLPT level I wanted to apply to and confirmed my personal information, could that progress have been saved so I only needed to retry the failed step?</p><p>That sent me down another small rabbit hole around persisted state and safe retries.</p><p>Once I finally made it past the photo upload later that morning, I had no trouble with payment.</p><p>From my perspective, it was all one registration flow. Technically, it was likely several different systems working together.</p><p>And that is where the frontend becomes part of the reliability story too.</p><p>A backend service can know that something is overloaded, slow, unavailable or safe to retry. But unless that state is surfaced clearly through the interface, the user is left guessing what to do next.</p><h4>Waiting can actually be the better experience</h4><p>Years ago, when I was working as a cinema manager, I was given one piece of advice: if something goes wrong, keep people informed. If a screening is delayed, tell them. Give them an idea of what to expect, and let them know you’ll update them again.</p><p>The same principle applies to software.</p><p>With the Pokémon lottery, I had to wait nine minutes, but I knew I was in a queue and roughly how long it would take.</p><p>With the JLPT registration, I kept hitting errors without knowing whether the next attempt would work, whether my previous progress had been saved, or whether I should simply come back later.</p><p>That difference is what I find most interesting.</p><p>A virtual queue is only one possible technical solution. The broader lesson is that infrastructure and frontend behaviour need to communicate with each other.</p><p>If the system is at capacity, the interface should help the user understand that. If something can safely be retried, the UI can make that clear. If progress has been saved, tell the user. If the best option is to wait, give them enough information to feel comfortable waiting.</p><p>From a frontend perspective, error states, retries, progress indicators and queue screens can look like small UI details.</p><p>But they are also the point where the behaviour of the underlying system becomes understandable to the person using it.</p><p>And sometimes, nine minutes of clearly explained waiting is a much better experience than hours of trying again.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b392503a660e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Testing AI Systems Feels Different From Traditional Software Development]]></title>
            <link>https://medium.com/@rferrandino.tokyo/testing-ai-systems-feels-different-from-traditional-software-development-58b70af18457?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/58b70af18457</guid>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Tue, 19 May 2026 10:52:48 GMT</pubDate>
            <atom:updated>2026-05-19T10:52:48.550Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*M2uZbo7GSZCRIxG0" /><figcaption>Photo by <a href="https://unsplash.com/@brett_jordan?utm_source=medium&amp;utm_medium=referral">Brett Jordan</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I am working on an internal MCP-based knowledge tool. One thing that surprised me is how different testing feels compared to traditional software development.</p><p>When working with APIs or frontend systems, failures are usually obvious. A request breaks, a test fails, TypeScript complains, or the UI visibly stops working. The feedback loop is relatively clear.</p><p>With LLM systems, the experience feels much more gradual and subtle.</p><p>The system still responds. It just responds slightly worse. And because the failures are subtle, they can easily go unnoticed until someone starts relying on the system for real work.</p><h3>When a Small Retrieval Fix Quietly Broke Something Else</h3><p>While adjusting retrieval scoring to improve results for a newly added record, I changed some keyword weighting logic and the new documents started surfacing correctly. At first glance, everything looked fine.</p><p>But after running a small retrieval evaluation suite I had started putting together, I realized several existing queries had quietly degraded. Questions that previously returned documents from one topic were now surfacing unrelated content from another.</p><p>Nothing had technically broken. The system still functioned. But the quality had shifted in ways that would have been very difficult to notice through occasional manual testing.</p><p>That was the moment I started thinking more seriously about evaluation workflows. I had read about the concept before, but seeing it play out in a real project felt entirely different.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/468/1*OhrKe9kYw8cRgLPDLB59jg@2x.png" /><figcaption>Simplified retrieval + evaluation flow.</figcaption></figure><h3>What My “Eval Harness” Actually Looks Like</h3><p>The term “eval harness” sounded intimidating to me initially, but in practice what I built is much simpler than the name suggests.</p><p>At its core, it’s just structured QA for AI systems.</p><p>I created a small set of expected queries and outputs, then started running them automatically after changes to the retrieval logic or document corpus.</p><p>At the moment I mainly have two layers:</p><ul><li>Retrieval checks, where I verify whether the system surfaces the expected document sources for a query. This is simple text search and it doesn’t require any Claude API call to work.</li><li>End-to-end response evaluations, where an LLM scores whether the final answer contains the expected information</li></ul><p>The specific tooling matters less than the habit. I’m currently using <a href="http://medium.com/r?url=https%3A%2F%2Fwww.promptfoo.dev%2F">Promptfoo</a>, which makes iteration easy, but the more important shift was committing to testing consistently instead of relying on intuition.</p><p>In many ways, it reminded me of frontend regression testing. Once the system grows beyond a certain point, it becomes difficult to mentally track how one change affects everything else.</p><h3>Writing the Test Before the Feature</h3><p>One workflow that ended up helping a lot was writing evaluation cases before expanding the system.</p><p>For example, before adding a new knowledge base record, I would first write the retrieval queries I expected to work. Initially they would fail, because the documents had not been ingested yet. Then after adding the new data and adjusting retrieval logic, I could rerun the same checks and see whether the results improved.</p><p>That process feels similar to test-driven development, even if the outputs are probabilistic rather than deterministic.</p><p>The interesting difference is that with AI systems, the assertions themselves become more flexible. Instead of checking a strict return value, you are often checking whether:</p><ul><li>the right information appears</li><li>the correct source is retrieved</li><li>the answer remains relevant and grounded</li></ul><p>That ambiguity required me to think about testing in a much less binary way than I was used to. Less about asserting correctness, more about building confidence incrementally.</p><h3>The Parts That Still Feel Fragile</h3><p>At the same time, there are still several areas that feel much harder to evaluate reliably.</p><p>One example is multilingual retrieval. Some of the documents in my corpus are in Japanese, some of them are in English, whereas the majority of queries come in English. Right now the retrieval pipeline translates Japanese queries before keyword matching, which works reasonably well but still introduces inconsistencies.</p><p>Another challenge is tool selection between connected MCP servers. Sometimes the model chooses the wrong tool simply because the descriptions overlap semantically.</p><p>A query mentioning Shopify B2B, for example, may trigger a development-oriented MCP (Shopify Dev MCP) instead of the knowledge tool.</p><p>Those issues are much harder to catch through retrieval tests alone because they sit at the prompting layer rather than the retrieval layer itself.</p><p>I’ve found that this type of behavior still requires a lot of manual exploration and observation, and test case adjustments when you find something that is consistently failing.</p><h3>What This Project Changed for Me</h3><p>The biggest shift for me was realizing that building AI systems feels less like traditional feature development and more like managing system behavior over time.</p><p>The challenge is not only making something work once. It’s making sure the system continues behaving reasonably as:</p><ul><li>new documents (sometimes with different formats) are added</li><li>prompts evolve</li><li>retrieval logic changes</li><li>tools expand</li><li>contexts overlap</li><li>users start using the product in unpredictable ways</li></ul><p>That also changed the way I think about confidence in software systems.</p><p>With traditional development, passing tests often feels binary. With LLM systems, confidence feels much more probabilistic and incremental. The goal becomes reducing uncertainty rather than eliminating it entirely.</p><p>I’m still very early in learning this space, but one thing I’ve already found interesting is how quickly AI tooling pushes frontend and application development closer to areas like retrieval systems, evaluation design, and information architecture. <br>It’s a much broader set of problems than I expected to work on, but that’s also part of what has made the project so interesting to explore.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=58b70af18457" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Storefront Is No Longer Just for Humans]]></title>
            <link>https://medium.com/@rferrandino.tokyo/the-storefront-is-no-longer-just-for-humans-82e95a2f03f3?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/82e95a2f03f3</guid>
            <category><![CDATA[ai-driven-commerce]]></category>
            <category><![CDATA[frontend-development]]></category>
            <category><![CDATA[shopify]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Thu, 07 May 2026 10:08:08 GMT</pubDate>
            <atom:updated>2026-05-07T10:08:08.565Z</atom:updated>
            <content:encoded><![CDATA[<p>Last year <a href="https://medium.com/@rferrandino.tokyo/from-shopify-catalog-to-chatgpt-traffic-how-ai-is-rewriting-seo-c33620b471c1">I wrote</a> about how AI tools like ChatGPT were beginning to reshape search and product discovery. At the time, much of the conversation still felt speculative. There was a strong focus on GEO (Generative Engine Optimization), conversational search, and the idea that AI assistants might eventually become a new acquisition channel for e-commerce.</p><p>Listening to <a href="https://www.linkedin.com/in/eduardofrias/">Eduardo Frias</a>’ recent talk on AI-driven commerce at a Shopify event in Tokyo, what struck me is how much more concrete the discussion has become over the past year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1ptOWLAvmzVGCtmU-D25_w.jpeg" /><figcaption>Eduardo Frias in Tokyo</figcaption></figure><p>The focus is no longer only on the idea that “AI will change e-commerce”, but on the implementation details behind that shift: structured product data, APIs, retrieval systems, machine-readable storefronts, and the way AI systems actually consume commerce experiences.</p><h3>From Browsing to AI Discovery</h3><p>For a long time, e-commerce UX has been designed primarily around human browsing behavior. Storefronts guide users through navigation menus, collections, campaigns, and conversion funnels. Discovery often starts from the homepage and gradually narrows toward a purchase decision.</p><p>AI-driven commerce introduces a different interaction model. Users increasingly arrive directly on product pages through conversational interfaces and AI recommendations rather than traditional browsing flows. In that context, structured and unambiguous information becomes much more important because AI systems do not interpret context in the same way humans do.</p><p>One point from the talk stayed with me: unclear or incomplete data can immediately remove a store from consideration. If an AI system cannot confidently retrieve or understand information about a product, pricing, availability, or delivery conditions, it simply moves elsewhere.</p><p>That changes the role of the storefront itself. <strong>A product page is no longer only a visual interface for customers, but also a source of structured information consumed by recommendation systems, retrieval pipelines, and AI agents.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*ZeAdWHW9NFm83vUO11sUNw@2x.png" /><figcaption>A simplified view of how AI-driven commerce changes the discovery flow around storefronts.</figcaption></figure><h3>Structured Data, APIs, and the Expanding Role of Frontend</h3><p>Traditionally, frontend work has focused on interfaces for human users: layout, responsiveness, interactions, performance, and accessibility. Those concerns still matter, but AI-driven commerce introduces a second audience consuming storefronts alongside people: machines.</p><p>In some ways, this shift is not entirely new. Frontend developers have already been working with structured data and machine-readable information for years through SEO and metadata implementations. I remember working on JSON-LD schemas for store locator pages using Shopify Liquid to dynamically output the values required by SEO teams. At the time, the goal was primarily search engine visibility and richer search results. What feels different now is the scale and importance of those systems as AI moves closer to becoming an active layer in product discovery and recommendation.</p><p>As a result, information that was once treated almost as supporting content becomes increasingly central. Product specifications, metadata, FAQs, reviews, alt text, ingredients, delivery information, and return policies all contribute to how AI systems understand and evaluate a store.</p><p>Within Shopify ecosystems, this becomes especially important because storefronts are often connected to multiple external systems and apps. Inventory, pricing, availability, delivery conditions, and policies all need to remain synchronized across APIs, metafields, search systems, and frontend implementations. When those signals become fragmented or outdated, trust quickly breaks down, both for users and for AI systems relying on that information to surface products and recommendations.</p><p>In that sense, frontend work starts moving closer to information architecture and system design, not only visual presentation. The structure, consistency, and accessibility of data become just as important as the interface itself.</p><p>This also changes part of the conversation around performance and optimization. Frictionless checkout flows and accessible information are no longer only UX concerns. They increasingly influence whether AI systems can confidently recommend a storefront in the first place.</p><h3>Japanese Ecommerce Through the Lens of AI Commerce</h3><p>Working in Japan made me think about how some local e-commerce patterns already align surprisingly well with machine-readable commerce.</p><p>Japanese product pages are often heavily optimized for information density. Compared to many Western storefronts, Japanese PDPs frequently surface a large amount of detail directly on the page: specifications, shipping conditions, sizing guidance, FAQs, ingredients, compatibility notes, warnings, rankings, and customer reassurance elements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WN3esgHTPK5TuT623Lz7nA@2x.png" /><figcaption>Comparison between New Era Japan and Italy product pages. Japanese ecommerce interfaces often prioritize explicit product information and reassurance elements more heavily than Western storefronts.</figcaption></figure><p>Historically, some of these pages have been criticized for feeling visually crowded. But from the perspective of AI-driven commerce, they are already surprisingly aligned with what recommendation systems are looking for: explicit, structured information with minimal ambiguity.</p><p>Clear, scannable descriptions also become important here, especially since the talk emphasized that the first several thousand characters of a product page carry the most weight for AI systems interpreting content. At the same time, it raises an interesting question about where the limit is. How much information is useful before clarity starts to suffer? Designing for both human readability and machine readability may become an increasingly delicate balance.</p><p>The checkout experience presents an interesting contrast. Japanese e-commerce often introduces more friction into purchasing flows, with customer registration frequently required before completing checkout. Part of this comes from expectations around loyalty programs, customer management, and repeat purchasing behavior.</p><p>If AI-driven commerce increasingly prioritizes speed and low-friction transactions, it raises interesting questions about whether markets like Japan will gradually move toward more flexible or loginless checkout experiences in the future.</p><h3>The UX Tension</h3><p>Another aspect I keep thinking about after the event is the balance between visual uniqueness and machine readability.</p><p>Beautiful design still matters for human users, but AI systems prioritize clarity, structure, consistency, and accessible information. As storefronts increasingly serve both people and machines, semantic clarity may become just as important as branding and visual differentiation.</p><p>I don’t think there are clear answers yet, but it feels like an important transition to pay attention to.</p><h3>AI Commerce Is Becoming an Architectural Conversation</h3><p>The biggest thing I took away from the event is that AI-driven commerce is no longer only changing how users discover stores. It is also starting to influence how storefronts themselves are architected.</p><p>The conversation is moving beyond marketing and SEO into frontend systems, APIs, retrieval logic, and information architecture. For frontend developers, that creates a broader definition of what a storefront actually is.</p><p>If AI systems increasingly become part of the commerce journey, then building storefronts that are understandable both to humans and machines may become one of the most important challenges in ecommerce development.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82e95a2f03f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Prompting to Systems Thinking: My Experience with Anthropic’s Claude API Course]]></title>
            <link>https://medium.com/@rferrandino.tokyo/from-prompting-to-systems-thinking-my-experience-with-anthropics-claude-api-course-40dec465b4c8?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/40dec465b4c8</guid>
            <category><![CDATA[anthropic-academy]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Wed, 06 May 2026 13:22:54 GMT</pubDate>
            <atom:updated>2026-05-06T13:22:54.932Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VnE4p663yQePYus6" /><figcaption>Photo by <a href="https://unsplash.com/@hdbernd?utm_source=medium&amp;utm_medium=referral">Bernd 📷 Dittrich</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Over the past year, AI has increasingly become part of everyday developer workflows.</p><p>Where I work, we’ve also been encouraged to explore these tools more deeply and think about practical applications inside the company.</p><p>As part of that, I recently completed a series of courses from <a href="https://anthropic.skilljar.com/">Anthropic Academy</a>. One of the most comprehensive one is <a href="https://anthropic.skilljar.com/claude-with-the-anthropic-api">Building with the Claude API</a>.</p><p>Originally, I expected something mostly focused on prompts and API calls. What I found more interesting was how much the course approached AI from an engineering perspective: workflows, evaluation, retrieval systems, tools, and architectures.</p><p>More than anything, it gave me a clearer mental model of how AI systems actually work end to end.</p><h3>More Than Prompt Engineering</h3><p>The course is quite comprehensive, and I worked through it over a couple of weeks rather than trying to finish it quickly.</p><p>What helped a lot is that the content is divided into clear macro topics:</p><ul><li>API fundamentals</li><li>Prompt engineering and evaluation</li><li>Tool use</li><li>RAG systems</li><li>MCP (Model Context Protocol)</li><li>Agents and workflows</li></ul><p>Because of this structure, it was easy to pause and come back later without losing context.</p><p>One thing I appreciated is that the course treats prompting as only one part of a much larger system. A lot of online AI content focuses on finding the “perfect prompt”, while this course focuses much more on evaluation, iteration, workflows, retrieval pipelines, and designing systems that behave consistently.</p><p>The course examples are intentionally simplified, but I think that’s part of why the learning experience works well. It gives enough depth to understand the implementation patterns without becoming inaccessible.</p><h3>MCP and Internal Knowledge Systems</h3><p>The section that sparked the most ideas for me was MCP. I’m still learning and experimenting, but it changed the way I think about internal knowledge systems.</p><p>At agencies and companies, useful information often lives across many disconnected places:</p><ul><li>documentation</li><li>conversations</li><li>partner resources</li><li>internal notes</li><li>app documentation</li></ul><p>The MCP section made me think more deeply about how these resources could potentially be structured and retrieved more consistently.</p><p>I’m now starting to explore how internal knowledge could be organized using an MCP-style setup.</p><p>What I found especially interesting is how quickly real-world complexity starts to appear. For example, supporting a bilingual team where some people query a system in Japanese and others in English, while the underlying sources may exist in either language.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xi--2pu5lV5tpfeC" /><figcaption>Photo by <a href="https://unsplash.com/@akurakuu?utm_source=medium&amp;utm_medium=referral">kuu akura</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>That was the moment where the course shifted for me from “learning AI tools” to thinking more seriously about systems and information architecture.</p><h3>The Value of Good Technical Explanations</h3><p>Another thing I appreciated was the way the material was explained visually.</p><p>The diagrams throughout the course were genuinely useful for understanding workflows, agent patterns, RAG pipelines, and tool interactions. For my learning style, seeing how pieces connect together visually makes complex topics much easier to absorb.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P_4MaaxsJZFj_GEFYZzPWw@2x.png" /><figcaption>Screenshot of some of the diagrams used in the course. Source: Skilljar</figcaption></figure><p>Especially in AI, where terminology and abstractions can quickly become overwhelming, having clear visuals and structured explanations made the material feel approachable even when the concepts themselves weren’t simple.</p><p>Compared to other technical courses I’ve taken in the past, this one felt particularly well paced and easy to follow.</p><h3>Final Thoughts</h3><p>What I appreciated most about this course is that it approached AI as an engineering discipline rather than a collection of isolated tricks.</p><p>It didn’t remove the complexity, but it made the complexity feel easier to reason about.</p><p>I still have a lot to learn, but I now feel much more confident exploring practical AI applications and understanding how the different pieces connect together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40dec465b4c8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Recognition as a skill: lessons from building visibility in tech]]></title>
            <link>https://medium.com/@rferrandino.tokyo/recognition-as-a-skill-lessons-from-building-visibility-in-tech-9a969a5c4823?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/9a969a5c4823</guid>
            <category><![CDATA[women-in-tech]]></category>
            <category><![CDATA[career-growth]]></category>
            <category><![CDATA[wise-japan]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Sat, 22 Nov 2025 09:49:03 GMT</pubDate>
            <atom:updated>2025-11-22T09:49:03.956Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Recognition is the bridge between doing good work and being seen for it.</em></p><p>In October, I had the chance to co-host a <strong>Women in Software Engineering JP</strong> event at Code Chrysalis, in Tokyo.<br>The topic: <em>Getting recognition for your work.</em></p><p>As women in tech, many of us know how it feels to do great work quietly.<br>We fix problems, improve systems, help our teams move faster, but sometimes that impact stays invisible. This session was about exploring how we can make our work visible in authentic, inclusive ways that support both our growth and our community.</p><h4>Why Recognition Feels Difficult</h4><p>When we asked the group what felt hardest about getting recognition, the responses came quickly:</p><ul><li>“Hard to talk about myself”</li><li>“Sharing work publicly”</li><li>“Too busy to showcase my work”</li></ul><p>These are common challenges, especially in Japan, where humility and teamwork are deeply valued. Many engineers prefer to let results speak for themselves, but the truth is: results often need a little help being heard.</p><p>Recognition isn’t about self-promotion. It’s about <strong>communication</strong>:<strong> </strong>helping others understand the <em>impact</em> of your work.</p><h4>Rethinking Recognition</h4><p>During the session, we reframed recognition as more than praise.<br>It’s a system of connection between your effort and its impact, between your growth and your community.</p><p>Recognition creates:</p><ul><li><strong>Growth: </strong>new opportunities, projects, and trust.</li><li><strong>Impact: </strong>clarity on how your work changes outcomes.</li><li><strong>Confidence: </strong>motivation and self-belief.</li><li><strong>Community: </strong>shared learning and collaboration.</li></ul><p>When we look at it this way, recognition isn’t something to wait for. It’s something we can practice.</p><p>And it often starts with small, intentional actions.</p><h4>Frameworks for Making Work Visible</h4><p>We discussed four simple habits that help make technical work visible.</p><ol><li>Track it<br>Keep a “wins notebook.” Note things you’ve solved, learned, or improved, even small ones. These records help when writing reviews or preparing for promotion discussions.</li><li>Frame it<br>Instead of describing effort, describe <em>impact.</em><br> Not “I optimized the homepage,” but “Our team worked on performance. I handled image optimization, which cut homepage load time from 3 to 2s”</li><li>Share it<br>Visibility grows from consistency. Share small updates: a Slack message, a team demo, or a comment in a retro. These micro-moments of visibility build awareness over time.</li><li>Amplify it<br>Recognition is also collective. Highlight others’ contributions. Share teammates’ achievements in your updates or team channels. It builds a culture of visibility that benefits everyone.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lSYm9yvXMUXu9t3rTy4SUg@2x.png" /></figure><h4>Recognition in Japan</h4><p>One theme we discussed was how recognition fits into Japanese work culture. Three concepts are particularly important:</p><ul><li>和 <em>Wa</em> emphasizes harmony and group success.</li><li><em>根回し Nemawashi</em> shows how influence is built through informal discussions before decisions are made.</li><li>And the 先輩・後輩 <em>senpai–kohai</em> relationship means recognition often flows through mentors and senior colleagues.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*iCxeseYqpQGcrKx1NzCvAQ@2x.png" /></figure><p>The same principles apply to visibility:</p><ul><li>Frame your contributions in terms of group effort, but highlighting the part you played in the group</li><li>Talk about the “why” behind your projects in small conversations, not just formal reviews. Prepare the the ground quietly: share your ideas, progress, or challenges early so that when decisions or recognition moments come, others already see your value.</li><li>Share your progress early with a senpai or mentor.</li></ul><p>Visibility doesn’t have to mean standing in front of everyone.<br>It can also mean creating understanding through relationships, the quieter form of recognition that many workplaces in Japan value deeply.</p><h4>What We Heard from the Community</h4><p>When we asked attendees to share one technical accomplishment they were proud of but felt hadn’t been fully seen, the responses were honest and diverse, and they perfectly captured the kinds of invisible work that power every tech team.</p><p>Some engineers talked about <strong>deep technical wins</strong>, like:</p><ul><li>refactoring legacy code</li><li>improving testing methodologies</li><li>streamlining infrastructure</li></ul><p>Others mentioned <strong>cross-functional impact</strong>, such as:</p><ul><li>coaching or mentoring teammates</li><li>driving project completion on time</li><li>or helping fix a feature that wasn’t even their responsibility</li></ul><p>There were also many <strong>quiet enablers: </strong>the documentation writers, the code reviewers, the people archiving thousands of unused files or troubleshooting client issues that no one else sees.</p><p>Each of these contributions creates value. But as many participants shared, they often go unnoticed because they happen behind the scenes, or outside formal recognition systems.</p><p>Then we then shifted the conversation forward, asking everyone to think of one thing they could make <em>more visible</em> this month. Some committed to sharing learnings on Slack, others to mentoring, and others to speaking up more in project updates. And using analytics to backup the impact of their contributions.</p><p>Seeing the group reflect and take action was a reminder that recognition isn’t a single big moment, it’s something we build slowly, through steady, authentic visibility.</p><h4>Recognition as Practice</h4><p>The biggest takeaway for me as a speaker was this:</p><blockquote><em>Recognition isn’t a reward you receive. It’s a dialogue you start.</em></blockquote><p>Every time we document, share, or connect our work to outcomes, we open that dialogue. With our teams, our managers, and the wider community.<br>Sometimes it feels uncomfortable, and sometimes it meets resistance. That’s okay. Growth rarely happens in comfort zones.</p><p>What matters is building the habit: making visibility part of how we work, not just something we do when asked to “show results.”</p><h4>Closing Thoughts</h4><p>At WiSE Japan, we talk a lot about <strong>technical excellence</strong> and <strong>inclusive leadership</strong>.<br>Recognition sits at the intersection of both: it’s how excellence becomes visible and how inclusion becomes real.</p><p>I’m grateful to everyone who joined the event, shared their stories, and reminded me that visibility isn’t about ego, it’s about connection.</p><blockquote><em>When we make our work visible, we don’t just advance our own careers. We make space for others to do the same.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*inZisI_eybO8h-_LAiGHlQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9a969a5c4823" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Mapping the Shopify App Ecosystem in Italy (2025)]]></title>
            <link>https://medium.com/@rferrandino.tokyo/mapping-the-shopify-app-ecosystem-in-italy-2025-203545ae26f5?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/203545ae26f5</guid>
            <category><![CDATA[shopify]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[shopify-apps]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Mon, 15 Sep 2025 08:12:01 GMT</pubDate>
            <atom:updated>2025-09-15T08:14:27.667Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gABG3wxjWKskw19g" /><figcaption>Photo by <a href="https://unsplash.com/@tata186?utm_source=medium&amp;utm_medium=referral">Tatiana Rodriguez</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>When we talk about the <a href="https://apps.shopify.com/">Shopify App Store</a>, the conversation often gravitates toward the biggest global players. But zooming into a specific country reveals a more nuanced picture, one shaped by local business needs, developer strengths, and cultural context.</p><p>Over the month of August, I set out to map the <strong>Shopify app ecosystem in Italy</strong>. I came across <a href="https://filippi.dev/reports/apps">Filippi.dev’s global app reports</a>, but noticed the dataset was slightly outdated. To get the most reliable snapshot, I built a <strong>Python crawler</strong> to scan Shopify’s own App Store sitemap. The global count had grown from 11,208 apps to over <strong>14,200 apps</strong>, and within that, I identified 120 apps developed by Italian companies.</p><p>The result was not just an exercise in data collection. It was also a chance to see familiar names of developers and agencies I’ve crossed paths with over the years. It reminded me how connected and collaborative this ecosystem is, even as it scales globally.</p><h3>Dataset &amp; Methodology</h3><p>To build this report, I took inspiration from <a href="https://filippi.dev/reports/apps">Filippi.dev</a>’s Shopify app dataset, which originally tracked around 11,200 apps. Since that dataset was several months old, I developed a Python crawler to scan the Shopify App Store sitemap directly and compile an updated dataset.</p><p>The current count (as of August 27th, 2025) is 14,205 apps globally. From this larger pool, I filtered out the apps associated with developers based in Italy, using their self-reported addresses.</p><p>For each app, I collected the following fields:</p><ul><li>App name</li><li>Developer name</li><li>Developer address (region in Italy)</li><li>Reviews count</li><li>Categories</li><li>Multi-language availability</li><li>Description</li><li>Built for Shopify badge</li></ul><p>A quick note on limitations:</p><ul><li>Developer addresses are self-reported, so there may be edge cases (international companies with Italian branches, or developers who moved).</li><li>“<a href="https://shopify.dev/docs/built-for-shopify">Built for Shopify</a>” achievement is a clear signal, but not the only indicator of quality.</li><li>Language support is taken at face value from the app listings.</li></ul><p>Despite these caveats, the dataset offers a reliable snapshot of the Italian Shopify app ecosystem today, with enough depth to highlight meaningful trends and opportunities.</p><h3>Who Builds Shopify Apps in Italy?</h3><p>Italy’s ecosystem is diverse. Some developers are large agencies, others are individuals building niche tools.</p><ul><li><strong>Top players</strong>: <a href="https://nextools.tech/">Nextools</a> leads the ecosystem with 10 apps, more than double the second-place developer, <a href="https://apps.shopify.com/partners/prtdev">PRT</a> (4 apps). Other notable contributors include <a href="https://apps.shopify.com/partners/myappify">MyAppify</a> (4 apps), <a href="https://apps.shopify.com/partners/idt-s-p-a">IDT S.p.A.</a> (3 apps) and <a href="https://apps.shopify.com/partners/easycom-italia">EASYCOM ITALIA</a> (3 apps).</li><li><strong>Developer mix</strong>: Looking beyond the top names, the ecosystem spreads across 6 major professional agencies, 15 specialized software or tech companies, 14 e-commerce and logistics firms, 14 digital agencies, 5 industry-specific companies (fashion, logistics, automotive), 26 startups, 8 individual developers, and 2 communication service providers.</li></ul><p>What stood out to me in this mix is how balanced it feels.</p><p>You see everything from solo freelancers experimenting with new ideas, to established software houses running multiple apps with steady revenue. Logistics is clearly a shared strength — no surprise in a market where Italian merchants often sell abroad — but there’s also specialization in areas like fashion and automotive that feel very “Italian.” It’s not a uniform landscape, and that variety is what makes it interesting.</p><p>Reviews add another layer to this picture. Most Italian apps have just a handful of reviews, which shows how fragmented and early-stage the market still is. But a smaller group of apps — often in shipping, dropshipping, or marketing — stand out with far higher review counts. These apps usually come from the same agencies already leading in volume, suggesting that scale and merchant adoption tend to reinforce each other.</p><p>Geographically, developers are spread across the country. Milan, Rome, and Turin unsurprisingly appear as hubs, but there are also developers in smaller cities like Biella, Lucca, and Bolzano. Much like Italy’s broader SME landscape, the Shopify app community isn’t centralized — it thrives across regions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B1CV24mp3jN5RysSNOd8qg.png" /><figcaption>Top provinces for Shopify Apps in Italy</figcaption></figure><h3>What Do These Apps Focus On?</h3><p>The strongest categories reflect Italy’s e-commerce priorities:</p><ul><li><strong>Shipping</strong>: 18 apps</li><li><strong>Inventory sync</strong>: 13 apps</li><li><strong>Dropshipping</strong>: 7 apps</li></ul><p>This emphasis is no coincidence. Italian merchants often sell abroad and need reliable logistics, compliance with EU regulations, and tools for cross-border trade.</p><p>A smaller but notable vertical is <strong>product feed integrations</strong>. Around 7 apps focus on generating and optimizing feeds for marketplaces and comparison engines such as Idealo, Trovaprezzi, and Drezzy — showing how Italian developers address local sales channels that global apps often overlook.</p><p>Other categories — such as marketing, analytics, and loyalty — are present but less developed compared to other markets. This creates an opportunity for developers to branch out beyond logistics.</p><h3>Patterns in Descriptions</h3><p>Looking at how Italian developers describe their apps reveals recurring themes. Terms like <em>spedizione</em> (shipping), <em>fatturazione</em> (invoicing), <em>gestione ordini</em> (order management), and <em>fidelizzazione</em> (loyalty) appear frequently.</p><p>The language dimension adds another layer. According to the dataset, <strong>42 apps are localized into more than one language</strong>, while <strong>78 remain monolingual</strong>. Within those, <strong>29 are Italian-only</strong>, typically focused on compliance, invoicing, and other domestic needs. The other <strong>49 are English-only</strong>, usually in categories with international appeal like shipping, dropshipping, and multi-language. This makes language less about translation and more about positioning — developers are choosing the audience they want to serve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y7zTY0mq0BcP9Odfttdb6g.png" /><figcaption>Chart displaying the app description language</figcaption></figure><p>Descriptions overall reinforce the idea that many apps present themselves as practical problem-solvers for day-to-day merchant challenges, rather than purely growth-driven tools.</p><h3>Built for Shopify</h3><p>Only <strong>8 Italian apps</strong> currently carry the badge, while 112 do not. Interestingly, 7 of those certified apps come from <strong>Nextools</strong> — a developer that has emerged as a leader in the Italian ecosystem. Earlier this year, they also won the <a href="https://nextools.tech/supaeasy-wins-shoptank-at-shopify-edition-25/"><strong>Shoptank event</strong></a><strong> at Shopify Editions.dev</strong> with their SupaEasy app, further cementing their visibility on a global stage. This shows both the opportunity and the challenge: while Nextools sets the standard, other developers still have room to catch up in terms of quality and recognition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UkbY90V7yY8HfYab3Zax8Q@2x.png" /><figcaption>Built for Shopify introduction from the Shopify website</figcaption></figure><h3>Conclusion</h3><p>Italy’s Shopify app ecosystem is smaller than global hubs, but it is active, diverse, and growing. The apps reflect both universal merchant needs (shipping, inventory, dropshipping) and uniquely Italian priorities (electronic invoicing, fiscal compliance, loyalty in fashion).</p><p>One detail that stood out to me is how the ecosystem feels almost <em>bilingual by design</em>. Some apps are fully localized, while others deliberately remain monolingual — <strong>29 apps written only in Italian</strong> to meet domestic requirements, and <strong>49 only in English</strong> to reach an international audience. It’s a strategic choice, and a reminder that localization isn’t just about adding translations, but about deciding who your app is really for.</p><p>For me, mapping this ecosystem was rewarding beyond the numbers. I recognized familiar names — agencies I’ve worked alongside, indie developers I’ve met — and it underscored how connected this community is.</p><p>If you’re a Shopify app developer in Italy, I’d love to hear your experience. How do you see the ecosystem evolving? And if you’re an international developer, what would make you consider expanding into the Italian market?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=203545ae26f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Same course, different takeaways: a reflection on learning in tech.]]></title>
            <link>https://medium.com/@rferrandino.tokyo/same-course-different-takeaways-a-reflection-on-learning-in-tech-579119e91d67?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/579119e91d67</guid>
            <category><![CDATA[shopify]]></category>
            <category><![CDATA[devlife]]></category>
            <category><![CDATA[continuous-learning]]></category>
            <category><![CDATA[shopify-development]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Tue, 24 Jun 2025 10:07:36 GMT</pubDate>
            <atom:updated>2025-06-24T10:07:36.717Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*arCBL4q19snqW7eC" /><figcaption>Photo by <a href="https://unsplash.com/@brett_jordan?utm_source=medium&amp;utm_medium=referral">Brett Jordan</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Today I started watching a course on building Shopify apps.</p><p>It’s designed for developers who want to follow best practices, but even though I’ve worked with Shopify for years (mostly in storefront development using Liquid and JavaScript), I found myself researching what others might consider foundational.</p><p>The first video already goes deep into the boilerplate structure, touching on topics like serverful vs serverless models, tunneling for embedded apps, Prisma databases in dev mode, and the setup of developer tooling.</p><p>I’d heard some of these terms before, in blog posts, in docs, in dev banter, but I never really paused to understand what they exactly are and how they fit together. And when you’re delivering features for projects on a tight deadline, the incentive to pause is small.</p><p>Now that I’m between jobs and waiting for my visa, I finally have the headspace to explore areas of the Shopify ecosystem that aren’t part of my day to day role, and I’m realizing how many foundational layers I’ve only skimmed until now.</p><p>This course, aimed at app developers, doesn’t assume you know everything. But it <em>does</em> assume you’re willing to fill in the blanks.</p><p>One thing that’s made this process smoother: ChatGPT. It’s been surprisingly helpful in filling in the gaps, whether I’m asking for a 101 explanation of tunneling or trying to understand how Next.js + Prisma setup differs from Node + MongoDB orRemix.</p><p>It’s helped me not only find answers faster, but also ask better questions. Which is half the battle when you’re navigating unfamiliar territory.</p><h3>Everyone Brings a Different Stack of Experience</h3><p>I’ve seen the same pattern in language learning (I’m studying Japanese). Two people might read the same textbook, but their takeaways will be wildly different based on their background.</p><p>Same goes for developers.</p><p>You can go through the same course, same material, same video and still walk away with a completely different understanding, depending on your foundation.</p><h4>Slower Isn’t Worse</h4><p>So here’s what I’m holding onto:</p><ul><li>You don’t know what you don’t know — until you need to.</li><li>Revisiting foundations isn’t a setback, it’s an investment.</li><li>Learning slowly, but with context, is still progress.</li></ul><p>I’m not rushing through this course. I’m building a scaffolding that will support the kind of developer I want to become. One who understands what’s under the hood, not just what’s visible on the storefront.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=579119e91d67" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Shopify Catalog to ChatGPT traffic: how AI is rewriting SEO]]></title>
            <link>https://medium.com/@rferrandino.tokyo/from-shopify-catalog-to-chatgpt-traffic-how-ai-is-rewriting-seo-c33620b471c1?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/c33620b471c1</guid>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[shopify]]></category>
            <category><![CDATA[shopify-catalog]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Wed, 18 Jun 2025 05:58:11 GMT</pubDate>
            <atom:updated>2025-06-18T05:58:11.444Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*erhFP94ARb6GnrXC" /><figcaption>Photo by <a href="https://unsplash.com/@lureofadventure?utm_source=medium&amp;utm_medium=referral">Ali Kazal</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>When I looked at my travel blog analytics recently, I noticed something unexpected: a visit referred by ChatGPT.</p><p>It caught my eye not just because it was unusual, but because it felt like a glimpse into what’s already happening.</p><p>The way people search is shifting. Search Engine Optimization (SEO), once centered solely around Google, is evolving in response to the growing influence of AI tools like ChatGPT, Perplexity, and Copilot.</p><p>In this article, I’m building on insights shared by others in the industr, particularly those closely following Shopify’s Catalog API and applying them to areas I personally care about, such as Shopify development and content discoverability for blog creators.</p><h3>The new search interface: AI</h3><p>For years, getting organic traffic meant optimizing for Google’s crawlers: writing helpful content, filling in your meta descriptions, and building backlinks.</p><p>But now, people are asking AI tools for answers, and those tools are beginning to generate responses, pulling data from the web in real time, sometimes even appearing before traditional links, like Google’s AI-generated overviews now displayed at the top of many search results.</p><p>That means:</p><ul><li>Your content may show up <strong>inside</strong> a chatbot reply, not just as a link</li><li>You might get traffic from ChatGPT or Perplexity even if you’re not ranking on Google</li><li>The way your content is structured (semantic data, headings, clarity) matters more than ever</li></ul><p>For e-commerce, this is a huge opportunity. And a big challenge.</p><h3>Shopify and the future of product discovery</h3><p>One of the most exciting developments in this space is Shopify’s new <a href="https://changelog.shopify.com/posts/introducing-shopify-catalog"><strong>Catalog</strong></a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kNJysL5k9TPBbqU7i8LKHA.png" /><figcaption>Screenshot from the demo of Shopify Catalog on <a href="https://namastudio.it/en/blogs/blog/shopify-editions-summer-2025-horizon">Editions 2025</a> website</figcaption></figure><p>Currently available only to a limited group of pilot partners, this feature gives AI agents like Perplexity access to product data across all Shopify stores. We’re not talking about scraping pages or guessing what’s in stock. This is a machine-readable feed of:</p><ul><li>Product titles</li><li>Prices and stock levels (Shopify will update these in real-time while other data are updated less frequently)</li><li>Enriched product attributes (materials, dimensions, etc.)</li><li>Taxonomy and metadata</li></ul><p>This API will let large language models surface actual products, and probably in the future even enable checkout across multiple brands.</p><p>Shopify’s Catalog relies on a standardized taxonomy to classify products across stores. Developers can view and implement this taxonomy via <a href="https://shopify.github.io/product-taxonomy/">Shopify’s GitHub page</a>, ensuring consistency in how items are categorized and discovered. And merchants can assign categories using the <a href="https://help.shopify.com/en/manual/products/details/product-category">product category field in the Shopify admin</a>, which supports AI-assisted suggestions.</p><h3>From search to suggestion: your product as the answer</h3><p>What does this shift mean for e-commerce store:</p><ul><li>Your product could be recommended directly inside an AI chat</li><li>Relevance and clarity in product data will rule</li><li>A unified cart for multiple Shopify stores is now a realistic future</li></ul><p>Access to Shopify Catalog is currently invite-only, with the first pilot partner being <strong>Perplexity AI</strong>. More will come soon, according to Vanessa Lee, VP Product at Shopify.</p><h3>Advice for merchants and developers</h3><p>If you develop for Shopify or maintain a store, this is the time to take action. Gavin Mckew in his <a href="https://sherocommerce.com/shopify-catalog-api/">deep dive on the Catalog API</a> provides a cool checklist. My relevant points that I will explore, take action on or recommend are as follows:</p><h4>For merchants</h4><ul><li>Treat product content like <strong>performance marketing</strong>: clear, benefit-driven, and keyword-rich</li><li>Replace custom fields with <strong>standard metafields</strong> so your product data aligns with Shopify’s Catalog API</li><li>Populate <a href="https://apps.shopify.com/shopify-knowledge-base"><strong>Knowledge base</strong></a><strong> metaobjects</strong> to support AI tools with accurate and contextual product info</li></ul><h4>For developers</h4><ul><li>Audit and migrate legacy schema to align with Shopify’s structured data model</li><li>Build storefront extensions or UI components that <strong>surface rich product data</strong> in a machine-readable way</li><li>Implement analytics to monitor AI-driven referrals and track visibility in tools like Copilot or ChatGPT</li></ul><h4>For content creators</h4><p>Whether you run a blog or documentation site:</p><ul><li>Add structured data and schema.org metadata to posts</li><li>Write with clarity and hierarchy: <strong>scannable headings, bullet points, and plain language</strong> help</li><li>Assume AI agents are your first reader — <strong>optimize for comprehension, not just keywords</strong></li></ul><h3>Catalog API: What’s Still Evolving</h3><p>While the Catalog API opens major opportunities, there are limitations developers should be aware of:</p><ul><li><strong>Localization and Shopify Markets support</strong>:<br> AI agents like Perplexity can pass user context (locale, currency), but when that’s missing, the API currently defaults to US settings. Support for multi-market contexts is expanding</li><li><strong>B2B workflows</strong>:<br> The Catalog API is currently focused on consumer discovery. B2B support is not yet available and will likely follow a different structure and timeline</li><li><strong>Data abuse concerns</strong>:<br> Shopify is limiting access to <strong>invite-only partners</strong>, who are vetted based on GMV upside and risk of misuse. Live inventory and price data are only unlocked for <strong>trusted integrations</strong> to avoid abuse like price-matching bots or checkout scraping</li></ul><h3>For content creators: how to stay relevant in an AI cearch world</h3><p>Whether you run a blog, a documentation site, or a personal portfolio, here’s what I’m starting to prioritize to stay visible as the search landscape shifts. Of course if I am writing a <a href="https://nottheusualtourists.com/">travel blog article</a> for the pleasure of telling my story, I still go with my writing flow, but I will optimize the alt text of the images to make it more discoverable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PR91-PETIm03cbrrXZoqcw.png" /><figcaption>Homepage of my blog Not the usual tourists</figcaption></figure><ul><li><strong>Write like you’re answering a question</strong><br>AI tools love clarity. Make your headers specific (e.g., “How to Prepare for Umeshu Season in Japan”) and your paragraphs concise.</li><li><strong>Use structured metadata</strong><br>Add schema.org markup, article tags, and semantic HTML whenever possible.</li><li><strong>Keep your information current and factual</strong><br>AI agents pull trusted, up-to-date data. Refresh your content periodically, especially product pages or articles with changing context (like travel guides or SEO how-tos)</li><li><strong>Optimize for AI summaries and previews</strong><br>The first 2–3 sentences of your article matter more than ever, they’re likely to be quoted or summarized. Make them clear, unique, and answer-focused</li><li><strong>Think beyond your site</strong><br>Tools like Perplexity and ChatGPT don’t just link, they summarize. Include internal links and contextual references that help the AI “connect the dots” across your own content.</li></ul><h4>For portfolios</h4><ul><li>Add job titles, skills, and cities in a scannable format</li><li>Use project summaries with bullet points and clear titles (e.g., “Hydrogen Store Redesign for Ethical Brand”)</li><li>Include a small FAQ or “About this project” section for context AI tools might lift</li></ul><p>These are small shifts, but together they’ll help your content surface in new, conversational ways, whether from a blog, product catalog, or personal website.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c33620b471c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Notes from Toronto: Shopify Editions.dev and the power of a connected ecosystem]]></title>
            <link>https://medium.com/@rferrandino.tokyo/notes-from-toronto-shopify-editions-dev-and-the-power-of-a-connected-ecosystem-f0ae6f610e91?source=rss-1317749d29b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0ae6f610e91</guid>
            <category><![CDATA[shopify]]></category>
            <category><![CDATA[developer-event]]></category>
            <category><![CDATA[devlife]]></category>
            <dc:creator><![CDATA[Rossella Ferrandino]]></dc:creator>
            <pubDate>Wed, 11 Jun 2025 04:51:07 GMT</pubDate>
            <atom:updated>2025-06-11T12:22:23.672Z</atom:updated>
            <content:encoded><![CDATA[<p>From May 27th to June 1st I was in Toronto, where I attended Shopify <a href="https://editions.dev/pages/editions-dev">Editions.dev</a>, a multi-day event that brought together developers, agencies, product teams, and partners from around the world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KtGc5e_d2ffNvcEuJga2Cw.jpeg" /><figcaption>Toronto landscape</figcaption></figure><p>For those less familiar with Shopify: it’s a commerce platform used by millions of brands to run their online stores. From small founder-led businesses to major global retailers, Shopify powers everything from product pages to checkout, shipping, POS, and more. And twice a year, the company releases Editions, a set of product updates and platform improvements. For the second year in a row, they decided to bring the developer community together in person — and the result was a packed, energetic week of learning and connection.</p><h3>A week built around learning, not just announcements</h3><p>The event wasn’t just a single conference. While Editions.dev itself was a full-day event (plus an opening night party), the entire week was filled with activity. Other companies and communities in the ecosystem hosted their own events around it: <a href="https://techtonic.heymantle.com/">Techtonic</a> (organized by <a href="https://heymantle.com/?ref=techtonic">Mantle</a>) offered agency and app-focused talks, <a href="https://www.meetup.com/spdy-meetup-toronto/">SPDY meetup Toronto</a> hosted a web performance meetup with the Shopify Growth team ad Mozilla HQ, and there were smaller community gatherings like the<a href="https://shopdevalliance.com/"> ShopDev Alliance</a> meetup. It felt like a city-wide celebration of the Shopify developer ecosystem.</p><p>And of course, the main Editions.dev day: 2000 developers packed into a huge venue, attending live workshops, panel discussions, networking, and browsing booths from different product teams.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9VWAgjdnkeY5f4fPqYfQ5Q.png" /><figcaption>Schedule of the events I attended in the week</figcaption></figure><p>As a theme developer working with Shopify for the past few years, I’ve followed the evolution of the platform closely. But this week gave me something different: a sense of what it means to be part of a wider, collaborative ecosystem. From freelancers to app developers, from core product engineers to agency leads, everyone there was openly sharing what they’d learned, what they’d built, and what’s next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2WVSmapbpHSjBnchSC4c4A.jpeg" /><figcaption>ShopDev Alliance group photo</figcaption></figure><h3>Liquid Hackathon, Sidekick and the future of theme development</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M_gPrn-JSzWfa6zR8yy0FA.png" /></figure><p>One of the most fun parts of the week was the Liquid Hackathon at the Editions opening party, where participants had 20 minutes to build something creative on top of Shopify’s new Horizon theme. Watching it live was a mix of admiration and curiosity — so I decided to <a href="https://www.linkedin.com/posts/rossella-ferrandino_my-solo-take-on-the-editionsdev-liquid-activity-7338100359829090304-rePU?utm_source=social_share_send&amp;utm_medium=member_desktop_web&amp;rcm=ACoAAAVXQjoBkjbJSKuS2JjcjzoIx8-H-0WlalY">try the challenge on my own</a> afterwards.</p><p>With Sidekick (Shopify’s new AI assistant), I asked it to build a “Shop the Look” section. In minutes, it returned a fully working block: JavaScript, scoped styles, and usage instructions included. I also had it help me set up a color scheme and used ChatGPT to generate placeholder product images. It wasn’t perfect, but it was fast and surprisingly coherent.</p><p>More than anything, it reminded me that the tools are evolving.</p><p>And so are the ways we build.</p><h3>Shopify in the real World: visiting Toronto stores</h3><p>Outside of the venue, I visited a few physical stores powered by Shopify: Cozey, Mejuri, and Kotn.</p><p>Cozey’s storefront had a cozy, minimalist aesthetic, complete with an in-store planner for customizing sofas. I’m currently in the market for one, and their product page stood out, especially the feature showing real people sitting on the sofa, with their height listed. It’s such a thoughtful way to help users picture the product in context.</p><p>At Mejuri, I had a chance to view their Shopify POS in action. A staff member helped me create a draft order and sent it to my email. Tablets were available to browse specs and prices, and QR codes let customers scan and shop directly. Their site, built with Shopify’s headless framework Hydrogen, feels sleek and dynamic. Their store locator and appointment scheduling experience is one of the best I’ve seen.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*Bu5u4VqEc9NtWxlkiYaLyQ.jpeg" /><figcaption>Mejuri store</figcaption></figure><p>Kotn, meanwhile, is a brand I’ve followed for years due to their sustainability focus. I didn’t chat much with the staff (the store was quite busy), but knowing that one of the co-founders is also a product director at Shopify made it all the more interesting. Their site is polished and aligned with best-in-class practices (no surprise there).</p><h3>Merchants in the spotlight</h3><p>One of the most charming touches during Editions.dev? The snacks and drinks were all from brands running on Shopify. QR codes on every can or bar let attendees explore their stores, making the merchant experience tangible and central. It was a small detail, but one that really drove home the connection between what we build as developers and how it shows up in real life.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bioFvzNEvqqpz-xspEgZfg.jpeg" /><figcaption>Each of the snacks in the cafeteria had barcodes to the stores (powered by Shopify)</figcaption></figure><h3>Final Thoughts</h3><p>There was a lot to take in during the week, from admin-intents and AI-generated theme blocks to conversations with agency owners and Shopify employees I’ve followed online for years. But what I’m taking with me is more than notes. It’s energy, perspective, and a deeper appreciation for the people behind the platform.</p><p>Shopify’s product philosophy is clearly evolving. So are the tools, and so is the community. And after a week like this, it’s hard not to feel excited about what comes next.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0ae6f610e91" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>