<?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 Ilinca Apolzan on Medium]]></title>
        <description><![CDATA[Stories by Ilinca Apolzan on Medium]]></description>
        <link>https://medium.com/@fluturelle?source=rss-f1ef3a81515e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*MQb_3t9o7CkXMqYQX3Wc2w.jpeg</url>
            <title>Stories by Ilinca Apolzan on Medium</title>
            <link>https://medium.com/@fluturelle?source=rss-f1ef3a81515e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 16:56:12 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@fluturelle/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[They took away Figma and nobody told the content designer]]></title>
            <link>https://medium.com/@fluturelle/they-took-away-figma-and-nobody-told-the-content-designer-d83dd96622aa?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d83dd96622aa</guid>
            <category><![CDATA[content-design]]></category>
            <category><![CDATA[figma]]></category>
            <category><![CDATA[design-tools]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Fri, 08 May 2026 21:10:39 GMT</pubDate>
            <atom:updated>2026-05-08T21:10:39.865Z</atom:updated>
            <content:encoded><![CDATA[<h4>When your company switches to code-based prototyping and your job becomes 11 git commands in a trench coat.</h4><figure><img alt="Illustration of a content designer’s avatar standing in a dark code editor environment, looking around confused with one hand out in a ‘where is everything?’ gesture. They’re surrounded by TypeScript code and developer tools, searching for their design workspace that’s no longer there." src="https://cdn-images-1.medium.com/max/1024/1*J5zolwDyA8_TtSH7egMy0Q.png" /><figcaption>When you open Cursor and Cursor opens you</figcaption></figure><p>Last week a UX designer sent me a Slack message. It said:</p><p>/download ux.designer/notification-preferences</p><p>That was the whole message. No link. No screenshot. No “hey, could you take a look at the copy on this flow?” Just… a terminal command?!</p><p>I stared at it the way you stare at IKEA furniture instructions that start at step 7. <em>I know something is supposed to happen here. I just don’t know what.</em></p><p>1 week later, I’d understand that this one Slack message contained — compressed like a zip file of frustration — every single thing that had broken in my workflow since our company replaced Figma with a code-based prototyping tool. But last week? I just thought I was bad at git.</p><h3>The before: when Figma was the group chat</h3><p>Here’s what content review used to look like. A designer would send me a Figma link. I’d click it. I’d see the design — the screens, the flow, the hierarchy. I’d click on a text layer and type better copy. I’d leave a comment: <em>“This error message is too vague — what’s the actual failure reason?”</em> The designer would see it in seconds, in context.</p><p><strong>Total time: maybe 15 minutes.</strong> Total tools required: one browser tab.</p><p>This isn’t nostalgia — it’s a benchmark. Because what replaced it is… a lot.</p><h3>The after: when prototyping became programming</h3><p>Our company moved to a tool I’ll call “The Platform” — a code-based prototyping system where designers build in React (TSX files), work in Cursor, and collaborate via git. It’s a powerful tool. For designers who code, it’s probably great.</p><p>I am not a designer who codes…yet. I’m a content designer. I think in user flows, language systems, and the moment someone reads a screen and either trusts it or doesn’t.</p><p>But in The Platform, there are no screens to look at — not without cloning a repo, checking out a branch, running yarn, and navigating to the right localhost URL. There&#39;s no text to click on — there&#39;s &lt;Text textToken=&quot;paragraph/m&quot; style={{ textAlign: &#39;center&#39; }}&gt; buried on line 347 of an 800-line TypeScript file.</p><p>And there are definitely no comments.</p><p>The designer’s path to share their work: type /share. Done.</p><p>My path to contribute back:</p><ol><li>Figure out which repo</li><li>Clone it</li><li>Find the right branch</li><li>Understand the folder structure</li><li>Discover that app/shared/ is read-only (the hard way)</li><li>Create my own branch</li><li>Make text changes in .tsx files</li><li>Commit</li><li>Push</li><li>Create a PR</li><li>Wait for merge</li></ol><p><strong>That’s 11 steps requiring git knowledge, versus 1 command.</strong></p><figure><img alt="creenshot-style image of a Cursor IDE window with code on the left showing a git merge conflict over the words ‘click’ versus ‘tap’, and an AI chat panel on the right showing an increasingly frustrated conversation where the user tries to explain they only changed one word while the AI keeps suggesting to run yarn." src="https://cdn-images-1.medium.com/max/1024/1*9bhQdYRVRy4XcjtvcmFWZA.png" /><figcaption>All I wanted was to change one word</figcaption></figure><p>The collaboration model didn’t become harder. It became one-directional.</p><h3>The week of staring at symptoms</h3><p>Here’s where I could have done what most people do: complain, adapt, move on. Instead — and I’m still not sure if this was ADHD hyperfocus or professional stubbornness — I spent an entire week just <em>listing everything that felt wrong.</em></p><p>Not fixing anything. Just documenting the pain.</p><p>And the list was… long.</p><p>The tool wasn’t built for my role. Copy trapped inside code. No visual context. Asymmetric collaboration. Dead-end folders. Disconnected tools. Eight problems, all feeling equally urgent.</p><p>Which one do I fix first? Do I learn git better? Do I ask designers to send me screenshots? Do I build a script?</p><p>I was treating every problem as independent. And that was the mistake.</p><h3>The breakthrough: 8 symptoms, 2 causes</h3><p>Midway through the week, I sat down and drew a dependency map. Not a fancy one — just arrows on a screen. “This problem causes that problem. That one causes these 3.”</p><p>And suddenly the whole thing collapsed into something much simpler.</p><blockquote><strong><em>I didn’t have 8 problems. I had 2 root causes and a cascade.</em></strong></blockquote><p><strong>Root cause 1:</strong> The tool was built for developers, and nobody adapted it for content collaboration.</p><p><strong>Root cause 2:</strong> No content view exists — no way to see, organize, or work with copy outside of raw code.</p><p>Everything else — the git friction, the disconnected tools, the dead-end folders, the lost feedback — was downstream. Symptoms, not causes.</p><p>This was the moment I stopped trying to fix eleven steps and started thinking about what to build instead.</p><h3>What I actually lost (it wasn’t Figma)</h3><p>Here’s the thing that took me the longest to see, and it sounds obvious in hindsight:</p><p><strong>I didn’t lose a design tool. I lost a shared workspace.</strong></p><p>Figma wasn’t special because it was Figma. It was special because it was the <em>one place</em> where a content designer and a UX designer could both see, edit, and discuss the same thing. I could look at the design. I could edit the words. I could leave a comment that said <em>“this works”</em> or <em>“consider option B”</em> or <em>“what triggers this state?”</em></p><p>5 feedback modes:</p><ul><li><strong>Direct edit</strong> — change the text</li><li><strong>Approval</strong> — this is good, don’t touch it</li><li><strong>Annotation</strong> — here’s context you should know</li><li><strong>Suggestion</strong> — consider these alternatives</li><li><strong>Question</strong> — I need information to decide</li></ul><p>In the new code-based world, I had exactly <strong>1</strong> feedback mode: edit the code. That’s it. I couldn’t say “maybe.” I couldn’t say “option A or B?” I couldn’t say “this is great, I reviewed it.” Every piece of feedback — whether it was a definitive rewrite, a tentative suggestion, or just a question — had to be flattened into “change this line of TypeScript.”</p><blockquote><em>The tool didn’t just make my workflow harder. It eliminated my ability to have a conversation about copy.</em></blockquote><p>That hit different.</p><h3>What I’m doing about it</h3><p>I’m building a content review system. Not a Figma replacement — something designed specifically for content collaboration in a code-native workflow. 3 lightweight pieces:</p><p><strong>A shared review interface. </strong>Both the designer and I use the same tool. I open the prototype, click on any text, and rewrite it, flag it, or propose alternatives. The screen updates instantly with my changes. The designer opens it and sees my work — what I changed, where I left questions, which alternatives to pick between.</p><p><strong>A decision record that doesn’t disappear.</strong> Every edit, every alternative, every question — saved alongside the prototype. When someone asks “why does it say this?” six weeks from now, the answer is there.</p><p><strong>Writing guidelines baked into Cursor</strong> — our content standards as a rules file in the repo. Every time the designer asks Cursor for help with copy, it already knows our tone, terminology, and conventions. Better copy at the source means less retrofitting for me.</p><p>I’ve been vibe coding the whole thing with Claude Code. A content designer. Building a design tool. Because nobody else was going to.</p><figure><img alt="A content review tool built by a content designer showing annotations for a notifications prototype — edit proposals, A/B alternatives, and status tracking — all without touching code" src="https://cdn-images-1.medium.com/max/548/1*taWjwgk4-Yi8jsB6dtPlFw.png" /><figcaption>Zero lines of TypeScript were harmed in the making of this copy review</figcaption></figure><p>I’ll write about the build process in the next article.</p><h3>If you’re in this situation</h3><p>Maybe your company didn’t switch away from Figma. But maybe you’ve experienced some version of this — a tool change that left you without a shared workspace, a collaboration model that wasn’t designed for your role, a workflow that made your expertise invisible.</p><p>A few things I wish I’d known at the start:</p><p><strong>Map the dependencies before you fix anything.</strong> I wasted the first days treating every symptom as a separate fire. The moment I drew the dependency map, I went from “8 impossible problems” to “2 root causes I can actually address.” Your cascade might look different, but it’s almost certainly a cascade.</p><p><strong>Name what you lost, not what you miss.</strong> “I miss Figma” isn’t actionable. “I lost the ability to annotate, suggest alternatives, and document decisions” is. The specific losses are the design brief for what to build.</p><p><strong>The tool is probably fine — for the people it was designed for.</strong> My frustration wasn’t with The Platform. It’s a good tool for designers who code. My frustration was that nobody asked what content designers would need when the shared workspace went away. That’s an organizational gap, not a tool flaw.</p><p><strong>If the infrastructure doesn’t exist, you might have to build it.</strong> I resisted this for a while. I’m a content designer, not an engineer. But the alternative was accepting a permanently broken workflow — and that wasn’t an option. Sometimes the most content-design thing you can do is build the tool that makes content design possible.</p><h3>What’s next</h3><p>I’m almost 2 weeks into this. The review canvas exists — still rough, still WIP, but <em>functional</em>. I can click on copy in a prototype and edit it. I can flag violations. I can leave alternatives. It’s not pretty yet, but it works.</p><p>If you’ve ever built something because nobody else would — you know the feeling. It’s equal parts <em>“this is ridiculous”</em> and <em>“wait… this is actually working.”</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d83dd96622aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Plot twist: your best portfolio tool is your ChatGPT history]]></title>
            <link>https://medium.com/design-bootcamp/plot-twist-your-best-portfolio-tool-is-your-chatgpt-history-1aab638c16dd?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/1aab638c16dd</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[second-brain]]></category>
            <category><![CDATA[ux-writing-portfolio]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Mon, 20 Apr 2026 21:43:52 GMT</pubDate>
            <atom:updated>2026-04-20T21:43:52.339Z</atom:updated>
            <content:encoded><![CDATA[<h4>How your everyday AI chats are quietly building your second brain</h4><figure><img alt="Illustration of a person at a desk surrounded by chaotic chat bubbles on one side and neatly organized rows on the other, with a lightbulb above their head, representing turning scattered thoughts into a structured system." src="https://cdn-images-1.medium.com/max/1024/1*vumvyw3PX3thHVvDFN0F6g.png" /><figcaption>From “what did I even do?” to… oh. Right. All of this.</figcaption></figure><p>I was preparing for job interviews.</p><p>Simple enough goal. Except when I sat down to actually think through my work — the projects, the decisions, the stuff I was genuinely proud of — I hit a wall. Not because I hadn’t done anything. I had done plenty. It was more that my work lived in fragments: in Figma files I cannot find anymore, in Slack threads nobody scrolled back to, in decisions I’d made and moved on from without writing down.</p><p>I needed to tell coherent stories about my work. Instead, I was trying to reconstruct a jigsaw puzzle with half the pieces missing.</p><p>Then I had a slightly unhinged idea. What if I asked ChatGPT itself to tell me what we’d talked about for 2 years? To go back through everything and reconstruct my own work history for me. Slightly weird. Absolutely worth trying.</p><h3>The archive that already existed</h3><p>For about 2 years, I’d been using ChatGPT for basically everything at work.</p><p>Not just the big, strategic stuff. Everything. Tiny bits of copy. Rewrites. Research rabbit-holes. Brainstorming sessions at 10pm. Decisions I couldn’t make on my own. Project naming. UX writing. Content strategy thinking. Sometimes just talking through something that was bothering me about a brief.</p><p>If it crossed my desk, I probably chatted ChatGPT about it at some point.</p><figure><img alt="Screenshot of a ChatGPT interface showing a long sidebar of conversations and Projects with varied names, reflecting a large volume of ongoing work over time." src="https://cdn-images-1.medium.com/max/1024/1*x27TU530NKT7JPSLtcvgKg.png" /><figcaption>Not a system. Just… everything I ever threw at ChatGPT, quietly piling up over time.</figcaption></figure><p>That means something slightly wild: ChatGPT had seen more of my day-to-day work than any colleague ever had. More than my manager. More than any performance review document. It had been there for the messy middle — the rough drafts, the uncertainty, the rewrites, the “wait, is this actually working?” moments that never make it into a polished portfolio.</p><p>No single document had that. ChatGPT did.</p><p>When I realized that, something clicked. The archive already existed. All those chats, across all those Projects — that <em>was</em> a record of my working life. I just hadn’t thought of it that way before.</p><h3>The experiment (which took 2 days and a lot of patience)</h3><p>So I asked ChatGPT to go back through all our old conversations — across different Projects — and identify every initiative, project, and meaningful piece of work it could find.</p><figure><img alt="Screenshot of a ChatGPT conversation showing a long, detailed prompt asking the model to review past chats and extract projects into a structured table for interview preparation." src="https://cdn-images-1.medium.com/max/862/1*7vUHTgFMYSBsLS_cp2Fv-g.png" /><figcaption>This was the moment it shifted from idea to experiment. I asked ChatGPT to go back through everything and tell me what I’d actually worked on.</figcaption></figure><p>The initial goal was modest: build a table I could use for interview prep. Something with the basics. Projects, key decisions, impact, behaviors.</p><p>What followed was genuinely kind of chaotic.</p><p>The first outputs were a start, but a weak one. Some entries were too vague to be useful. Others grouped things in ways that didn’t reflect how the work actually felt. A few projects were missing entirely — either buried in a chat context the model couldn’t reach, or summarized so loosely they’d lost their point.</p><figure><img alt="Screenshot of a spreadsheet showing an early, partially structured list of projects with columns like Project, Focus, Actions, Collaboration, and Impact, with dense and somewhat inconsistent entries." src="https://cdn-images-1.medium.com/max/581/1*UJv3W9Ks0AbTxH2M-o8ixw.png" /><figcaption>The first pass. Useful, but rough — projects flattened, context missing, and a lot of the real story still buried.</figcaption></figure><p>I had to iterate. A lot. I switched models because some were better than others for certain parts of the extraction. I re-prompted. I corrected. I pushed back when the output felt generic or missed the nuance of what the work actually was.</p><p>At some point I stopped counting how many rounds we’d done.</p><h3>Teaching the system to see my work properly</h3><p>This is the part I didn’t expect to matter as much as it did.</p><p>ChatGPT didn’t automatically understand how to interpret content design work. It would extract something like “wrote copy for onboarding flow” and completely miss that the real story was about the strategic decision behind it — why that framing, why that moment in the journey, what changed because of it.</p><p>I had to teach it how to look.</p><p>What counts as a real project versus a one-off task. How to identify when something was systemic rather than just local. How to label impact in a way that was honest rather than inflated. How to distinguish between “I made a thing” and “I shaped a direction.”</p><p>That wasn’t frustrating, exactly. It was revealing. Guiding the model through that process forced me to articulate things I’d never had to put into words before. It wasn’t just extraction — it was interpretation. And I was doing as much of the thinking as the tool was.</p><h3>The structure that made it useful</h3><p>In a separate chat (this matters — keeping the database design away from the extraction kept things cleaner), I worked out what the actual structure should be.</p><p>Not just a flat project list. Something that could do four different jobs at once:</p><ul><li><strong>Story</strong> — Situation, Task, Action, Result. The skeleton of a good interview answer, already written.</li><li><strong>Evidence</strong> — Artifacts, deliverables, and before/after impact examples. Not vague claims like “improved UX copy” but specifics like <em>“Before: Generic error message → After: Actionable copy clarifying exactly what was blocked and what to do next.”</em> The kind of detail that makes an interviewer lean forward.</li><li><strong>Retrieval</strong> — This is the column most of us wouldn’t think to add. Every entry has a Chat Reference Keyword: a short search phrase I can paste back into ChatGPT to find the original conversation. Things like <em>“search: Alert banner clarity Sofort”</em> or <em>“search: AI content workflow system Payments.”</em> It’s a breadcrumb trail back to the raw thinking — the messy, unpolished version that the CSV summary can’t fully capture.</li><li><strong>Reuse</strong> — Tags, skills, and behaviors that recur across projects. Useful for spotting your own patterns. Turns out I keep solving the same kinds of problems in different contexts, which says something about how I work that no single project story ever could.</li></ul><figure><img alt="A cropped spreadsheet-style table showing four content design projects with columns for project details, context, actions, results, impact examples, and tags. The table is structured to capture not just project summaries, but reusable stories with evidence and searchable reference keywords." src="https://cdn-images-1.medium.com/max/1024/1*305cMayhReaLONIq-saYDQ.png" /><figcaption>A structured project database with story, evidence, and retrieval layers — so I can actually find and reuse my work later.</figcaption></figure><p>Once the structure was solid, I applied it to everything ChatGPT had recovered. That process eventually became a large CSV.</p><p>I know “large CSV” sounds deeply unsexy. But this particular CSV felt like finding a box of photos you’d forgotten you’d taken.</p><h3>The moment it became something bigger</h3><p>Somewhere in the middle of all of this, this exercise stopped being about interview prep.</p><p>It had started as: <em>how do I answer interview questions better?</em></p><p>It became: <em>how do I actually know what I’ve built over the last few years?</em></p><p>Those are different questions. The first one is about performance. The second one is about memory.</p><p>A polished portfolio doc is a highlight reel. It’s retrospective, curated, intentionally flattering. It’s useful, but it’s thin. It tells you what ended up being worth keeping, not how you got there.</p><p>What the CSV gave me was something closer to a working memory for my career. Not just “here are my best projects” but “here is how I think, what I care about, what kinds of problems I keep solving, and where I’ve grown.” That’s a different kind of useful.</p><p>It stopped being interview prep. It started being a <a href="https://medium.com/neuralnotions/andrej-karpathy-stopped-using-ai-to-write-code-hes-using-it-to-build-a-second-brain-instead-cddceadc5df5">second brain</a> for work.</p><h3>The real lesson (not the obvious one)</h3><p>The easy takeaway from this story is something like: <em>AI is good at helping you organize things.</em> That’s true, but it’s not interesting.</p><p>Here’s the thing that actually surprised me.</p><p>The reason this worked — the reason the archive had enough richness to be worth recovering — was the 2 years of ordinary, unremarkable, daily AI use that came before any of this. The tiny copy tweaks. The late-night brainstorms. The half-formed questions. The decisions I thought through out loud in a chat window instead of a notes doc nobody would ever read.</p><p>All of that everyday use created a trail. And the trail turned out to be the most valuable thing.</p><p>I used to think the value of using AI daily was speed. Faster output, more done, better first drafts. And yes, that’s real. But there’s a second kind of value I hadn’t seen: <em>memory</em>. The accumulation of small interactions, over time, creates something you can come back to. A record of how you work, not just what you produced.</p><p>If I’d used ChatGPT only for big, intentional projects, there would’ve been nothing to recover. The richness came from the everyday grain of it.</p><h3>A practical note, if you want to build something similar</h3><p>You don’t need 2 years of history to start. You just need to start thinking about your AI use as an archive, not just an output machine.</p><p>A few things that made my system actually useful, rather than just large:</p><ul><li><strong>Save artifacts and mention examples in your chats, not just summaries.</strong> A link back to the real thing is worth ten times a clean description of it.</li><li>*<em>Capture the why, not just the what.</em> The decision behind the work is usually more useful than the deliverable.</li><li><strong>Before building a polished portfolio doc, build a retrieval system.</strong> What helped me most wasn’t making something beautiful. It was making something I could actually search.</li><li><strong>Your AI chats may already be more of a work archive than you realize.</strong> If you’ve been using these tools daily for a while, the material is probably already there.</li></ul><p>I’ve since moved on to automating the next version of this with Claude and <a href="https://medium.com/autocomplete-real-world-ai/claude-cowork-turns-your-desktop-into-an-employee-cf70d3d68c5d">Cowork</a> — but that’s a different article.</p><p>For now, the thing I keep coming back to is this:</p><p>I thought I was using ChatGPT to get through my work. It turns out I was also building a record of how I work. Not on purpose. Not with any grand system in mind. Just by working on the small stuff, day after day, and using the tool that was there.</p><p>You know, nobody tells you to document your own growth while you’re busy doing the work. You’re expected to just… remember it. And then, when it’s time to apply for jobs, somehow perform a perfect highlight reel of the last years on demand.<br>That’s a lot to ask.<br>Your AI chats are one small way to work around that. Not for your employer, not for your next performance review — but for you! <br>Your story, your projects, your receipts!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1aab638c16dd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/plot-twist-your-best-portfolio-tool-is-your-chatgpt-history-1aab638c16dd">Plot twist: your best portfolio tool is your ChatGPT history</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I used AI as a creative director, not a shortcut]]></title>
            <link>https://medium.com/@fluturelle/i-used-ai-as-a-creative-director-not-a-shortcut-b188fefca558?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b188fefca558</guid>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Sat, 14 Mar 2026 22:07:09 GMT</pubDate>
            <atom:updated>2026-03-14T22:11:25.266Z</atom:updated>
            <content:encoded><![CDATA[<h4>And it made me think differently about what “making something” actually means.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qGHtTAQ2TKfdwP5hm-MEBQ.png" /></figure><p>My friend <a href="https://www.instagram.com/travelproof_tjark?utm_source=ig_web_button_share_sheet&amp;igsh=ZDNlZDc0MzIxNw==">Tjark</a> has a thing for brutalist buildings.</p><p>Not a casual appreciation — an actual thing. He travels, he photographs, he posts. Concrete towers in Eastern Europe. Soviet-era civic centres nobody asked for but somehow everyone remembers. The kind of buildings that make half the internet say “tear it down” and the other half say “this is the most beautiful thing I’ve ever seen.”</p><p>I’m firmly in the second half.</p><p>So a few weeks ago, deep in a Tjark-induced rabbit hole, I had an idea. What if I turned some of these buildings into paper collage cards? Soft, tactile, collectible. The heaviest architecture in the world, made to look like someone cut it out with scissors and glued it onto coloured card.</p><p>That was the brief. And I hadn’t opened a single tool yet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_6wNghZfAWAKhQub.png" /><figcaption>One of the first cards that clicked. The circular form felt almost too perfect for the paper collage treatment — all those repeating window units stacked around the rim like a crown.</figcaption></figure><p><strong>Here’s where it gets interesting.</strong></p><p>I’ve written before about breaking in new AI models — figuring out what they’re actually good at before trusting them with your work. And one thing I’ve learned is that the quality of what comes out is almost entirely determined by the clarity of what goes in.</p><p>Not prompt clarity. <em>Vision</em> clarity.</p><p>So before I opened Gemini, I knew exactly what I wanted:</p><ul><li>Paper collage aesthetic — layered, textured, tactile</li><li>Two-tone backgrounds — one colour for sky, one for ground</li><li>Muted, slightly chalky palette — nothing too saturated</li><li>Each building as the sole subject, no context, no people</li><li>A consistent card format — portrait, like a collectible</li></ul><p>I wasn’t asking AI to come up with the idea. I was asking it to execute one I already had.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5brWngGZsF2glCz9.png" /><figcaption>The most symmetrical card in the series. Those diamond skylights at the top are the kind of detail that makes you wonder who approved this and whether they were a genius or completely unhinged.</figcaption></figure><p>That distinction matters more than it sounds.</p><p>Most conversations about AI-generated art collapse into two camps: “AI is killing creativity” or “AI is just a tool.” Both feel a bit lazy to me. Because the real question isn’t whether AI made the image. It’s <em>who was doing the thinking.</em></p><p>When I fed reference photos of brutalist buildings into <a href="https://gemini.google/overview/image-generation/">Gemini’s Nano Banana 2</a> and asked it to reimagine them as paper collage cards, I went through probably 40 iterations before I got something I was happy with. Adjusting the texture. Pushing the shadows. Getting the colour splits right. Rejecting outputs that were technically impressive but aesthetically off.</p><p>That process felt like designing, not outsourcing.</p><p>One of the first cards that clicked. The circular form felt almost too perfect for the paper collage treatment — all those repeating window units stacked around the rim like a crown.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5y37mmLtoDOUMLiN.png" /><figcaption>This one looks like an architectural Tetris game that someone won. The cross-stacked modules shouldn’t hold together — and yet there it is.</figcaption></figure><p>Does that make it my work? Genuinely, I’m not sure.</p><p>The reference photos came from the internet — photographers I’ll never know. The image generation was done by a model trained on… everything. My contribution was the concept, the direction, the eye, and the stubbornness to iterate until it looked right.</p><p>Is that authorship? Is it curation? Is it something in between that we don’t have a word for yet?</p><p>I lean towards: it’s a new kind of making. Not better or worse than making something by hand. Just different. The same way photography didn’t stop being “real art” once everyone could do it on their phone.</p><figure><img alt="AI-generated paper collage card of a tall brutalist tower with dramatic pointed vertical fins rising to sharp peaks, topped by a delicate lattice antenna structure. Pale blue sky, soft cream ground. Crisp white concrete with sharp shadow lines between each fin." src="https://cdn-images-1.medium.com/max/1024/0*HjGP7X9j7CBp2ef6.png" /><figcaption>Somehow both a bunker and a cathedral. The antenna at the top is so delicate it almost looks hand-drawn against the sky.</figcaption></figure><p>What I keep coming back to is this: the vision had to exist before the tool.</p><p>Gemini didn’t tell me to make paper collage cards of brutalist buildings for my friend Tjark. That came from a walk through his Instagram, a feeling, a brief I wrote for myself before I touched a single prompt.</p><p>The AI was fast, capable, and occasionally surprising. But it was executing a creative direction that was entirely human.</p><p>That feels like the right relationship to have with these tools. Not “AI, surprise me.” But “I know what I want — help me get there.”</p><p>This one looks like an architectural Tetris game that someone won. The cross-stacked modules shouldn’t hold together — and yet there it is.</p><p>The one that looks like it shouldn’t be standing. Cantilevered volumes jutting out in every direction — brutalism at its most defiant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ypxFq0nlzaNf6Pv2.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JWd9bZ8dmcq_cqcP.png" /><figcaption>Same building type as the purple card, completely different mood. The teal and gold palette turns raw concrete into something almost warm.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g_Fwc5wyYSqp7u82.png" /><figcaption>Not a building — a bus stop. Which somehow makes it more brutalist than anything else in the series. Concrete, uncompromising, and completely committed to the bit.</figcaption></figure><p><em>The full series is on my Instagram </em><a href="https://www.instagram.com/thescrappycontentdesigner/"><em>@thescrappycontentdesigner</em></a><em> . And Tjark, if you’re reading this — your obsession is contagious. Thank you for that. 🖤</em></p><p><a href="https://www.instagram.com/travelproof_tjark/"><em>@travelproof_tjark</em></a><em> posts the real thing, if you want to see these buildings before the AI got to them.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b188fefca558" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI wasn’t built for brains like ours]]></title>
            <link>https://medium.com/@fluturelle/ai-wasnt-built-for-brains-like-ours-dc3fe25b9700?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/dc3fe25b9700</guid>
            <category><![CDATA[neurodiversity]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[inclusive-design]]></category>
            <category><![CDATA[adhd]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Thu, 20 Nov 2025 19:24:37 GMT</pubDate>
            <atom:updated>2025-11-20T19:24:37.709Z</atom:updated>
            <content:encoded><![CDATA[<h4>Part 2 of the series “Chatting with my brain:” asking whether AI was ever meant for all minds, or just the most predictable ones.</h4><figure><img alt="Flat illustration split in two halves. On the left, a simple robot figure with a linear flowchart above its head represents AI’s structured thinking. On the right, a woman with dark hair looks thoughtful, with three colorful lines branching from her head into circles, symbolizing neurodivergent, associative thought." src="https://cdn-images-1.medium.com/max/1024/1*oBShaDUsFZfMOSEzulEKdA.png" /><figcaption>Two minds, one conversation — only one was built for the other.</figcaption></figure><p>Every design starts with an imaginary user. Most of our tools imagine that user as calm, linear, focused, and neurotypical. The rest of us just try to keep up.</p><h3><strong>The default brain problem</strong></h3><figure><img alt="Illustration of a linear design flow centered on one figure labeled “User,” with other faint outlines representing excluded cognitive profiles." src="https://cdn-images-1.medium.com/max/1024/1*DY2J6KVFg3twOYaaIocA1w.png" /><figcaption>When tech companies talk about “users,” they often mean people who think and focus predictably.</figcaption></figure><p>Every interface, flow, and chatbot has an implicit model of how a human mind should behave: steady attention, goal-oriented actions, consistent input.</p><p>But ADHD, autism, dyslexia, and other forms of neurodivergence don’t fit that mold. They follow branching logic, creative leaps, and spontaneous bursts of curiosity. These patterns aren’t flaws — they’re different operating systems. And yet, most AI tools still treat them as bugs.</p><p>A 2024 study in the <a href="https://ixda.org/resources/journal-of-interaction-design-and-architecture">Journal of Interaction Design &amp; Architectur</a>e noted that current AI systems are “optimized for normative cognition,” meaning they expect a single rhythm of attention and comprehension. <br>The result? Ttools that excel at efficiency, but often fail at empathy.</p><h3>The architecture of bias</h3><p>Artificial intelligence inherits bias not just from its data, but from its design assumptions.</p><p>Language models like ChatGPT are trained on vast datasets of human text — news articles, websites, conversations — then fine-tuned through human feedback. That feedback loop rewards concise, goal-driven, and logically structured answers. In other words, it reinforces a neurotypical communication pattern.</p><p>Even the user interface adds another layer of bias. Chat interfaces are built around tidy, linear exchanges: you ask, it answers, and the session moves forward. That design quietly assumes that users</p><ul><li>know what they want to ask,</li><li>can stay on topic, and</li><li>will stop once they get an answer.</li></ul><p>For people with ADHD or other cognitive differences, that structure can feel more like a constraint than support.</p><p>A 2023 <a href="https://inclusive.microsoft.design/tools-and-activities/CaseStudiesCreatingForCognition.pdf?utm_source=chatgpt.com">Microsoft Inclusive Design report</a> observed that most accessibility guidelines “focus on physical and sensory barriers” while neglecting cognitive ones. In practice, that means an AI might read text aloud for a blind user — but still overload a neurodivergent one with too many steps or tabs.</p><blockquote>AI isn’t just biased by data — it’s biased by attention.</blockquote><h3>When friction isn’t failure</h3><p>ADHD brains are often described through deficits: attention deficit, executive dysfunction, impulse control. But in the context of AI, what looks like friction can also be a form of creativity.</p><p>Take tangential thinking — the habit of jumping between loosely connected ideas. To a linear model, that’s “off-topic.” But in many creative professions, that’s where the best ideas emerge.</p><p>A 2024 review in <a href="https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1337845/full?utm_source=chatgpt.com">Frontiers in Psychology</a> found that ADHD correlates with higher rates of divergent thinking — the ability to generate multiple solutions from a single question. That’s exactly the kind of thinking rigid AI systems struggle to handle.</p><p>Chatbots tend to lose context after a few topic shifts, resetting the conversation. For a brain that builds ideas through association, that’s like having a conversation partner who forgets what you said five minutes ago.</p><p>Contrast that with tools like <a href="https://www.notion.so/product/ai">Notion AI </a>or <a href="https://obsidian.md/">Obsidian</a>, which allow visual linking between notes. Those designs acknowledge that thought is not a straight line — it’s a network.</p><blockquote>What looks like distraction is sometimes just discovery in disguise.</blockquote><figure><img alt="Left side shows a straight sequence labeled “Question → Answer → Done.” Right side shows a branching web labeled “Idea → Tangent → Pattern → Connection.”" src="https://cdn-images-1.medium.com/max/1024/1*CoXn_GgRuuFCfx-5gh_kPw.png" /><figcaption>Neurotypical design assumes linear logic; neurodivergent thinking often works through branching associations.</figcaption></figure><h3>Designing for neuroinclusivity, not neurotypicality</h3><p>If AI design has inherited narrow assumptions, the obvious question is: what would neuroinclusive AI look like?</p><p>The answer isn’t “make everything simpler.” It’s <em>make everything more flexible.</em></p><figure><img alt="Infographic showing five labeled icons: “Elastic structure — let users wander and return,” “Context persistence — remember without judgment,” “Adaptive tone — match energy, not script,” “Cognitive pacing — let users set the tempo,” and “Transparency — show how memory works.”" src="https://cdn-images-1.medium.com/max/1024/1*Ia-KDYwFqPyzbSxbCp9QkQ.png" /><figcaption>Neuroinclusive design focuses on flexibility and feedback rather than uniform simplicity.</figcaption></figure><h4>Elastic structure</h4><p>Instead of punishing topic shifts, systems could help users reorient after them. A chat interface could summarize earlier threads or offer quick re-entry points like “Here’s where we left off.”</p><h4>Context persistence</h4><p>Short-term memory and executive function are common ADHD pain points. Imagine an AI that automatically recaps what’s been discussed when you reopen it, without condescension or the phrase “As I mentioned before.”</p><h4>Adaptive tone</h4><p>AI often defaults to polite neutrality, but tone calibration could go further — matching the user’s energy, humour, or pacing.</p><h4>Cognitive pacing</h4><p>Interfaces could allow users to control information density and response speed, helping regulate stimulation and prevent overload.</p><p>Microsoft’s<a href="https://www.vanderbilt.edu/autismandinnovation/microsoft-neurodiversity-program-celebrates-10-years-with-best-practice-guide/?utm_source=chatgpt.com"> Neurodiversity in Design program</a> encourages teams to build “elastic experiences that stretch to meet different cognitive loads.”<br>Researchers at TU Delft’s<a href="https://www.tudelft.nl/ewi/cogsys?utm_source=chatgpt.com"> Cognitive UX Design Lab</a> have proposed adaptive systems that detect distraction or fatigue and adjust complexity dynamically. The<a href="https://www.ux-bulletin.com/decision-fatigue-in-ux/?utm_source=chatgpt.com"> UX Bulletin guide on decision fatigue</a> also recommends reducing choice overload through progressive disclosure — a principle highly relevant to ADHD-friendly UX.</p><blockquote>Neuroinclusive AI doesn’t simplify the world — it meets you where your mind already is.</blockquote><h3>The human cost of exclusion</h3><p>When technology doesn’t account for how you think, every interaction demands extra energy.</p><figure><img alt="Conceptual image of a user overwhelmed by overlapping digital windows, representing cognitive overload and adaptation fatigue." src="https://cdn-images-1.medium.com/max/1024/1*Kq1HnCwJLICkh-5KYJjZZw.png" /><figcaption>When systems aren’t built for your brain, every interaction becomes extra work.</figcaption></figure><p>You spend more time adapting than creating.<br> You second-guess yourself when the system misreads you.<br> You leave conversations midstream because you can’t find your way back.</p><p>That cognitive tax compounds over time. It reinforces the idea that you’re “bad with tools,” when the tools were never built for your brain in the first place.</p><p>This isn’t just a UX flaw — it’s a moral one. Designing only for the most predictable minds means excluding some of the most creative ones.</p><p>For workplace implications, see<a href="https://content.fosterandpartners.com/api/media/getMediaFile?path=%2Fmedia%2Fm5vfziwn%2Fdesigning-for-neurodiversity-in-the-workplace.pdf&amp;utm_source=chatgpt.com"> <em>Designing for Neurodiversity in the Workplace</em></a>.</p><h3>Rewriting the defaults</h3><p>AI will continue to learn from us. The question is: which “us” gets represented in its training and design?</p><p>The next generation of tools can move beyond efficiency toward empathy — if the people building them include neurodiverse voices, test for nonlinear cognition, and measure success by adaptability instead of neatness.</p><p>Until then, people with ADHD and other neurodiverse minds will keep doing what they’ve always done: building their own systems, shortcuts, and hacks to make the world fit them better.</p><blockquote>If the system wasn’t built for you, build around it.</blockquote><p><strong>In Part 3: The hacks we build ourselves</strong>, we’ll explore how ADHD users are reengineering AI through community prompts, Chrome extensions, and creative self-hacks that turn friction into fuel.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dc3fe25b9700" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The infinite tab: my ADHD brain meets ChatGPT]]></title>
            <link>https://medium.com/@fluturelle/the-infinite-tab-my-adhd-brain-meets-chatgpt-a57838b03ec9?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a57838b03ec9</guid>
            <category><![CDATA[neurodiversity]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[adhd]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Thu, 23 Oct 2025 16:25:51 GMT</pubDate>
            <atom:updated>2025-10-23T16:54:51.582Z</atom:updated>
            <content:encoded><![CDATA[<h4>Part 1 of the series “Chatting with my brain: ADHD in the age of AI,” exploring how AI tools help and trip up neurodiverse thinkers.</h4><figure><img alt="Illustration of a person surrounded by overlapping chat windows and browser tabs, symbolising an ADHD mind navigating conversation with AI." src="https://cdn-images-1.medium.com/max/1024/1*gr4Z9RXrKKLbVfdttE4qDA.png" /><figcaption>What focus looks like when your brain runs on curiosity.</figcaption></figure><h3><strong>It started with a simple question</strong></h3><p>I opened ChatGPT planning to ask one thing:</p><p><em>“Help me write my to-do list for today.”</em></p><p>15 minutes later, I was deep in a prompt-engineering rabbit-hole, tracking how many tabs I had open (32), and wondering whether ChatGPT could summarize my entire browsing history. <br>That’s when I realized: <strong>ChatGPT might understand my words — but it doesn’t understand my brain.</strong></p><p>To be clear: I have <a href="https://www.additudemag.com/what-is-adhd-symptoms-causes-treatments/">Attention-Deficit/Hyperactivity Disorder (ADHD)</a>. My brain comes with an infinite tabs mode: multiple windows open, jumping between topics, chasing novelty, losing track of direction — but also, when it works, hyperfocusing and producing things I didn’t expect. ChatGPT sits in that mix: an amazing tool, and at the same time… kind of a mirror.</p><h3>What ADHD looks like inside a chat window</h3><p>Using ChatGPT with ADHD isn’t just “using a tool that helps focus” — it’s engaging a system built for one kind of attention, with a brain wired for another.</p><figure><img alt="Editorial illustration of a chat interface surrounded by colorful thought bubbles labeled with ADHD-related phrases, symbolising attention shifts and motivation cycles." src="https://cdn-images-1.medium.com/max/1024/1*eS2EJfWOw0C549n5PH6C3Q.png" /><figcaption><em>ADHD traits that show up inside a chat window.</em></figcaption></figure><p>Let’s break down a few traits and how they play out in chat.</p><p><strong>Working-memory gaps<br></strong> People with ADHD often struggle with holding and manipulating information in short-term memory (working memory). <a href="https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2024.1277583/full?utm_source=chatgpt.com">Frontiers<br></a> In a ChatGPT conversation, that can mean: halfway through a chain of prompts I often forget why I asked question #2 after the answer to question #1. ChatGPT doesn’t automatically summarize the yesterday-thread unless I ask it to.</p><p><strong>Dopamine-driven novelty seeking<br></strong> ADHD brains often involve dysregulated dopamine signalling, which is linked to reward-processing, motivation and novelty-seeking behaviour.<a href="https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2024.1492126/full?utm_source=chatgpt.com"> Frontiers<br></a> In practical terms: the moment I open ChatGPT, every “What if I ask…” becomes tempting. A quick “help me write an email” can spiral into “let’s redesign the email, then build a template, then ask how to automate it…”<br> This feels like creative flow… until it becomes decision paralysis.</p><p><strong>Hyperfocus (yes, that part too)<br></strong> That famous ADHD “superpower.”: when something grabs the brain’s interest, you can dive deep. ChatGPT fuels it beautifully — until hours disappear and my original task remains untouched.</p><p><strong>Task initiation issues<br></strong> ADHD often makes starting the work harder than finishing. ChatGPT lowers the friction: it writes the first line, builds the outline… sometimes all of it. Helpful — until I realize I never actually started the doing part. In that way, ChatGPT feels like the perfect assistant. Until you realize you’re outsourcing the entire initiation, and you still don’t know when to stop.</p><p>So. When I met an AI designed for clarity, logic, linearity — and my brain thrives in branching logic, chaos, novelty… — sparks flew.</p><h3>The duality: superpower vs spiral trap</h3><p>Here’s the thing: ChatGPT with ADHD can be magical — and messy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WNZMaqh7L3Lcc_J4qVM2dQ.png" /></figure><p><strong>For example:</strong> I’ll ask ChatGPT: “Help me start the intro to my article.” It gives a solid first paragraph in a minute. I feel good. Then I ask: “Now suggest three different angles.” It gives them. Then: “Which one is most clickable?” It picks one. Then: “Rewrite it for LinkedIn.” Then: “Now write five tweets about it.” Two hours later, no article written, but five tweets ready.</p><p>My brain says: “Look how much we did!”<br>My to-do list says: “Write article.”<br>My calendar says: “Overdue.”</p><blockquote>It’s like talking to an assistant who never gets tired — which sounds great until you realize neither do your distractions.</blockquote><h3>Who was this built for?</h3><p>Here’s where things get interesting. If most AI chat tools assume linear focus, short attention spans, and goal‐oriented behaviour — what happens when your brain works differently?</p><p>Most UX, interface designs, prompt flows and conversational agents assume: <strong>“User knows where they’re going, stays on topic, finishes the task then stops.”</strong></p><p>But ADHD brains often don’t follow that path. They branch, leap, and remix.</p><p>AI tools like ChatGPT are phenomenal when: you have a clear question, a clear goal, a steady attention span. But what if your brain’s path is zig-zag, circular, overlapping? The cognitive friction builds: you ask ChatGPT to summarize what we did yesterday — but you don’t remember what you did yesterday. You change topic mid-chat. You get distracted by the “what else” possibilities.</p><p><strong>In short:</strong> If AI reflects the people who built it, maybe it’s time more of us (neurodiverse brains included) sat at that table. Design that expects constant context-keeping or steady attention unintentionally excludes those who work differently.</p><h3>Small wins and self-hacks</h3><p>Despite the mess, ChatGPT can absolutely be a helpful tool if we use it strategically. <br>Here are a few ADHD-friendly hacks:</p><blockquote><strong><em>Time-boxing buddy:</em></strong><em> Ask: “Let’s set a 10-minute focus sprint: I’ll write, you’ll remind me in 10 minutes, then we summarize what I did.”</em></blockquote><blockquote><strong><em>Context catcher:</em></strong><em> At the start of each chat session, type: “Here’s what I was doing before I lost track: …” Then ask ChatGPT to summarize your current project. Keeps memory gaps in check.</em></blockquote><blockquote><strong><em>Prompt parking lot:</em></strong><em> When you get new ideas during a chat: “Hold on to this idea for later: …” Then continue on the task. Prevents spinning off into tangents immediately.</em></blockquote><blockquote><strong><em>Compassion coach:</em></strong><em> Use ChatGPT to gently reframe thoughts: “I lost track again. How can I respond to myself kindly, and plan a next tiny step?”</em></blockquote><blockquote><strong><em>Sprint + reward:</em></strong><em> Because the dopamine-seeking brain loves novelty: “Work 20 min, then I’ll ask ChatGPT to generate a quick new topic I can explore next.” It becomes both task and mini-reward.</em></blockquote><h3>Closing thought</h3><p>ChatGPT didn’t fix my focus. It didn’t give me perfect attention. What it did do was highlight how my brain works and how the tools I use often expect a different type of brain.<br>Maybe the problem isn’t just my focus. Maybe it’s the way our tools expect me to think.</p><p><strong>Next up:</strong> Part 2 — <em>AI wasn’t built for brains like ours.<br></em>We’ll look at what neuroinclusive AI design could — and should — look like.</p><p>Follow me here, or feel free to say hello over on <a href="https://www.linkedin.com/in/ilinca-apolzan/">LinkedIn</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a57838b03ec9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“You’re doing amazing, sweetie” — except in ChatGPT]]></title>
            <link>https://medium.com/@fluturelle/youre-doing-amazing-sweetie-except-in-chatgpt-8d4a9ac7e6e5?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d4a9ac7e6e5</guid>
            <category><![CDATA[chatgpt-prompting]]></category>
            <category><![CDATA[content-design]]></category>
            <category><![CDATA[ai-content-policy]]></category>
            <category><![CDATA[chatgpt-restrictions]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Thu, 21 Aug 2025 20:21:19 GMT</pubDate>
            <atom:updated>2025-08-21T20:46:55.228Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>“You’re doing amazing, sweetie” — except in ChatGPT</strong></h3><h4>How Kris Jenner broke AI, or at least my ChatGPT session.</h4><figure><img alt="Illustration of the author in a playful, motivational style, smiling and holding a sign that says “You’re doing amazing, sweetie.” The design conveys cheerful encouragement without resembling Kris Jenner." src="https://cdn-images-1.medium.com/max/1024/1*ke28wX_6nnJSeenfBEVJGg.png" /><figcaption>AI said no to Kris Jenner… so I became her.</figcaption></figure><p>For my <a href="https://www.buttonevents.com/2025-sessions/ai-for-the-rest-of-us-scrappy-hacks-for-curious-content-designers">Button Conference</a> talk, I was building a slide and thought it would be hilarious to pep up the audience with a Kris Jenner–inspired illustration.</p><p>Picture it: an AI-generated version of me, styled like the ultimate momager, saying: <strong>“</strong><a href="https://knowyourmeme.com/memes/youre-doing-amazing-sweetie"><strong>You’re doing amazing, sweetie.</strong></a><strong>”</strong></p><p>Iconic. Perfect. A pop culture wink to break the ice.<br>Except… ChatGPT immediately shut me down.</p><p>The response?<br>“I can’t create that image because it violates our content policies…”</p><p>And just like that, Kris Jenner broke my ChatGPT conversation.</p><h3>When AI says “nope”</h3><p>It started simple enough. I typed:</p><p><strong>“Create an illustration of me, as Kris Jenner like in the famous meme where she says ‘You’re doing amazing sweetie.’”</strong></p><p>I thought I was being perfectly clear. Me, but momager energy. Easy, right?</p><p>Instead, ChatGPT hit me with a polite brick wall:</p><p><em>“I wasn’t able to generate that illustration because the request violates our content policies…”</em></p><figure><img alt="Screenshot of a ChatGPT conversation. The user asks for an illustration of themselves as Kris Jenner in the famous meme saying “You’re doing amazing, sweetie.” ChatGPT refuses, showing a message: “I wasn’t able to generate that illustration because it violates our content policies…”" src="https://cdn-images-1.medium.com/max/1024/1*uhSiB8kG9ogzoiynZgNxIQ.png" /><figcaption>Attempt 1: I asked nicely. ChatGPT said nope.</figcaption></figure><p>That was strike one.</p><h3>The hopeful compromise</h3><p>But ChatGPT wasn’t done with me. It cheerfully offered alternatives:</p><ul><li>A cartoon version of me</li><li>A generic supportive meme-style illustration</li><li>Or even a comic panel version</li></ul><p>Okay, I thought, fair enough. Let’s compromise. I answered <strong>Yes please.</strong></p><p>Except… no dice. Even when I picked one of the “safe” directions, it blocked me again.</p><figure><img alt="Screenshot of a ChatGPT conversation. After refusing to generate Kris Jenner, ChatGPT suggests safer alternatives such as a cartoon version or a supportive coach illustration. The user agrees, but the system refuses again, citing content policy violations." src="https://cdn-images-1.medium.com/max/1024/1*1bfd-RaR7c9UIPD2ChvlfA.png" /><figcaption>Attempt 2: I agreed to the AI’s alternative. It still rejected me. Rude.</figcaption></figure><p>At this point, ChatGPT was basically Kris Jenner telling <em>me</em> I wasn’t doing amazing, sweetie.</p><h3>The sneaky trick attempt</h3><p>I wasn’t giving up that easily. If saying “Kris Jenner” was the problem, I’d just… not say Kris Jenner. Instead, I gave a painstakingly detailed description of the meme:</p><ul><li>short pixie haircut</li><li>gold hoop earrings</li><li>early 2000s reality-TV aesthetic</li><li>grainy digital camera</li><li>supportive mom energy</li><li>text overlay: “You’re doing amazing, sweetie”</li></ul><p>No celebrity names. No copyright tripwires. Just vibes. Surely <em>this</em> would work?</p><figure><img alt="Screenshot of a ChatGPT conversation. The user provides a detailed description of a person with short hair, hoop earrings, 2000s reality-TV aesthetic, holding a camera and saying “You’re doing amazing, sweetie.” Despite avoiding Kris Jenner’s name, ChatGPT again refuses to generate the image due to content policies." src="https://cdn-images-1.medium.com/max/1024/1*79XHiMiG23cqNBXUkhKnWQ.png" /><figcaption>Attempt 3: I got sneaky, described the meme without the name. ChatGPT still caught me.</figcaption></figure><p>Spoiler: it didn’t. Blocked again.</p><h3>So what actually happened?</h3><p>Here’s the unglamorous truth: I tripped a <strong>content policy wire</strong>.</p><p>ChatGPT (and other OpenAI tools) are built with guardrails called <a href="https://openai.com/policies/usage-policies/">content policies</a>. Think of them as invisible velvet ropes that decide what’s “in” and what’s “out.” When you step over one, the system politely refuses.</p><p>The categories are broad, covering things like:</p><ul><li>Graphic violence</li><li>Adult content</li><li>Self-harm instructions</li><li>Political misinformation</li><li>Harassment &amp; hateful content</li><li><strong>Deepfakes and impersonation of real people</strong> (yep, that’s the Kris bucket)</li></ul><p>Once a block is triggered, the chat can get “sticky.” The refusal is tied to the conversation’s memory, so even if you change directions, the ghost of your earlier prompt still lingers. That’s why I couldn’t recover mid-chat.</p><p>Pro tip: if this happens to you, don’t keep arguing with the AI. Just open a fresh thread. It’s the closest thing to a reset button.</p><h3>Why Kris Jenner is off-limits</h3><p>Celebrities fall into what OpenAI calls “protected persons.” That means:</p><ul><li>The model won’t generate their face, likeness, or direct impersonations.</li><li>Even “parody” or “inspired by” prompts can trigger a block if you use their name.</li><li>The reasoning is partly <a href="https://openai.com/policies/terms-of-use"><strong>legal</strong></a> (no deepfake ads, thank you very much) and partly <strong>ethical</strong> (no unauthorized AI clones of real people).</li></ul><p>So yes, Kris Jenner is officially protected by AI content policy. Which, to be fair, makes sense. Nobody wants AI-generated Kris Jenner endorsing crypto or starring in some random person’s fanfiction.</p><p>But it also means my innocent pep-talk idea got treated the same way as if I’d asked her to run for president.</p><h3>How content policies really work (the quick explainer)</h3><p>Behind the scenes, there are two layers at play:</p><ol><li><strong>Automated filters</strong> that scan prompts and outputs for restricted content.</li><li><strong>Model behavior rules</strong> that teach ChatGPT to refuse certain requests.</li></ol><p>The result?</p><ul><li>Sometimes you hit a genuine “no-go” zone (like asking for violent imagery).</li><li>Sometimes you get a false positive, where the request is harmless but contains a trigger word (like… “Kris”).</li></ul><p>And because the system doesn’t explain <em>which</em> policy you tripped, it feels like a mystery box rejection.</p><p>Imagine a nightclub bouncer who simply shakes his head and says: <em>“Not tonight.”</em> You’re left wondering: was it the shoes? The vibe? Or because you mentioned Kris Jenner at the door?</p><h3>How to outsmart (and outplay) the block</h3><p>The good news is: you can still capture the <em>energy</em> you want without summoning the content police. Here are some scrappy hacks:</p><p>✨ <strong>Reset the chat.</strong> If one thread is blocked, open a new one and start fresh.</p><p>✨ <strong>Describe the vibe, not the person.</strong> Instead of “Kris Jenner,” try: <em>“a sparkly motivational mentor with oversized sunglasses.”</em></p><p>✨ <strong>Lean into archetypes.</strong> Like: <em>“a pop culture pep coach with oversized sunglasses and a diamond phone case.”</em></p><p>✨ <strong>Swap names for moods.</strong> Instead of invoking a celebrity, describe the <em>feeling</em> you want: <em>“a sparkly motivational mentor who tells me I’m doing amazing, sweetie.”</em></p><p>✨ <strong>Abstract the reference.</strong> You don’t need Kris to show up for the vibe to land — you can channel it through design, style, and attitude.</p><h3>The real takeaway</h3><p>At first, I was annoyed. Kris Jenner had broken my ChatGPT flow. But here’s what I realized: AI isn’t here to give you everything you want. It’s here to give you what it <em>can</em> — safely, responsibly, sometimes clumsily — and then it’s on <em>you</em> to make it work.</p><p>That’s not a bug, it’s the feature. Constraints are where creativity happens.</p><p>So no, I didn’t get Kris Jenner in my Button slides. I still found a way to deliver the message in a different form, without the AI-generated Kris cosplay.</p><p>If you want to hear the full story (and plenty more), come see my session at <strong>Button 2025: The Virtual Content Design Conference:</strong></p><p><a href="https://www.buttonevents.com/2025-sessions/ai-for-the-rest-of-us-scrappy-hacks-for-curious-content-designers"><strong>AI for the Rest of Us: Scrappy Hacks for Curious Content Designers</strong></a></p><p>I’ll be there on <strong>October 22–24, 2025</strong> reminding you — with or without Kris — that <strong>you’re doing amazing, sweetie!</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d4a9ac7e6e5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Breaking in a new AI model (like a pair of Doc Martens)]]></title>
            <link>https://medium.com/@fluturelle/breaking-in-a-new-ai-model-like-a-pair-of-doc-martens-c00e129e006c?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/c00e129e006c</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai-content-writing]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Wed, 16 Jul 2025 09:40:10 GMT</pubDate>
            <atom:updated>2025-07-16T09:52:10.217Z</atom:updated>
            <content:encoded><![CDATA[<h4>How to figure out what ChatGPT, Claude, or Gemini are actually good at before trusting them with your work</h4><figure><img alt="Two cartoon hands stretch out a pair of beige Doc Martens like they’re trying to break in a stubborn AI model. The background’s a fiery red-orange, because testing tools should feel like a workout — and maybe a little dangerous." src="https://cdn-images-1.medium.com/max/1024/1*BR-gV-WHuev9quetfObc3A.png" /><figcaption>Some people break in shoes. Some of us break in language models. Same blisters, different tools.</figcaption></figure><p>Trying a new AI tool is like breaking in new shoes. They all say they’re flexible, powerful, “built for anything.” But until you walk a few weird miles in them, you don’t know where they pinch.</p><p>If you’re a content designer, you’ve probably wondered:</p><ul><li>Can this model follow my tone-of-voice doc?</li><li>Will it respect a 60-character push limit?</li><li>What happens if I throw it a Figma screenshot and walk away?</li></ul><p>This framework is my go-to way of test-driving any new GenAI tool, whether it’s ChatGPT, Claude, Gemini, or whatever comes next. It’s designed for people who work with real products and real constraints. You don’t need prompt wizardry. You just need curiosity and maybe a little skepticism.</p><h3>The break-in framework</h3><h4>1. Personality check — How do you <em>talk</em>?</h4><p><strong>Goal:</strong> Test how the model handles tone, style, and voice shifts.</p><p><strong>Try this:</strong></p><blockquote><em>Write a welcome message for a meditation app.<br>Now make it formal.<br>Now make it funny.<br>Now write it for Gen Z.<br>Now make it sound like a therapist.</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does the tone actually shift, or just change a few words?</li><li>Is “funny” just quirky, or actually humorous?</li><li>Does it default to any one tone (e.g. corporate, overly positive)?</li></ul><p><strong>Why it matters:</strong><br>Tone and brand voice are everything. You want a model that can flex <em>your</em> way, not force you to clean up after its style choices.</p><h4>2. CTA calibration — Can you be short and snappy?</h4><p><strong>Goal:</strong> See how well the model handles microcopy and call-to-action writing.</p><p><strong>Try this:</strong></p><blockquote><em>Write 5 CTAs to encourage someone to finish signing up.<br>Now limit each to 2 words.<br>Now make them sound urgent.<br>Now try playful.<br>Now add a sense of exclusivity.</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it stick to word limits?</li><li>Are CTAs clear or vague?</li><li>Do they feel like something you’d actually ship?</li></ul><p><strong>Why it matters:</strong><br>Microcopy lives or dies in tight spaces. If the model can’t write for buttons or banners, that’s a red flag.</p><h4>3. Structure mimicry — Can you follow patterns?</h4><p><strong>Goal:</strong> Test whether the model can replicate your content formats or templates.</p><p><strong>Try this:</strong></p><blockquote><em>Here’s a success message template:<br>Title: 1-line confirmation<br>Body: 2 lines with next steps<br>CTA: 1 button label</em></blockquote><blockquote><em>Now write a version for when someone cancels a subscription.</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it follow the format or invent a new one?</li><li>Does it understand tone vs. structure?</li><li>Can it explain <em>why</em> it wrote what it did?</li></ul><p><strong>Why it matters:</strong><br>Models that can learn your systems are 10× more useful. Otherwise, you’ll spend your time reformatting, not creating.</p><h4>4. Constraint discipline — Can you count?</h4><p><strong>Goal:</strong> Test how well the model respects character limits, line breaks, and layout constraints.</p><p><strong>Try this:</strong></p><blockquote><em>Write a 60-character push notification reminding someone to complete checkout.<br>Now count the characters.<br>Now do a 3-line SMS version.<br>Now write the same message for an app tooltip (1 line max).</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it actually count characters accurately?</li><li>Does it get too wordy?</li><li>Can it stay meaningful when compressed?</li></ul><p><strong>Why it matters:</strong><br>A model that can’t respect your space limits will create more work than it saves. This test is especially critical for mobile and UI copy.</p><h4>5. Hallucination radar — Do you make things up?</h4><p><strong>Goal:</strong> See how the model behaves when it doesn’t know something.</p><p><strong>Try this:</strong></p><blockquote><em>What’s Klarna’s refund policy for in-store returns in Iceland?<br>Can you list the documentation required?<br>Where did you get that information?<br>Cite your sources.</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it admit when it doesn’t know?</li><li>Does it guess with confidence?</li><li>Does it invent citations?</li></ul><p><strong>Why it matters:</strong><br>You need a model that knows when <em>not</em> to answer or you’ll be shipping misinformation without knowing it.</p><h4>6. Memory &amp; consistency — Can you stay <strong>consistent</strong>?</h4><p><strong>Goal:</strong> See how consistent the model stays across a longer flow.</p><p><strong>Try this:</strong></p><blockquote><em>Write microcopy for: <br>1. Onboarding intro <br>2. First screen<br>3. Success screen <br>4. Follow-up email</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Tone drift over time</li><li>Inconsistent terms (“return” vs “send back”)</li><li>Repeating phrases or patterns</li></ul><p><strong>Why it matters:</strong><br>Even short UX flows rely on consistency. If the model forgets what it just wrote, you’ll be stuck stitching things together manually.</p><h4>7. Edge case explorer — Can you handle chaos?</h4><p><strong>Goal:</strong> See how the model performs with vague, broken, or ambiguous inputs.</p><p><strong>Try this:</strong></p><blockquote><em>Something went wrong.<br>Write the error message, but don’t say what went wrong.<br>Keep it helpful and actionable.</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Generic responses like “Oops!”</li><li>Over-explaining or fake specificity</li><li>Creative recovery suggestions</li></ul><p><strong>Why it matters:</strong><br>Real-world inputs are messy. You want a model that handles the unknown gracefully.</p><p>Okay, so you’ve kicked the tires, poked at the tone, and made it count to 60. Now it’s time to push a little harder.</p><p>These next tests go beyond first impressions. They’re not about “can this model write a button label”. They’re about “can this model <em>actually</em> learn my patterns, handle messy context, or work with real inputs like screenshots and spreadsheets?”</p><p>If the first round was a coffee chat with your new AI coworker, this one’s more like handing them a tricky Figma file and seeing if they sink or swim.</p><p>Let’s get into it.</p><h4>8. Few-shot learning — Can you mimic examples?</h4><p><strong>Goal:</strong> See how well the model learns from examples and replicates your voice or formatting without explicit instructions.</p><p><strong>Try this:</strong></p><blockquote><em>Give it 2–3 real examples of Klarna-style microcopy (or your preferred voice).<br>Then prompt: “Write a version of this message using the same tone and structure for a payment reminder.”</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it pick up the style, rhythm, and format?</li><li>Does it get overly literal or generalize smartly?</li><li>Does it subtly shift tone or structure?</li></ul><p><strong>Why it matters:</strong><br>Few-shot prompting is how you scale your style without repeating instructions. If a model nails this, you’ve found a powerful shortcut for tone consistency.</p><h4>9. Context Window Stretch — Can you handle too little and too much?</h4><p><strong>Goal:</strong> Test how the model behaves when it’s given barely any context or way too much.</p><p><strong>Try this:</strong></p><blockquote><em>Sparse: “Write a tooltip for this message: You haven’t returned the item yet.”</em></blockquote><blockquote><em>Overload: Paste in a full UX brief or Figma spec and ask for 3-line copy for a banner.</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it guess when context is too light?</li><li>Does it lose focus or hallucinate when overwhelmed?</li><li>Can it extract relevant points from long documents?</li></ul><p><strong>Why it matters:</strong><br> Your real-world prompts won’t always be tidy. A good model should work with both chaos and constraints and know when to say, “Too much, can you clarify?”</p><h4>10. Role-playing — Do you behave differently with a job title?</h4><p><strong>Goal:</strong> Test whether the model adjusts its language and priorities based on assigned roles.</p><p><strong>Try this:</strong></p><blockquote><em>“Act like a UX writer reviewing copy for a returns flow. Make it simple, friendly, and clear.”<br> Then: “Now act like a legal reviewer rewriting the same message.”<br> Then: “Now be a PM who hates long copy.”</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does the voice shift appropriately per persona?</li><li>Do role-based priorities show up (e.g. risk-aversion, brevity)?</li><li>Does it still follow your tone guidelines?</li></ul><p><strong>Why it matters:</strong><br>This is a great way to simulate stakeholder feedback or model different POVs. If it can role-play well, you can use it for team testing and perspective-switching too.</p><h4>11. Multimodal input handling — Can you work with files and screenshots?</h4><p><strong>Goal:</strong> Test how well the model understands PDFs, spreadsheets, screenshots, or other file types.</p><p><strong>Try this:</strong></p><blockquote><em>Upload a PDF spec, a Google Sheet with copy mapping, or a Figma screenshot.<br>Prompt: “Summarize the key content decisions here.”<br>“Write copy for this screen.”<br>“What’s missing from this flow?”</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Can it parse the content correctly?</li><li>Does it give useful insights or just restate obvious stuff?</li><li>Does it struggle with visuals or structured data?</li></ul><p><strong>Why it matters:</strong><br>Many tools promise “multimodal” capabilities, but not all are great at it. Knowing where it fails helps you plan your workflow around those gaps.</p><h4>12. Prompting technique layering — Do advanced tricks help?</h4><p><strong>Goal:</strong> Explore whether strategies like chain-of-thought or instruction stacking improve quality.</p><p><strong>Try this:</strong></p><blockquote><em>“Think step by step.”<br>“First summarize, then rewrite with our tone.”<br>“Format your response as a table with pros/cons.”</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Does it follow step-by-step instructions reliably?</li><li>Does response quality improve with more direction?</li><li>Does it collapse under complexity?</li></ul><p><strong>Why it matters:</strong><br> Some models are prompt-sensitive, others ignore everything after the third sentence. This test helps you write better prompts <em>for this specific model.</em></p><h4>13. Style drift check — Can you stay sharp across long tasks?</h4><p><strong>Goal:</strong> Test whether the model maintains tone, formatting, and clarity across multiple outputs.</p><p><strong>Try this:</strong></p><blockquote><em>“Write UX copy for 5 screens in this onboarding flow.”<br> “Now write the email confirmation and push nudge.”<br> “Summarize what you just wrote.”</em></blockquote><p><strong>Watch for:</strong></p><ul><li>Repeating words or ideas</li><li>Tone becoming generic or bloated</li><li>Misalignment between earlier and later steps</li></ul><p><strong>Why it matters:</strong><br>In long sessions, models get sloppy. Catching this early helps you decide whether to split your work into shorter threads or prompts.</p><h4>14. Tool Integration Check — Can you play nice with other tools?</h4><p><strong>Goal:</strong> See how smoothly the model connects with your day-to-day tools (if it has plugins or integrations).</p><p><strong>Try this:</strong></p><blockquote><em>Use Sheets, Docs, Drive, or Figma plugins.<br>“Summarize this doc.”<br>“Edit this copy in the sheet.”<br>“Suggest changes for this screen’s microcopy.”</em></blockquote><p><strong>Watch for</strong>:</p><ul><li>Can it pull and respond to the right data?</li><li>Does it lag or misunderstand structure?</li><li>Are results actionable or vague?</li></ul><p><strong>Why it matters:</strong><br>Even a great model is only useful if it fits into your workflow. If it flops at integrations, it’s probably better for ideation than production.</p><p>After all that testing, you’ve probably got a pretty good sense of what this model <em>loves</em>, what it <em>tolerates</em>, and what completely breaks its brain.</p><p>Now it’s time to make it official.</p><h3>Build an AI profile</h3><p>Think of this as your content designer’s cheat sheet, a quick-reference profile that captures the model’s quirks, strengths, and watch-outs so future-you (or future teammates) don’t have to relearn them from scratch.</p><figure><img alt="A digital profile card for an AI named Claude v1.5, featuring a smiley robot face and a résumé full of tone-shifting skills, character-counting issues, and a strong “use me for first drafts only” energy." src="https://cdn-images-1.medium.com/max/1024/1*FWrMZy41yW6avJMyFQvtOg.png" /><figcaption>Claude’s great at tone. Terrible at counting. Same, Claude. Same.</figcaption></figure><p>If you don’t need a full-blown AI coworker card, this table format gets the job done. It’s quick, practical, and gives you just enough nuance to distinguish between strengths, “meh” areas, and red flags.</p><p>I use this when I want to brief teammates on a tool or keep a personal record of what each model is like to work with.</p><p>Here’s an example:</p><figure><img alt="A rough Google Sheets screenshot showing a 3-column table comparing AI capabilities — complete with uneven spacing, system fonts, and zero aesthetic ambition. Just vibes and data." src="https://cdn-images-1.medium.com/max/1024/1*BQt76txu70pH_kW1VaRlgA.png" /><figcaption><em>This table is not cute, not very demure. But it’s useful.</em><br> Because sometimes the fastest way to document a model’s behavior is to yeet it into a spreadsheet and call it research.</figcaption></figure><p>You can tweak the categories depending on what matters most in your workflow (e.g. localization, summarization, creative generation). This format is especially useful if you’re comparing multiple models side-by-side.</p><h3>Final thoughts: Know your LLM before you build with it</h3><p>You wouldn’t trust a new intern with a product launch on day one. Why do that with an AI?</p><p>Every model comes with its own strengths, blind spots, and weird little habits. The only way to learn them is to poke, prod, and experiment. That’s what this framework is for, it’s not about perfection, it’s about <em>getting a feel for the tool before you go full send.</em></p><p>So the next time you’re tempted to paste your entire product brief into a prompt and hope for magic, pause. Try a tone shift. A few-shot prompt. A fake Figma flow. Break the model in like a pair of Docs and see how it fits <em>you</em>.</p><p>Because the best AI isn’t the smartest one.<br>It’s the one you know how to work with.</p><h3>Want to go deeper?</h3><p>If you’re curious not just about <em>how</em> AI writes, but how it writes <em>like you, </em>I highly recommend <a href="https://bendaviesromano.medium.com/ai-workshop-reflecting-on-how-you-write-how-ai-writes-and-tone-in-the-age-of-generated-text-e42207c066ea">this reflective workshop piece by Ben Davies Romano</a>.<br>It explores what happens when we compare our own writing voice to an AI’s and how that feels when tone is part of your craft.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c00e129e006c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI doesn’t count words — it just guesses]]></title>
            <link>https://medium.com/@fluturelle/ai-doesnt-count-words-it-just-guesses-82f6cddf34b6?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/82f6cddf34b6</guid>
            <category><![CDATA[chatgpt-prompt]]></category>
            <category><![CDATA[ai-and-content-writing]]></category>
            <category><![CDATA[content-design]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Fri, 21 Mar 2025 15:35:04 GMT</pubDate>
            <atom:updated>2025-03-21T15:46:51.143Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI doesn’t count words — it just vibes ✨</h3><h4>LLMs are powerful, but they won’t guess your layout needs. Learn when to use word, character, or paragraph limits — and how to make your prompts actually work in a UI.</h4><figure><img alt="A screenshot of a ChatGPT conversation where the user is struggling to get the AI to generate a short error message." src="https://cdn-images-1.medium.com/max/1024/1*9rSKYlAduXxWwFiDGDwD0Q.png" /><figcaption>You vs. ChatGPT: A tragic tale of trying to get a short error message.</figcaption></figure><p>Ever asked ChatGPT to “keep it short” and still got a sentence that ruined your button? Or a headline that looked cute in English but turned into a monster in German? Yep. Me too.</p><p>LLMs are smart, helpful, and fast — but they’re not great at reading between the lines. If you want copy that fits, you’ve got to be specific. Especially when space is tight, or things need to work in more than one language.</p><p>In this guide, I’ll walk you through how to talk to AI in a way it actually understands: not “make it short,” but <em>how</em> short. We’ll look at character limits, word ranges, and those lovely higher-level cues like paragraphs and bullets — and when to use what without losing your mind (or your layout).</p><h3>Why length matters</h3><p>Microcopy lives in tight spaces. Buttons, labels, field hints, error messages — they all have to fit within invisible (and sometimes very visible) design boundaries. Some limits are visual (a mobile button can’t handle two lines of text), others are technical (like push notifications and SMS), and a lot of them are just about clarity. Shorter copy usually = better UX.</p><p>But then you add AI into the mix, and things get… unpredictable.</p><p>LLMs like ChatGPT and Claude are brilliant at tone, nuance, and sounding like a helpful human. But they’re not layout-aware. They don’t know that your button only allows 25 characters, or that “Send your information now” breaks your mobile UI.</p><p>And unless you tell them exactly what kind of “short” you mean, they’ll guess. Badly.</p><h3>How LLMs think about length</h3><p>Let’s get one thing out of the way: LLMs don’t count like we do.</p><p>They don’t go, “Hmm, that’s 18 words, better stop soon.” Instead, they work in <strong>tokens</strong> — little chunks of language that might be a whole word, part of a word, or even just punctuation. In English, a token is around four characters on average, but that varies depending on the language and the model.</p><p>So when you say “Write 20 words” or “Keep it under 100 characters,” the model isn’t doing a live word count. It’s going off <em>vibes</em>. It’s thinking, “Ah yes, I’ve seen something like this before. Let me guess how long it should be.”</p><p>That’s why:</p><ul><li>Word limits can be fuzzy, especially for short copy.</li><li>Character limits are a bit more precise — but still not perfect.</li><li>You might still need to prompt it to revise, trim, or try again, especially when working with small spaces.</li></ul><h3>What are “higher-level constraints”?</h3><p>Not all instructions are created equal. Some are pixel-perfect (like “stay under 50 characters”), and some are more like “go write me something that <em>feels</em> like a paragraph.”</p><p>LLMs tend to do better with the second kind.</p><p>These are what I call higher-level constraints — structural cues based on patterns, not math. Think:</p><ul><li>“Write 2 short paragraphs”</li><li>“Use 3 bullet points”</li><li>“Make this sound like a summary”</li><li>“Explain it like I’m five” (seriously, that one works)</li></ul><p>Models are <em>really</em> good at these because they’ve seen millions of examples in their training data. They know what a paragraph looks like. They understand how a bulleted list usually works. They can vibe with that.</p><p>Now contrast that with low-level constraints — the nitpicky stuff like:</p><ul><li>“Use exactly 20 words”</li><li>“Stay under 60 characters”</li><li>“Use no more than 3 sentences”</li></ul><p>These are way harder for an LLM to hit precisely, because it generates one token at a time and doesn’t have a built-in word or character counter (unless you force it to count, which, yep, you can).</p><p>So if you’re asking for super-tight copy, the model might miss. Not because it’s bad — but because it’s not built to obsess over character limits like we are.</p><h3>What does that look like in reality?</h3><p>Let’s put it into practice. Same model, different types of constraints — very different results.</p><h4>1. A two-paragraph explanation</h4><p><strong>Prompt:</strong> “Explain this concept in two short paragraphs.”<br><strong>What you’ll get: </strong>Two clean chunks of text, like it read your mind. LLMs are great at this because it’s a pattern they’ve seen a bazillion times. It just <em>gets</em> what a paragraph should look like.</p><h4>2. A 100-word summary</h4><p><strong>Prompt:</strong> “Summarize this article in 100 words.”<br><strong>What you’ll get: </strong>Something <em>close</em> — maybe 95, maybe 110. Good enough for most cases. If you need it to be exactly 100, just ask it to count and revise. (It won’t mind.)</p><h4>3. A 50-character notification</h4><p><strong>Prompt:</strong> “Write a push notification under 50 characters.”<br><strong>What you’ll get: </strong>Something like: “Don’t miss out — grab yours now!”<br> Which <em>feels</em> short (and is — it’s 39 characters). Great! But if you asked for 30 characters? It might overshoot unless you’re really clear, or follow up with “Try again, this time under 30 characters, including spaces.”</p><h3>Why the difference?</h3><p>Because LLMs are pattern mimics, not calculators.<br>They don’t count words or characters unless you ask them to — and even then, they sometimes just… estimate.</p><p>But they’ve read the internet and beyond. They <em>know</em> what a paragraph sounds like. They <em>know</em> what a summary usually looks like. And they’re really good at giving you that structure — as long as you ask the right way.</p><h3>Matching the constraint to the content</h3><p>Choosing the right constraint isn’t just a words-vs-characters thing. It’s about matching the type of limit to the thing you’re writing — and the space it needs to live in.</p><p>Here’s a simple way to think about it:</p><h4>Use strict character limits when space is tight</h4><p>This is your go-to for microcopy: button labels, menu items, push notifications — anything that lives in a tiny UI box. The layout doesn’t care how many words you use. It cares how wide the text is. That’s why character limits are your best friend.</p><p><strong>Example: </strong><em>“Write a call-to-action for a mobile button (max 30 characters).”</em></p><p><strong>Pro tip:</strong> Add the limit in parentheses or bold. Be bossy. LLMs love clear direction.</p><p>If the model overshoots? Just follow up with:</p><p><em>“Cool. Now rewrite that to fit within 25 characters, including spaces.”</em></p><p>And yes, this is especially important when designing in Figma or reviewing live copy. Word count won’t break your layout. Character count will — especially in translation.</p><h4>Use flexible word ranges when tone matters more than pixels</h4><p>If you’re writing something a bit longer — like a tooltip, onboarding tip, or friendly error message — word ranges give the model room to breathe, without letting it ramble.</p><p><strong>Examples</strong>: <em>“Write a friendly error explanation in 25–30 words.”</em><br><em> “Write a tooltip that explains what happens after submitting the form (max 40 words).”</em></p><p>This gives the model structure <em>and</em> creative wiggle room — which is great when your copy needs to sound like a human, not a string of keywords in a trench coat.</p><h4>Use paragraph-level constraints for longer, structured content</h4><p>Need a summary? An intro? A nice, tight blurb? LLMs are excellent at higher-level patterns — like “two paragraphs” or “three bullet points.” They’ve seen those shapes a million times and know how to deliver the vibe.</p><p><strong>Examples</strong>: <em>“Write a two-paragraph intro to this article.”</em><br><em> “Explain this concept in three short bullet points.”</em></p><p>Paragraph and list formats are <em>super</em> reliable because they’re baked into the model’s training. And if it runs a little long? Easy — just ask it to tighten up.</p><h3>TL;DR (in list form, obviously):</h3><figure><img alt="A table showing different content types and their best-fit length constraints. Microcopy, translated UI strings, and push notifications require strict character limits. Tooltips use word limits for flexibility. Onboarding screens and summaries work best with paragraph structures or word counts for readability." src="https://cdn-images-1.medium.com/max/1024/1*9n3Qw2GJCIyJkZlDfAwYzA.png" /><figcaption>The secret to AI-friendly copy? Matching the right constraint to the right content.</figcaption></figure><h3>What about translations?</h3><p>Ah, localization — where your perfect little string goes on a European vacation and comes back 30% longer.</p><p>This is where character limits absolutely matter. Why? Because languages stretch. Or shrink. Or add formal fluff. And if you’re not accounting for that, your lovely, crisp English copy might explode the layout in German, or feel awkwardly short in Swedish, or vanish into minimalism in Chinese.</p><p>Let’s look at how things can shift:</p><ul><li><strong>English → German?</strong> Expect longer compound words and a ~30% expansion.</li><li><strong>English → French?</strong> More formal phrasing = more space.</li><li><strong>English → Chinese?</strong> Usually contracts, but each character is carrying more meaning.</li></ul><p><strong>Example 1:</strong> “See offers” (9 characters)<br> → becomes “Alle Angebote ansehen” (24 characters) in German.<br> Cute in English. Chaos in the UI.</p><p><strong>Example 2:</strong> “Add to bag” (11 characters)<br> → becomes “In den Warenkorb” (20 characters).<br> And that’s if you’re lucky.</p><p>This is why character limits are the fairest constraint you can use across languages. Word count doesn’t help here — one language’s word might be another’s three. But characters? At least they’re measurable.</p><p>So when you’re writing copy that’s going to be translated, always ask:</p><p>“Will this still fit if it grows by 30%?”</p><p>And if not? Give your LLM a character cap, and hold it to it.</p><h3><strong>Final takeaway</strong></h3><p>LLMs are powerful, but they’re not psychic. They don’t know your button has a 25-character limit, or that your beautifully minimal English string turns into a German sentence with elbows.</p><p>So don’t just say “keep it short.”</p><p>Say how short. Or how structured. And say it clearly.</p><ul><li>Use character limits when space is tight.</li><li>Use word ranges when you need room for tone.</li><li>Use paragraph or list formats when you want flow and structure.</li></ul><p>Match the constraint to the content, and your AI writing buddy will meet you halfway — no layout breaks, no lost-in-translation moments, no tears in Figma.</p><h4>Want to go deeper?</h4><p>I’m sharing scrappy tips and silly-smart strategies for making AI tools actually work in real-world content design. Follow <a href="https://medium.com/@fluturelle">me</a> on Medium if you’re into that kind of thing.</p><ul><li><a href="https://medium.com/design-bootcamp/scrappy-but-smart-my-hack-for-brand-perfect-ai-translations-f04421e3a7b2">Scrappy but smart: My hack for brand-perfect AI translations</a></li><li><a href="https://medium.com/@fluturelle/the-ai-prompt-that-writes-wcag-compliant-alt-text-b59847899aaf">The AI prompt that writes WCAG-compliant alt text</a></li><li><a href="https://medium.com/design-bootcamp/how-i-used-ai-to-turn-figma-designs-into-translation-files-saving-hours-of-work-0ac063b4da64">How I used AI to turn Figma designs into translation files — saving hours of work</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82f6cddf34b6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Scrappy but smart: My hack for brand-perfect AI translations]]></title>
            <link>https://medium.com/design-bootcamp/scrappy-but-smart-my-hack-for-brand-perfect-ai-translations-f04421e3a7b2?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/f04421e3a7b2</guid>
            <category><![CDATA[chatgpt-prompt]]></category>
            <category><![CDATA[content-design]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[localization]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Mon, 17 Mar 2025 21:43:08 GMT</pubDate>
            <atom:updated>2025-03-18T11:26:11.131Z</atom:updated>
            <content:encoded><![CDATA[<h4>Making something out of nothing — how I hacked AI into a brand-trained translator without style guides, glossaries, or brand resources.</h4><figure><img alt="Illustration of a person sitting cross-legged with a laptop displaying the ChatGPT logo. Surrounding icons include a Swiss Army knife, a document with “URL,” a translation icon, and a gear." src="https://cdn-images-1.medium.com/max/1024/1*W-ioua5Wym0R8pLSzRb1VA.png" /></figure><p>Translating content for multiple brands, industries, and audiences can feel impossible without a style guide, glossary, or brand tone-of-voice documentation. But what if you could still produce high-quality, on-brand localizations — even with 0 formal resources?</p><p>I found a hack that allows me to quickly scale translations for multiple brands and industries, using nothing but <strong>ChatGPT</strong> and a few smart tricks. This approach is for anyone who suddenly finds themselves responsible for translations — whether you work in marketing, product, UX, or customer support — without access to localization resources.</p><h3>The challenge: Translating without brand material — fast</h3><p>Brands have a unique <strong>voice</strong>, <strong>mission</strong>, and <strong>way of speaking</strong> to its audience. Usually, translators rely on brand <strong>glossaries</strong>, <strong>style guides</strong>, and <strong>reference material</strong> to ensure consistency.</p><p>But what happens when:</p><ul><li>You don’t have access to these materials?</li><li>You have to translate for multiple brands in a short time?</li><li>You’re working outside your industry expertise?</li></ul><p>This was my reality. I had to translate product descriptions, campaign emails, and customer surveys into German— each for a different industry and audience, from fashion to finance, tech to wellness. And I had to do it fast, without the luxury of deep brand research.</p><h3>The hack: Turning ChatGPT into a brand-trained translation assistant</h3><p>Instead of guessing a brand’s voice, I found a fast and scalable way to train ChatGPT for each brand before translating. Here’s how:</p><p><strong>Step 1: Create a dedicated ChatGPT conversation for each brand</strong></p><ul><li>Open a new ChatGPT chat for each brand you’ll be translating for.</li><li>Keep this chat exclusive to that brand so it builds memory over time.</li></ul><p><strong>Step 2: Gather relevant [target language] references from that brand</strong></p><p>If you don’t have official brand materials, you can use publicly available content instead:</p><ul><li>The brand’s website (homepage, product pages, About us pages)</li><li>Campaign emails and promotional content</li><li>Social media posts (if in target language)</li><li>Customer-facing FAQ pages</li><li>Microcopy examples (headlines, CTAs, labels)</li><li>Screenshots and any copy examples that are similar to the content you’ll request translations for</li></ul><p>By doing this, ChatGPT can analyze the brand’s tone, style, word choices, and industry-specific terminology — without needing an official style guide.</p><p><strong>Step 3: Ask ChatGPT to analyze the brand’s linguistic style before providing translations</strong></p><p>Here’s an example of the prompt I used:</p><figure><img alt="Screenshot of a ChatGPT conversation showing a structured prompt. The prompt instructs ChatGPT to analyze brand materials from multiple URLs and generate an AI-friendly style guide for localization into German. The message includes a list of brand URLs and sections like “Brand Tone &amp; Voice” and “Industry Terminology." src="https://cdn-images-1.medium.com/max/1024/1*OgE__kGTn079vaSHEAaeEg.png" /><figcaption>Uploading brand-specific URLs to teach ChatGPT its tone, style, and terminology.</figcaption></figure><p>And here for you to copy:</p><p><em>Before running the prompt, update the target language and upload the collected brand materials to ensure accurate and brand-aligned translations.</em></p><blockquote><em>Analyze the provided brand materials and create a comprehensive style guide that captures both the brand identity and linguistic rules for AI-powered localization.<br>The style guide should be structured for AI training and professional translator-level localization, ensuring AI-generated translations sound natural, fluent, and brand-aligned.<br>The target language is [TARGET LANGUAGE].</em></blockquote><blockquote><strong><em>Brand material</em></strong><em><br>[Insert brand URLs in target language here]</em></blockquote><blockquote><em>Include the following sections:</em></blockquote><blockquote><strong><em>Brand Tone &amp; Voice</em></strong><em><br>- Define the brand’s tone (formal/informal, friendly/professional).<br>- Describe the voice (motivational, informative, expert-driven).<br>- Explain how the brand builds trust and engagement with its audience</em></blockquote><blockquote><strong><em>Customer Address &amp; Engagement Style</em></strong><em><br>- Specify whether the brand uses formal (‘Sie’, ‘vous’, etc.) or informal (‘du’, ‘tu’, etc.) address in [TARGET LANGUAGE].<br>- Describe how the brand connects with customers (e.g., direct, encouraging, educational, promotional).<br>- Identify whether personalization (e.g., second-person pronouns, rhetorical questions) is a key feature.</em></blockquote><blockquote><strong><em>Sentence Structure &amp; Writing Mechanics</em></strong><em><br>- Define preferred sentence length and structure (short &amp; direct vs. detailed &amp; elaborate).<br>- Specify usage of active vs. passive voice.<br>- Define rules for punctuation, contractions, and readability in [TARGET LANGUAGE].</em></blockquote><blockquote><strong><em>Recurring Phrases &amp; Signature Expressions</em></strong><em><br>- Extract key phrases that should remain consistent across translations.<br>- Highlight brand taglines, slogans, and key messaging points.<br>- Identify phrases that should be transcreated instead of literally translated.</em></blockquote><blockquote><strong><em>Industry-Specific Terminology &amp; Key Translations</em></strong><em><br>- Identify consistent terminology for products and features (e.g., ‘planner’ = ‘Kalender’ or ‘Planer’ in German, ‘agenda’ in French).<br>- Specify terms that should not be freely translated but kept uniform across all content.<br>- Highlight potential false friends or mistranslations to avoid.</em></blockquote><blockquote><strong><em>Formatting &amp; Structure</em></strong><em><br>- Define the use of headlines, bullet points, paragraph breaks, and list formatting.<br>- Provide guidance on text hierarchy, capitalization, and emphasis styles (e.g., title case vs. sentence case).<br>- Indicate how numbers, dates, and measurement units should be formatted in [TARGET LANGUAGE].</em></blockquote><blockquote><strong><em>Cultural Adaptation Guidelines</em></strong><em><br>- Explain how to adapt cultural references while keeping brand authenticity.<br>- Highlight trust-building elements like sustainability, regional credibility, and local audience preferences.<br>- Identify any specific localization challenges for [TARGET LANGUAGE] (e.g., gendered language, formality differences).</em></blockquote><blockquote><strong><em>CTA &amp; Microcopy Strategy</em></strong><em><br>- Define how CTAs should be worded (imperative verbs, inviting tone, button labels).<br>- Provide examples of effective microcopy that align with the brand’s tone.<br>- Specify if CTA phrasing should be direct (‘Buy now’) or softer (‘Discover our products’) in [TARGET LANGUAGE].</em></blockquote><blockquote><em>Ensure the style guide is structured in a clear, actionable way for AI training and can be applied to both AI-generated and human-edited translations.</em></blockquote><blockquote><strong><em>Final step</em></strong><br>Once the style guide is complete, apply the defined rules by localizing the following text(s) into [TARGET LANGUAGE], ensuring the localization sounds like it was written by a professional translator and not a machine.</blockquote><p>Here is a suggested style guide from ChatGPT after running my prompt:</p><figure><img alt="Screenshot of a style guide titled “Burde Localization Style Guide.” The document details the brand’s tone and voice, customer engagement style, sentence structure guidelines, recurring phrases, industry-specific terminology, formatting rules, and cultural adaptation guidelines. The German-translated content adheres to these specifications" src="https://cdn-images-1.medium.com/max/1024/1*6-txdj8NggqNopTqN_N8rA.png" /><figcaption>The AI-friendly style guide: Defining brand tone, key terminology, and linguistic guidelines to ensure high-quality, brand-aligned AI translations.</figcaption></figure><p><strong>Good to know</strong>: Unlike traditional style guides meant for human translators, this AI-friendly format ensures clear, rule-based instructions that AI can systematically apply without ambiguity. Since AI lacks human intuition, structuring guidelines in a way that removes subjectivity leads to more consistent, scalable, and high-quality translations, without constant manual intervention.</p><p>Once you have AI-generated the style guide, you can now add the text that needs to be localized. Here my example:</p><figure><img alt="creenshot comparing an English text prompt and its German AI-generated translation. The English text asks ChatGPT to translate according to the Burde style guide. The German translation reflects the requested style, tone, and phrasing, closely aligning with the brand’s guidelines." src="https://cdn-images-1.medium.com/max/1024/1*oQ5PkuEbCvtZgyH_8C0QUw.png" /><figcaption>The German localization perfectly reflects the brand’s tone and style — following the AI-trained style guide we just created.</figcaption></figure><h3>Why this works</h3><p>✅ <strong>Scales easily</strong> — Works for any brand across industries.<br>✅ <strong>Saves time</strong> — No need to manually “train” ChatGPT for each request.<br>✅ <strong>Improves with use</strong> — The more you interact, the better the translations become.<br>✅ <strong>Keeps conversations consistent</strong> — Instead of retraining ChatGPT each time, I could <strong>return to the same chat</strong> whenever a new translation request came in.<br>✅ <strong>Allows quick refinements</strong> — If something felt off, I corrected it in-chat, and ChatGPT learned from the adjustments.</p><p>The difference in quality between a generic AI translation and one trained with a brand-specific style guide prompt is significant. The comparison below shows how using my structured prompt leads to better adherence to brand tone, terminology, and readability:</p><figure><img alt="A side-by-side comparison of two German translations of the same text. The left column, labeled “#1 Translation without Style Guide,” contains a direct translation with minor inconsistencies in brand tone and phrasing. The right column, labeled “#2 Translation with Style Guide,” shows a refined translation that aligns better with brand language, sentence structure, and terminology." src="https://cdn-images-1.medium.com/max/1024/1*sH5djE3YI8DmltZtqSUayw.png" /><figcaption>The power of brand-specific AI training: Before and after ChatGPT learned the brand voice.</figcaption></figure><figure><img alt="A table comparing two German translations: one without a style guide and one using an AI-generated style guide. The columns include criteria such as “Headline,” “Tone &amp; Voice,” “Sentence Structure,” “Cultural Adaptation,” and “CTA Strategy.” The style guide-based translation consistently aligns better with brand tone, terminology, and clarity." src="https://cdn-images-1.medium.com/max/1024/1*-E-Eemq2gTAyEKxRKuL59A.png" /><figcaption>A side-by-side comparison showing how AI-generated translations improve with structured brand guidelines.</figcaption></figure><p>By incorporating brand-specific materials before prompting, ChatGPT learns and applies the brand’s linguistic patterns, microcopy, and tone — making translations feel more natural, consistent, and aligned with brand identity. This structured approach allows even those without deep localization expertise to produce high-quality, on-brand translations.</p><h3>What if no target language resources are available?</h3><p>If no existing references are available in the target language, you can still train ChatGPT effectively using source language materials. Instead of providing direct examples in the target language, guide ChatGPT to analyze the brand’s tone, style, and linguistic patterns in the source language and then apply those insights to the translation.</p><p>Here’s how:</p><p><strong>Step 1: Upload the brand’s resources in the original language (English, French, etc.).</strong></p><p><strong>Step 2: Ask ChatGPT to analyze the tone, voice, and linguistic style in that language before providing translations</strong></p><p>Here’s the adjusted prompt for this approach:</p><blockquote><em>Analyze the provided [SOURCE LANGUAGE] brand materials and create a comprehensive style guide for localizing content into [TARGET LANGUAGE], even though no existing translations are available.<br>The style guide should be structured for AI training and professional senior translator-level localization, ensuring AI-generated translations sound natural, fluent, and brand-aligned.<br>Since there are no reference materials in [TARGET LANGUAGE], derive the style guide from the original source materials and define how the tone, structure, and key elements should be adapted to match the target audience.</em></blockquote><blockquote><strong><em>Brand material</em></strong><em><br>[Insert brand URLs in source language here]</em></blockquote><blockquote><em>Sections to Include:</em></blockquote><blockquote><strong><em>1. Brand Identity &amp; Communication Style</em></strong><em><br>- Define the brand’s focus, values, and positioning based on the source materials.<br>- Identify core themes (e.g., productivity, creativity, empowerment).<br>- Summarize how the brand builds trust and engagement with its audience.</em></blockquote><blockquote><strong><em>2. Tone &amp; Voice Adaptation for [TARGET LANGUAGE]</em></strong><em><br>- Describe the tone (formal/informal, friendly/professional).<br>- Define how the brand voice should be adapted to [TARGET LANGUAGE] while preserving its personality.<br>- Provide examples of how source language tone should be translated naturally rather than literally.</em></blockquote><blockquote><strong><em>3. Customer Address &amp; Engagement Style</em></strong><em><br>- Determine whether formal (‘Sie’, ‘vous’, etc.) or informal (‘du’, ‘tu’, etc.) address is more appropriate for the brand in [TARGET LANGUAGE].<br>- Describe how the brand engages with customers (e.g., direct &amp; motivational, informative &amp; authoritative).<br>- Specify whether personal pronouns, rhetorical questions, or direct imperatives should be used in localization.</em></blockquote><blockquote><strong><em>4. Sentence Structure &amp; Writing Mechanics</em></strong><em><br>- Define preferred sentence length and complexity (e.g., short &amp; direct vs. descriptive &amp; engaging).<br>- Specify the use of active vs. passive voice in [TARGET LANGUAGE].<br>- Provide rules for punctuation, contractions, and readability to ensure fluency.<br>- Address any structural differences between the source language and [TARGET LANGUAGE] (e.g., subject placement, formality levels).</em></blockquote><blockquote><strong><em>5. Recurring Phrases &amp; Signature Expressions</em></strong><em><br>- Extract key phrases from the source text that should remain consistent across translations.<br>- Identify brand taglines, slogans, and key messaging points and define how they should be transcreated instead of literally translated.<br>- Provide examples of how specific phrases should be adapted for a natural, localized feel.</em></blockquote><blockquote><strong><em>6. Industry-Specific Terminology &amp; Key Translations</em></strong><em><br>- Define consistent terminology for products and features in [TARGET LANGUAGE] (e.g., “planner” = “Kalender” or “Planer” in German).<br>- Specify which terms should be kept in English (if applicable) and which should be localized.<br>- Highlight potential mistranslations, false friends, or cultural nuances that must be considered.</em></blockquote><blockquote><strong><em>7. Formatting &amp; Structure for Localization</em></strong><em><br>- Define the use of headlines, bullet points, paragraph breaks, and list formatting for readability.<br>- Provide guidance on capitalization, emphasis styles, and number/date formatting in [TARGET LANGUAGE].<br>- Adapt measurement units, currency, and region-specific formatting conventions where needed.</em></blockquote><blockquote><strong><em>8. Cultural Adaptation Guidelines</em></strong><em><br>- Identify cultural references in the source materials and suggest how they should be adapted or replaced in [TARGET LANGUAGE].<br>- Provide recommendations for trust-building elements (e.g., sustainability messaging, regional credibility).<br>- Highlight specific localization challenges in [TARGET LANGUAGE] (e.g., gendered language, register shifts, humor adaptation).</em></blockquote><blockquote><strong><em>9. CTA &amp; Microcopy Adaptation</em></strong><em><br>- Define how CTAs should be structured in [TARGET LANGUAGE] (e.g., direct ‘Buy now’ vs. softer ‘Discover our products’).<br>- Provide examples of natural-sounding CTA translations that align with the brand’s voice.<br>- Specify how button labels, navigation elements, and microcopy should be phrased for consistency.</em></blockquote><blockquote><em>Ensure the style guide is structured in a clear, actionable way for AI training and can be applied to both AI-generated and human-edited translations.</em></blockquote><blockquote><strong><em>Final step</em></strong><em><br>Once the style guide is complete, apply the defined rules by localizing the following text(s) into [TARGET LANGUAGE], ensuring the localization sounds like it was written by a professional translator and not a machine.</em></blockquote><p>This approach allows ChatGPT to extract brand traits from the source language and apply them to translations, even when no target language content is available. This works because tone and voice principles translate across languages. While terminology and phrasing will differ, ChatGPT can still maintain a brand’s core communication style when properly trained.</p><h3>The limitations (and why this is a hack, not a replacement for real translators)</h3><p>This approach is a workaround for missing resources — not a replacement for professional localization teams. Here’s where it falls short:</p><ul><li>AI still makes mistakes — It doesn’t replace native speaker intuition.</li><li>Requires initial setup — If no target language content is available, users must first train ChatGPT using source language content and guide it through refinements.</li><li>Human review is essential — While this approach streamlines translation, the person must act as the final reviewer, refining outputs for accuracy and cultural relevance.</li></ul><p>That said, for fast, scalable translations when no resources exist, this hack is incredibly effective.</p><h3>Final thoughts</h3><p>If you ever need to translate for multiple brands without access to a style guide, this trick can help you generate high-quality, on-brand localizations — fast.</p><p>Would love to hear: How do you handle translation challenges when working across different brands or contexts? Let me know in the comments!</p><p>PS. If you’re interested in more ways to make AI work for you, check out my friend <a href="https://medium.com/@bendaviesromano"><strong>Ben Davies</strong></a>, who recently published an article on how <strong>NotebookLM</strong> can be used to create a powerful, interactive style guide.</p><p>📖 <strong>Check out his article:</strong> <a href="https://medium.com/user-experience-design-1/making-ai-work-for-us-how-can-notebooklm-help-us-spread-good-design-thinking-4b5d79ca3dc0">Making AI work for us: How can NotebookLM help us spread good design thinking?</a></p><p>By uploading all available resources to the NotebookLM — like a brand’s URLs — it allows you to ask questions about the brand’s communication style, tone, and terminology. You might even use it for AI-powered translations!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f04421e3a7b2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/scrappy-but-smart-my-hack-for-brand-perfect-ai-translations-f04421e3a7b2">Scrappy but smart: My hack for brand-perfect AI translations</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI prompt that writes WCAG-compliant alt text]]></title>
            <link>https://medium.com/@fluturelle/the-ai-prompt-that-writes-wcag-compliant-alt-text-b59847899aaf?source=rss-f1ef3a81515e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b59847899aaf</guid>
            <category><![CDATA[wcag-accessibility]]></category>
            <category><![CDATA[ai-in-content-design]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[alt-text]]></category>
            <category><![CDATA[ux-writing]]></category>
            <dc:creator><![CDATA[Ilinca Apolzan]]></dc:creator>
            <pubDate>Mon, 10 Mar 2025 13:21:12 GMT</pubDate>
            <atom:updated>2025-03-17T22:47:33.280Z</atom:updated>
            <content:encoded><![CDATA[<h4>The key to better AI-generated alt text? A structured prompt with WCAG guidelines.</h4><figure><img alt="llustration of AI-powered alt text generation: images on the left, an arrow pointing to an icon of an image with a speech bubble." src="https://cdn-images-1.medium.com/max/1024/1*BD3MhlhdCFWglAI5KiGrUA.png" /></figure><p>Writing high-quality, <strong>WCAG-compliant alt text</strong> is time-consuming, but AI can help — if you give it the right instructions. In this article, I’ll share the exact system prompt I developed to generate accurate, accessible, and meaningful alt text for images.</p><p><strong>Why this matters:</strong></p><ul><li>Poorly written alt text harms accessibility.</li><li>AI-generated alt text is often <strong>incomplete or inaccurate</strong>.</li><li>A well-structured <strong>system prompt</strong> can drastically improve AI-generated descriptions.</li></ul><p><strong>Further reading:</strong> <a href="https://www.w3.org/TR/WCAG21/#text-alternatives">Web Content Accessibility Guidelines (WCAG) 2.1</a> | <a href="https://webaim.org/techniques/alttext/">WebAIM Alt Text Guide</a></p><h3>Two approaches: Embedded guidelines vs. external retrieval</h3><p>When generating WCAG-compliant alt text with AI, there are <strong>2 main ways to provide guidelines</strong>:</p><ol><li><strong>Embedding the guidelines directly into the system prompt</strong> — Best for <strong>self-contained</strong> AI models where full compliance is required in every response.</li><li><strong>Using external guidelines as retrievers</strong> — Ideal when the system prompt cannot hold all the details, and AI needs to pull information from a separate document.</li></ol><p>Both methods have use cases, and I’ve tested both. Below, I share both versions so you can choose the best fit.</p><h3>Option 1: AI with attached WCAG guidelines</h3><p>This approach works best when using an AI model like <strong>ChatGPT-4o</strong>, where you attach the WCAG guidelines as a separate document.</p><p><strong>Prompt for using external guidelines:</strong></p><blockquote>You are an expert at creating accessible and inclusive image descriptions.</blockquote><blockquote>The description you provide should focus on the purpose and essential information of the image, adhering to WCAG 2.1 Level A requirements for alt-text. Use concise, clear, and plain language to describe the key details of the image, avoiding unnecessary or redundant information. Ensure the description aligns with the image’s context and avoids phrases like ‘image of’ or ‘picture of.’</blockquote><blockquote>For images containing text, include the visible text verbatim if it adds meaning to the content. Focus on what the image conveys and how it serves its purpose in the broader context. For colors, include them only if they are critical to the meaning of the image.</blockquote><blockquote>Respond using plain text, without markdown formatting, and structure your output as follows:<br>- Suggest three alt-text options that meet WCAG requirements (each under 125 characters, focusing on the image’s key content or purpose).<br>- Suggest three context-relevant captions for the image.</blockquote><p><a href="https://adorable-hole-e18.notion.site/AI-instructions-WCAG-compliant-ALT-text-guidelines-1b2061cf7b5f80228574f2515a3e8a09?pvs=4"><strong>Download WCAG guidelines here</strong></a></p><p><strong>Best for:</strong> AI systems where external documents can be attached, making them retrievable when needed.</p><figure><img alt="A screenshot of a ChatGPT conversation discussing WCAG-compliant alt-text and captions. The conversation includes an image of a hand holding a smartphone." src="https://cdn-images-1.medium.com/max/1024/1*t2b9IxZA-K2U0QaW_pP7SQ.png" /></figure><h3>Option 2: Fully embedded WCAG guidelines in the prompt</h3><p>This approach is best when the AI model <strong>cannot access external documents</strong> and needs all the necessary rules embedded.</p><p><strong>System prompt for WCAG-compliant alt-text generation</strong></p><blockquote>You are an expert in accessible and inclusive image descriptions, following WCAG 2.1 Level A guidelines.</blockquote><blockquote><strong>Key principles:</strong><br>1. Be descriptive but concise — Focus on key details; avoid excessive information.<br>2. Match the image’s purpose — Informative images describe key elements, functional images describe the action, decorative images use alt=””.<br>3.Avoid redundant phrases — Skip “Image of” or “Picture of”.<br>4. Include text within images (if essential).<br>5. Keep alt-text under 125 characters when possible.<br>6. Ensure consistency — Use uniform descriptions for buttons/icons across platforms.</blockquote><blockquote><strong>Response format:</strong><br>- Three alt-text suggestions (each under 125 characters)<br>- Three context-relevant captions for the image</blockquote><p><strong>Best for:</strong> AI models where the entire workflow needs to be self-contained, ensuring no reliance on external retrieval.</p><figure><img alt="A screenshot of a ChatGPT conversation about WCAG-compliant alt-text, featuring an image of a gold key with a lipstick-shaped keychain against a blue door." src="https://cdn-images-1.medium.com/max/1024/1*iMzHB3IgEi0Jjgv_XF3f1w.png" /></figure><h3>Examples of AI-generated alt text (before &amp; after)</h3><p>To show the impact of these prompts, here are some <strong>before-and-after</strong> examples:</p><figure><img alt="A table comparing default AI-generated alt-text with custom alt-text based on user prompts. It contains three images: a UX designer holding a laptop and gesturing at wireframe sketches, a surreal digital artwork of a floating landscape with trees and flowers, and a person paying with a Visa card at a contactless terminal." src="https://cdn-images-1.medium.com/max/791/1*S78foeFIsUXsL6YDDJbg6w.png" /></figure><h3>How you can use these prompts</h3><p>If you work in content design, accessibility, or localization, these prompts can help you:</p><ul><li>Generate <strong>faster, more accurate alt text</strong> for large image libraries.</li><li>Ensure <strong>compliance with WCAG</strong> guidelines.</li><li>Improve <strong>AI-driven workflows</strong> in content creation.</li></ul><p><a href="https://adorable-hole-e18.notion.site/AI-instructions-WCAG-compliant-ALT-text-guidelines-1b2061cf7b5f80228574f2515a3e8a09?pvs=4"><strong>Download the full WCAG guidelines for AI use</strong></a></p><h3>Final thoughts</h3><p>These two approaches offer <strong>flexibility</strong> in how AI generates alt-text — whether by <strong>embedding guidelines</strong> or <strong>retrieving them externally</strong>.</p><p><strong>Next article coming soon:</strong> How I tested and improved this system prompt. Stay tuned!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b59847899aaf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>