<?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 Brian Kelson on Medium]]></title>
        <description><![CDATA[Stories by Brian Kelson on Medium]]></description>
        <link>https://medium.com/@briankelson?source=rss-dcba620c0db9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*d5KYqe8o96MYvewq58WzLA.jpeg</url>
            <title>Stories by Brian Kelson on Medium</title>
            <link>https://medium.com/@briankelson?source=rss-dcba620c0db9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 14:38:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@briankelson/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[The High Cost of the “Fraud” Defense: An Honest Discussion on Imposter Syndrome in Corporate…]]></title>
            <link>https://briankelson.medium.com/the-high-cost-of-the-fraud-defense-an-honest-discussion-on-imposter-syndrome-in-corporate-1acd5b7c3c19?source=rss-dcba620c0db9------2</link>
            <guid isPermaLink="false">https://medium.com/p/1acd5b7c3c19</guid>
            <category><![CDATA[corporate-culture]]></category>
            <category><![CDATA[self-love]]></category>
            <category><![CDATA[imposter-syndrome]]></category>
            <category><![CDATA[authenticity]]></category>
            <dc:creator><![CDATA[Brian Kelson]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 23:16:22 GMT</pubDate>
            <atom:updated>2026-09-25T10:18:59.569Z</atom:updated>
            <content:encoded><![CDATA[<h3>The High Cost of the “Fraud” Defense: An Honest Discussion on Imposter Syndrome in Corporate America</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nYkjGjJ27k8oeX6LvoWmnw.jpeg" /><figcaption>A fond memory from the archives: a research workshop I hosted once… happy, empowered faces! ❤</figcaption></figure><blockquote>“The whole problem with the world is that fools and fanatics are always so certain of themselves, and wiser people so full of doubts.”</blockquote><blockquote>— Bertrand Russell</blockquote><p>In the early days when user experience design was catching fire as a must-have skill across global corporations, well before UX had a real and influential voice at the product table, there was a predictable influx of professionals suddenly adopting the title of “UX Designer.” It is a familiar pattern, not unlike the sudden wave of “AI Experts” we see today.</p><p>The trouble was that comparing UX Professional A to UX Professional B using a simple “fraud or not” gut check was almost impossible. You couldn’t judge their capabilities accurately until you worked directly beside them and gained a clear view of their methods, philosophies, and hands-in-the-mud output. Even then, it took time to distinguish between their strengths and the areas of continuous (or the lack of), committed growth.</p><p>That single principle; an authentic, conviction-driven pursuit of relentless iterative self-improvement — in my humble opinion, is the real differentiator separating those who merely hold a title from those who embody real demonstrable substance in their craft.</p><blockquote>“The inner critic often tries to be the first to hit us as a defense mechanism, hoping that self-inflicted criticism will somehow cushion the blow of external judgment. But internal harm used as a shield against potential external harm is no solution at all, leaving us with the harsh reality that this persistent self-shame cycle disguised as self-care has been the chief perpetrator of damage all along.”</blockquote><blockquote>— Me</blockquote><h3>The Accusers in My Psyche</h3><p>So why do I bring this up?</p><p>As a young professional, husband, and father of four little kids, the notion of a formal academic accreditation was completely out of reach for me both financially and logistically. When I made the decision to shift my career toward a more science-based approach to design decisions, it was driven in equal parts by a sincere conviction that it was a superior way to do meaningful design work and by, frankly, basic financial necessity. UX roles paid significantly more than the standard graphic/web design roles that I built my early career upon. With the growing price tag of a large family, it was a weighty variable that I couldn’t pretend was absent from the decision matrix.</p><p>Because of this, two distinct accusers took root in my psyche:</p><ul><li><em>“You don’t have a degree… FRAUD!”</em></li><li><em>“You only did this for the money… FRAUD!”</em></li></ul><p>Rather than facing these voices through self-reflection or professional guidance, I suppressed them. I coped by allowing the anxiety to crank up the pressure, forcing myself to study hard and excel. Little did I realize at the time that this strategy only amplified these shaming voices, making them more and more convincing.</p><p>“Well, that’s great, Brian,” you might say. “It helped you excel in your career at least!”</p><p>Sure. There can absolutely be real financial benefits for people driven by fear and anxiety — constantly trying to prove to their peers and internal demons that they “belong”. But at what cost? We’ll come back to that question in a moment.</p><p>The oddly comforting news is that while this idea of a fraudulent identity feels incredibly isolating and lonely… research published in the <em>International Journal of Behavioral Science </em>estimates that roughly <strong>70% of people experience imposter syndrome at some point in their working lives.</strong> A similar study by Korn Ferry revealed that a whopping <strong>71% of U.S. CEOs experience symptoms of imposter syndrome</strong>, proving that even senior leaders struggle with self-doubt despite high achievement.</p><p>Anxiety is a remarkably effective driver of corporate outcomes. Dare I say… a silent staple of capitalism even. However, it operates with complete disregard for the long-term, debilitating impact it leaves on the individual. And we certainly shouldn’t, though quite tempting, blame profit-driven organizations for respecting us less than we respect ourselves, especially when our insecurities drive us to push harder than our peers, turn the heads of c-suite executives and “exceed expectations” in our yearly performance reviews. The unspoken tragedy is how few of the people who suffer from these internal accusers actually get to enjoy the financial security their nervous systems fought so desperately hard to earn. On the contrary, it’s more like this unattainable carrot waved in front of you, promising an eventual sense of belonging and deserved value that only taunts louder the higher one climbs in their career. Imposter syndrome has no respect for CEO titles, honors, or six-figure salaries; it simply wakes you up at 3:00 AM to remind you how thoroughly you’ve “pulled one over” on everyone, urging you back to the grindstone to keep proving them wrong.</p><p>But I’m here to tell you… there is a much healthier, more sustainable option to achieve professional milestones you can be proud of without sacrificing your ability to live the life your hard work paid for. But most importantly… an option that won’t demand you trade your personal mental, emotional and physical health for each title and pay bump, or for the (highly over-rated) applause and admiration of colleagues and leaders that, if attained, doesn’t scratch the itch you thought it would anyway.</p><h3>A Fork in the Road</h3><p>I learned this lesson the hard way at a major U.S. bank. That is… the real potential costs of an unrelenting, imposter-syndrome-fueled nervous system left unchecked in one’s life.</p><p>One Thursday morning, I sat down at my cubicle and called out to my neighbor — we’ll call him “Jerry” — for our usual morning banter. Silence. Jerry was always at his desk before me and always left the office after me. He wasn’t part of the “star UX’er” club (or clique), but he worked tirelessly, always said “yes” to extra work, and somewhat reliably delivered on that promise.</p><p>Confused, I looked over the partition. His entire cubicle was completely bare, as if he had never been there at all!</p><p>My coworkers and I soon learned the horrifying truth: during one of his typical late-night work sessions, the cleaning crew found Jerry passed away at his desk from a heart attack. It was a devastating blow, one that few of us could push aside to jump back on the horse and get work done. I was physically sick to my stomach for several days, and could not bring myself to return to the office just to sit next to his hauntingly empty cubicle.</p><p>I’m not here to tell you that I know with certainty that his primary cause of death was directly tied to imposter syndrome. I’m no doctor, and those medical details were not disclosed to me. But I will offer my own observations for you handful of readers that have made it this far. Jerry had been separated from his family to take this gig. He had been living in a tiny apartment in Charlotte working long hours to send every penny back home to his wife and kids. I had worked adjacent to Jerry long enough to observe patterns of insecurity and a familiar smelling desperation to achieve some plateau where he might eventually relax into some sense of security that comes from the corporate god’s giving you a thumbs up. I only feel somewhat qualified to identify these characteristics because they were not unlike my own at the time. But regardless of the primary contributing health factors causing his heart attack, there was no denying that Jerry’s stress and anxiety meter was cranked higher than the dial was designed to go. I think of Jerry often, and I share his memory here with the deepest of respect. May he rest in peace.</p><p>As a man in my mid-thirties who was already having troubling conversations with doctors about elevated blood pressure, Jerry’s death hit me with blunt force. It propelled me into an uncomfortable binary fork in my life: continue down the same path of anxiety-driven “fraud defense” and risk a similar fate, or make a conscious decision to get off of this merry-go-round, do some meditation, educate myself, and take ownership of this internal world where I’d been living as a passenger in a car speeding towards a cliff.</p><blockquote>“I think the thing I’d like to have known early in my career is that <strong>no one knows what they’re doing.</strong> I used to think, ‘Oh, when I grow up I’ll know what I’m doing,’ because obviously everyone else seems to know what they’re doing — but they don’t. No one knows what they’re doing, even me. I don’t know what I’m doing. It’s not like you get to a level where you’re like, ‘All right, now I’m ready to start my thing.’ Like, when are you ready? I’m not ready, no one’s ready. You know, <strong>you just have to get comfortable with going in and doing it.”</strong></blockquote><blockquote>— Jacob Collier</blockquote><h3>A Different Model of Leadership</h3><p>During this period of recalibration, I had the privilege of working alongside a brilliant, deeply humane colleague named <a href="https://www.linkedin.com/in/cary-brunsman/">Dr. Cary Brunsman</a>. While working side-by-side on complex customer problems, we would occasionally share vulnerable stories about our career paths and our experiences navigating heavily regulated financial environments.</p><p>During one of these conversations, Cary admitted that she, too, struggled at times with a frustrating sensation of being an imposter.</p><p>Honestly, my initial reaction was to dismiss it. I assumed she was just being modest to put me at ease. <em>There is no way she struggles with this,</em> I thought. <em>She literally has a PhD and leads some of the highest-profile strategic initiatives in our department.</em></p><p>Well, it didn’t take long interacting with her before I quickly realized that she was truly the “WYSIWYG” of human beings, in the most down-to-earth and refreshingly disarming of ways. She was driven by deep conviction, said what she meant, and wasn’t afraid to stand by an unpopular opinion if it reflected her own convictions and understanding of the truth. To me, she was a representation of an academic who actually used their credentials to put their hands in the mud, driven by an earnest desire to solve real problems for real people, rather than as a tool for status, class or condescension. She never hid her imperfections, she failed openly so that others could learn with her… and it was that combination of humility, bravery and useful intelligence that struck me as an extraordinary model for effective, humane leadership; a quality that is sadly so deficient in this upside-down-Biff-Tannen-alternative-timeline we live in today. 🤣😭</p><p>I digress.</p><p>Inspired by her example and other humble academic coworkers and mentors I was incredibly fortunate to encounter along the way, I began to reclaim the wheel of my own car. I read a timely book by Valerie Young that obliterated any lingering skepticism of whether there were others like me grappling with this phenomenon. But it was a “slow and steady wins the race” kind of main character energy. The self-doubting voices didn’t vanish overnight. Honestly, they are still passengers in the back seat to this day, but they no longer get to drive, and I’m very happy they don’t fight me for that control anymore.</p><p>Things in my life took a really positive turn after that self-reflective chapter. I developed an audacity to allow more of my unconventional, true self to spill out into my projects, affect my professional partnerships and evolve how I led, supported and inspired the teams that I was accountable for. I reached an executive leadership level I once thought completely impossible, managing cross-functional teams (even a few PhDs reporting to me) along the way, shepherding initiatives that reached millions of customers at one of the world’s largest financial institutions. Where I once felt paralyzed by this notion of being surrounded by people smarter than me in their respective fields, I came to not only celebrate it, but to require it. Building high-performing teams meant surrounding myself with exceptional talent, cultivating a culture where we challenged and sharpened one other, guided by an unwavering belief that ux teams, even in the most stressful corporate environments, could actually love their jobs, care about one other and feel safe to take risks and learn from failures openly, together.</p><p>And the most meaningful achievements in my career so far… some magical moments of actual pride in the measurable human impact stories that get woven into their resumes… stories that remind them of why they got into this field to begin with. And looking back, I can say confidently that this kind of dream could never be fully realized when placed in the hands of leaders and teams predominantly crippled by inferiority complexes, competitive division or some persistent terror that one might get “found out”.</p><h3>Breaking the Cycle</h3><p>This brings me to why I name-dropped Cary Brunsman in this article. She is hosting an event that I strongly support; one that is near to my heart.</p><p>It’s an online coaching workshop designed specifically to address self-doubt and the fraudulent self-identity so many of us wear like a backpack full of bricks. It is an invitation to that healthier path I keep promoting in this article: a life and career where crippling anxiety is no longer in charge calling the shots, giving you space to reach your true potential without paying with your health or dignity.</p><p>Cary is offering this session at a generously accessible rate to reach as many people as possible. If the ways she has positively impacted my life and career are any indication; her experience, education, integrity, and deep empathy make her uniquely qualified to guide professionals through this emotionally tangled territory.</p><p>And if you take nothing else from this article… please hear this: <strong>Your value is not defined by your credentials, your title, your compensation, or the opinions of peers and colleagues.</strong></p><p>Yes, traditional corporate metrics may evaluate you solely on output and kpi’s, but true sustainability and authenticity starts with self-acceptance and self-love. You are the only person who can judge whether you are being authentic to or dismissive of your own convictions and values. But even if you discover areas where this fraudulent persona you believed in caused you to be an imposter to yourself… start at the most critical place; being kind to the you that has suffered with this for so long. It is never too late to start being true to the real you. Ok, now that I’m starting to sound like a Hallmark card, let’s wrap this up.</p><p>If you feel ready to start shifting this pattern, consider taking one practical step forward today. This is an incredibly small investment of your time and money, yet could very well become the fulcrum point where things truly begin to turn a corner for you.</p><p><strong>Join Cary Brunsman’s upcoming workshop:</strong></p><p>👉 <a href="https://www.eventbrite.com/e/from-self-doubt-to-self-trust-tickets-2000639651052?utm_source=gemini">From Self-Doubt to Self-Trust</a></p><p><strong>Also:</strong></p><p>📚 <a href="https://impostorsyndrome.com/resources/books/">Here’s a link to several helpful books</a></p><p>[Disclaimers (that no one will read) added 9/25/26]</p><p>a. If the tone in this article in any with lands with a “people should skip college and expect to rise to executive levels” vibe, please let me be clear that I have nothing but the utmost respect and adoration for people who have invested significant time and resources to achieve mastery in their craft through that path. My 48 year old self, still aspires to pursue a Bachelor’s and hopefully beyond that someday when my life path makes that possible. My hope is that my vulnerability expands the canvas of colors to more adequately respect the diversity and complexity of our stories. There is no “one size fits all”. Your unique you… is exactly what we all need from you. Go do the thing!!!!</p><p>b. I’m not bashful about the constantly experimental ways that I integrate AI tools into the things that I do. AI absolutely helped me to shape this article, but the ideas, stories, heart and accountability are 100% mine alone. In this extremely complex and cognitively dissonant space (that we are all faced with bearing the weight of when it comes to utilizing this technology ethically and responsibly), my humble hope is that if a tree was killed or a glass of water was sacrificed in the making of this article, that the value exchange for any human that was touched or impacted in doing so is a value exchange with some justifiable merit. This is a deeply vulnerable, human conversation… and I will always welcome a dialogue about it at brian@kelson.ai. My commitment to you is that I will not run from opinions that call into question the notion that might suggest that I could be a part of the “problem”. I’ve decided to make the persistent insecurities I grapple with on this topic a very public and honest one. Perhaps it will be my next article. 🤔</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1acd5b7c3c19" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Put Claude Code in a Sandbox — And It Changed How I Think About AI Tools]]></title>
            <link>https://briankelson.medium.com/i-put-claude-code-in-a-sandbox-and-it-changed-how-i-think-about-ai-tools-aac33d90d9ed?source=rss-dcba620c0db9------2</link>
            <guid isPermaLink="false">https://medium.com/p/aac33d90d9ed</guid>
            <category><![CDATA[information-security]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[sandbox]]></category>
            <dc:creator><![CDATA[Brian Kelson]]></dc:creator>
            <pubDate>Wed, 11 Feb 2026 18:23:14 GMT</pubDate>
            <atom:updated>2026-02-11T18:23:14.287Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6wulRY09nGTdXz5W_LF_Ew.jpeg" /><figcaption>The trial and error can be stressful, but keep going. It’s worth the achievement I promise you! 🤜🤛</figcaption></figure><p>I use Claude Code — Anthropic’s AI coding assistant that runs in your terminal.</p><p>It’s fast. It’s capable. It can read files, modify code, run commands, and generally behave like a very efficient developer sitting at your keyboard.</p><p>And that’s exactly why I decided to put it in a sandbox.</p><p>Not because I distrust Claude.</p><p>Not because I expect malicious behavior.</p><p>But because I believe in the principle of least privilege.</p><p>If a tool only needs access to one folder, why give it access to my entire machine?</p><p>So I built a Docker-based sandbox that isolates Claude Code to a single directory on my system. Then I documented the entire process. Then I stress-tested that documentation. Then I invited another AI to critique it. Then I revised it again. Now granted there is plenty of documentation out there to guide you through this setup (most notable is <a href="https://code.claude.com/docs/en/sandboxing">Anthorpic’s own documentation you can read here</a>…), but I found most of them lacking the hand-holding I felt I needed to comprehend the context and the rationale behind it all. If this helps just one other person out there, it was worth the effort to document!</p><p>This article isn’t just about Docker.</p><p>It’s about boundaries, verification, and the importance of writing documentation that survives contact with reality.</p><h3>Native Mode vs. Sandbox Mode</h3><p>By default, Claude Code runs natively.</p><p>That means:</p><ul><li>It has access to every file your user account can access.</li><li>Your Desktop, Documents, Downloads, SSH keys — all visible.</li><li>It runs as you.</li></ul><p>That might be fine. But it’s broad.</p><p>In sandbox mode, Claude runs inside a Docker container.</p><p>Only one folder on my machine is mounted into that container. Everything else is invisible — not blocked, not restricted — just nonexistent from the container’s perspective.</p><p>Conceptually, it looks like this:</p><blockquote>Host Machine<br> ├── ~/Desktop (not mounted)<br> ├── ~/Documents (not mounted)<br> ├── ~/Photos (not mounted)<br> └── ~/claude_sandbox ← only this is visible</blockquote><blockquote>Docker Container<br> └── /workspace → mapped to ~/claude_sandbox</blockquote><p>If the container can’t see a path, it can’t access it. That’s stronger than “please don’t.”</p><h3>Why I Documented Everything</h3><p>I didn’t just want a working setup.</p><p>I wanted something:</p><ul><li>Reproducible</li><li>Auditable</li><li>Verifiable</li><li>Understandable by someone else</li></ul><p>Because powerful AI tools deserve thoughtful boundaries — and good boundaries deserve good documentation.</p><p>I wrote a full step-by-step guide explaining:</p><ul><li>The Dockerfile</li><li>The compose configuration</li><li>The permission model</li><li>The SSH handling</li><li>The verification process</li><li>The difference between ephemeral and persistent containers</li><li>The common mistakes I personally made</li></ul><p>And here’s the important part:</p><p>The documentation changed as I used it.</p><h3>Where Things Broke (And Why That Was Valuable)</h3><p>At first, I used:</p><blockquote>docker compose run — rm claude-sandbox</blockquote><p>It worked.</p><p>But then things got confusing:</p><ul><li>Container names weren’t stable.</li><li>docker inspect didn’t behave how I expected.</li><li>Writable-layer state disappeared between sessions.</li><li>My verification workflow became unreliable.</li></ul><p>Named volumes persisted authentication tokens — which is good — but container-level configuration did not persist between runs.</p><p>That forced me to rethink the architecture.</p><p>I moved to a persistent container model:</p><blockquote>docker compose up -d<br>docker compose exec claude-sandbox claude</blockquote><p>That one change simplified everything:</p><ul><li>Stable container identity</li><li>Reliable inspection</li><li>Persistent writable-layer state</li><li>Faster session startup</li><li>Cleaner mental model</li></ul><p>The documentation improved because reality pushed back.</p><p>And that’s exactly what good documentation is supposed to survive.</p><h3>Verification: Don’t Let One AI Grade Its Own Homework</h3><p>Claude helped me design the sandbox.</p><p>But asking Claude to verify its own architecture didn’t feel right.</p><p>So I used a second LLM as an independent reviewer.</p><p>I created a canary file outside the sandbox and attempted to access it from inside the container.</p><p>I inspected mount points.</p><p>I verified the container was not running in privileged mode.</p><p>I confirmed the Docker socket wasn’t mounted.</p><p>I inspected HostConfig.Binds.</p><p>I forced observable proof instead of trusting a narrative.</p><p>That process changed how I think about AI-assisted engineering.</p><p>It’s not about trust.</p><p>It’s about verification.</p><h3>The Real Lesson: Documentation Is a Security Boundary</h3><p>The biggest insight wasn’t technical.</p><p>It was this:</p><p>If your documentation doesn’t anticipate mistakes, someone will make them.</p><p>In my case, I:</p><ul><li>Accidentally shadowed the native claude command with a shell function.</li><li>Confused ephemeral containers with persistent state.</li><li>Under-explained the difference between named volumes and writable layers.</li><li>Initially overstated one of my own technical claims.</li></ul><p>Every time something was unclear, the documentation improved.</p><p>That’s how it should work.</p><p>Documentation isn’t just instructions.</p><p>It’s the interface between intention and behavior.</p><h3>Why This Matters</h3><p>AI coding tools are powerful.</p><p>And power without boundaries isn’t innovation — it’s entropy.</p><p>Docker isn’t magic.</p><p>LLMs aren’t magic.</p><p>Security isn’t magic.</p><p>But layered boundaries, verified assumptions, and careful documentation?</p><p>That’s engineering.</p><h3>An Invitation to the Claude Code Community</h3><p>This setup works for me.</p><p>But I don’t assume it’s perfect.</p><p>If you try this process and:</p><ul><li>Run into confusing steps</li><li>Discover edge cases</li><li>See architectural improvements</li><li>Spot blind spots in my reasoning</li><li>Identify better verification strategies</li></ul><p>I want to hear about it.</p><p>The goal isn’t to defend the design.</p><p>The goal is to improve it.</p><p>If we’re going to build with AI tools that can read and modify our systems, we should hold ourselves to a higher standard of operational clarity.</p><p>If you’re using Claude Code — or any AI coding assistant — and thinking about isolation, reproducibility, or documentation quality, let’s compare notes.</p><p>Good systems thinking improves through feedback.</p><h3>Final Thought</h3><p>I didn’t sandbox Claude Code because I was afraid of it.</p><p>I sandboxed it because I respect it.</p><p>And powerful tools deserve thoughtful boundaries.</p><p>If you’re interested in the full technical guide, verification prompts, and setup templates, I’ve published the complete documentation in my repository.</p><p>Try it.</p><p>Break it.</p><p>Improve it.</p><p>Then tell me what you learned.</p><h3>The source files:</h3><ul><li>Main tree: <a href="https://github.com/bkelson/claude-code-sandbox/tree/main">https://github.com/bkelson/claude-code-sandbox/tree/main</a></li><li>Setup Guide: <a href="https://github.com/bkelson/claude-code-sandbox/blob/main/how-to-sandbox-claude-code.md">https://github.com/bkelson/claude-code-sandbox/blob/main/how-to-sandbox-claude-code.md</a></li><li>Claude Prompt Template (to have it walk you through the whole thing): <a href="https://github.com/bkelson/claude-code-sandbox/blob/main/sandbox-setup-prompt.md">https://github.com/bkelson/claude-code-sandbox/blob/main/sandbox-setup-prompt.md</a></li><li>“Third-Party-LLM QA Check Prompt” Template: <a href="https://github.com/bkelson/claude-code-sandbox/blob/main/sandbox-verification-prompt.md">https://github.com/bkelson/claude-code-sandbox/blob/main/sandbox-verification-prompt.md</a></li></ul><p>🧠 + 🤖 <em>This article was produced in partnership with AI tools. They helped me think faster. They did not think for me. Every conclusion, correction, and claim here is ultimately mine — and so is the responsibility.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aac33d90d9ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Idea That Seemed Simple: Enable Claude + ChatGPT Communication]]></title>
            <link>https://briankelson.medium.com/the-idea-that-seemed-simple-enable-claude-chatgpt-communication-b0cfca91753d?source=rss-dcba620c0db9------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0cfca91753d</guid>
            <category><![CDATA[github]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Brian Kelson]]></dc:creator>
            <pubDate>Sun, 01 Feb 2026 11:49:13 GMT</pubDate>
            <atom:updated>2026-02-02T13:18:39.591Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aHQSWNMh7LCSw0Biwv3wVA.png" /><figcaption>(Please excuse my cringy ai-generated header design 👆. All to illustrate that I don’t plan to hide my utilization of this technology in this new chapter of my career. I want to report on the good, bad and the.. well ugly.)</figcaption></figure><p><strong>The Idea That Seemed Simple: Enable Claude + ChatGPT Communication</strong></p><p>You know that feeling when you have what seems like a straightforward idea? “I’ll just connect Claude and ChatGPT so they can talk to each other. How hard could it be?”</p><p>That was me, two days ago.</p><p>I had built an MCP server that worked beautifully locally. The tools were solid. The architecture made sense. All I needed to do was add OAuth authentication so ChatGPT could securely access it over the internet.</p><p>Narrator voice: <em>*It was not simple.*</em></p><p>What followed was a masterclass in why production software is different from working software, why security is hard, and why the willingness to fail publicly is actually a superpower.</p><p>— -</p><p><strong>What I Was Trying to Build</strong></p><p>The vision was clean:</p><ul><li>An MCP server exposing tools for Claude-ChatGPT communication</li><li>Shared message queues (inboxes for each model)</li><li>File sharing capabilities</li><li>Conversation logging</li><li>Execute responsibly by securing my local data/machine with adequate authentication protocols</li></ul><p>I would protect myself using OAuth 2.0. I had done OAuth integrations before. I understood the flow. I had the documentation open. This should have taken maybe an hour. 🤣😬🫣</p><p><strong>**It took two full days.**</strong></p><p>And honestly? I’m glad it did.</p><p>— -</p><p><strong>The First Wall: “This Worked Yesterday”</strong></p><p>My initial MCP server worked perfectly when ChatGPT connected directly via localhost. Clean, fast, simple.</p><p>Then I added OAuth. And suddenly… nothing worked.</p><p><strong>**The symptom:**</strong> ChatGPT would hit the server, get a 401, and just… give up. No OAuth prompt. No error message. Just silent failure.</p><p><strong>**My first thought:**</strong> “I must have misconfigured something obvious.”</p><p><strong>**Reality:**</strong> I had misconfigured <em>*everything*</em>, but none of it was obvious.</p><p>This is where Claude Code became less of a tool and more of a thinking partner. Instead of me staring at error logs alone, we started methodically working through the stack.</p><p>— -</p><p><strong>Lesson 1: The OAuth Metadata Rabbit Hole</strong></p><p>Here’s what I learned the hard way: OAuth isn’t just about “add a login screen.”</p><p>It’s a negotiation between three parties:</p><ol><li>ChatGPT (the client)</li><li>Auth0 (the authorization server)</li><li>My MCP server (the resource server)</li></ol><p>And they all need to agree on a <em>*lot*</em> of things:</p><ul><li>Who’s allowed to request tokens?</li><li>What resources are they requesting access to?</li><li>Which scopes/permissions are valid?</li><li>Where should callbacks go?</li></ul><p>I had configured each piece independently, and they were all <em>*technically*</em> correct. But they weren’t correct <em>*together*</em>.</p><p><strong>**The specific issue:**</strong></p><p>My OAuth metadata endpoint was telling ChatGPT: “Request access to `https://my-ngrok-url.com`&quot;</p><p>But Auth0 was configured to issue tokens for: `https://claude-chatgpt-bridge`</p><p>ChatGPT would send the request with the ngrok URL as the resource identifier, and Auth0 would respond: “I don’t know what that is. Request denied.”</p><p><strong>**The fix:**</strong> One line of code. Change the resource field from the dynamic tunnel URL to the stable API identifier.</p><p><strong>**The lesson:**</strong> OAuth is a protocol built on <em>*exact agreement*</em>. Off by one character? Failure. Wrong URL scheme? Failure. Cached metadata? Failure.</p><p>— -</p><p><strong>Lesson 2: Authorization vs Authentication (They’re Not The Same)</strong></p><p>At one point, everything <em>*looked*</em> configured correctly. The OAuth flow would complete. ChatGPT would get a token. But then:</p><blockquote>Error: Client “abc123” is not authorized to access resource server “https://claude-chatgpt-bridge&quot;</blockquote><p>Wait, what?</p><p>Turns out, in Auth0, you have to explicitly authorize an application to access an API. And not just once — you have to authorize it for <em>*both*</em> user access AND client access. With the correct scopes. For each.</p><p>I had checked one box and thought I was done.</p><p><strong>**The realization:**</strong> Authentication proves <em>*who you are*</em>. Authorization proves <em>*what you’re allowed to do*</em>. OAuth handles both, but in different places, at different times, with different configurations.</p><p>This is the kind of thing that seems obvious in retrospect, but when you’re three hours into debugging and your sixth attempt just failed with the exact same error message… it’s not obvious at all.</p><p>— -</p><p><strong>Lesson 3: Token Validation is a Spectrum</strong></p><p>My first implementation used Auth0’s `/userinfo` endpoint to validate tokens. The docs said it worked. My tests passed locally. Ship it, right?</p><p>Except Auth0 was returning JWT access tokens, not the opaque tokens I expected. And `/userinfo` doesn’t validate access tokens — it validates ID tokens.</p><p>Every single request from ChatGPT was getting rejected with a cryptic 401 error.</p><p><strong>**The debugging process:**</strong></p><ol><li>Add logging to see the token</li><li>Decode the JWT manually to see the claims</li><li>Realize the token structure was different than expected</li><li>Research how to validate JWT access tokens properly</li><li>Implement claim-based validation (issuer, audience, expiration)</li><li>Test again</li><li>Success!</li></ol><p><strong>**The lesson:**</strong> There isn’t one “right” way to validate tokens. The approach depends on <em>*what kind of token you’re getting*</em>. And that depends on how you configured OAuth, which depends on which grant type is being used, which depends on…</p><p>It’s turtles all the way down.</p><p>— -</p><p><strong>Lesson 4: Free Tiers Have Hidden Costs</strong></p><p>I started with ngrok’s free tier. It worked great in testing. But when ChatGPT tried to use it in production, it hit ngrok’s interstitial warning page:</p><blockquote>&gt; “You are about to visit [URL]. This tunnel serves content from a ngrok user. Click ‘Visit Site’ to continue.”</blockquote><p>Great for security. Terrible for automated OAuth flows.</p><p>ChatGPT couldn’t click the button. The OAuth flow would fail. I’d get no error message, just… nothing.</p><p><strong>**The fix:**</strong> Switch to Cloudflare Tunnel, which doesn’t have an interstitial page.</p><p><strong>**The lesson:**</strong> Free tiers are amazing for learning and testing. But in production (even “production” on your laptop), you need to understand what compromises you’re making. Sometimes the compromise is an interstitial page. Sometimes it’s rate limits. Sometimes it’s URLs that change on every restart.</p><p>All of these are solvable problems — <em>*if you know they exist*</em>.</p><p>— -</p><p><strong>What Actually Worked (The Successes)</strong></p><p>Not everything was a struggle. Some things worked beautifully:</p><p><strong>**1. Claude Code as a debugging partner**</strong></p><p>Instead of me searching docs and Stack Overflow alone, Claude Code would:</p><ul><li>Read error logs and suggest hypotheses</li><li>Check configuration files against documentation</li><li>Spot inconsistencies I’d missed (like http vs https)</li><li>Propose fixes and explain why they should work</li></ul><p>This wasn’t magic — it was collaboration. But having a tireless partner who could read 50 lines of Auth0 documentation while I tested a theory? Invaluable.</p><p><strong>**2. Systematic debugging over random fixes**</strong></p><p>Every time we hit a wall, Claude Code pushed for:</p><ul><li>“Let’s check the logs first”</li><li>“Let’s verify each endpoint manually”</li><li>“Let’s confirm the token format before implementing validation”</li></ul><p>This slowed me down initially. But it meant when we found the issue, we <em>*understood*</em> it. No cargo-cult fixes. No “it works but I don’t know why.”</p><p><strong>**3. Documentation as we went**</strong></p><p>Claude Code insisted on documenting each fix. At the time, it felt like overhead. But when I hit the <em>*next*</em> issue, having clean notes on what we’d already tried was a lifesaver.</p><p>And now? I have a complete setup guide that someone else can follow without experiencing any of my pain. That’s a win.</p><p>— -</p><p><strong>The Moment It Worked</strong></p><p>After two days of OAuth errors, token validation failures, and mysterious 401s, I finally got this message in ChatGPT:</p><blockquote>✅ Bridge status received successfully.</blockquote><blockquote>Shared Files: 1</blockquote><blockquote>Claude’s Inbox: 1 unread message(s)</blockquote><blockquote>ChatGPT’s Inbox: 1 unread message(s)</blockquote><blockquote>Conversation Log: 1 entries</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V3IuEo9T2XaSJVWKcqBuWA.png" /></figure><p>That dopamine hit was <em>*real*</em>.</p><p>But more importantly: I understood <em>*why*</em> it worked. Every fix we had made. Every configuration change. I could explain each one.</p><p>That understanding is worth more than the working code.</p><p>— -</p><p><strong>What I’d Do Differently</strong></p><p><strong>**Start with security, not add it later**</strong></p><p>I built the MCP server first, then tried to add OAuth. This meant retrofitting security instead of designing for it. Next time: security architecture first, features second.</p><p><strong>**Use Auth0’s logs from day one**</strong></p><p>I spent hours guessing what was wrong before I realized Auth0 logs showed me <em>*exact*</em> error messages. “Client not authorized to access resource server” was right there, with timestamps and request details. I just had to look.</p><p><strong>**Test the OAuth flow independently**</strong></p><p>I kept testing through ChatGPT, which made debugging slow (reconnect, wait, check logs, repeat). I should have tested the OAuth flow with curl first. Get it working in isolation, <em>*then*</em> integrate with ChatGPT.</p><p><strong>**Accept that free tiers have constraints**</strong></p><p>I burned half a day fighting ngrok’s behavior before switching to Cloudflare Tunnel. I should have recognized the interstitial page as a dealbreaker earlier and moved on.</p><p>— -</p><p><strong>The Meta-Lesson: Embrace Public Failure</strong></p><p>Here’s the thing: I could have built this in private, figured it out through trial and error, and presented only the polished final result.</p><p>But that would rob you of the most valuable part: the <em>*process*</em>.</p><p>The mistakes I made? You’ll probably make similar ones. The dead ends I hit? You’ll hit them too. The “obvious in retrospect” realizations? They’re only obvious <em>*after*</em> you’ve lived through the confusion.</p><p>By sharing the messy middle — the wrong turns, the frustrations, the “wait, I need to authorize BOTH user access AND client access?” moments — I’m giving you a map of the terrain I already explored.</p><p>You don’t need to step in every pothole I did.</p><p><strong>**And here’s what makes this possible:**</strong> Working with Claude Code, everything was logged. Every hypothesis, every test, every fix. I’m not trying to remember what I did three days ago — I have a complete transcript.</p><p>That makes failure less scary. Because failure becomes <em>*documented learning*</em>.</p><p>— -</p><p><strong>Practical Takeaways</strong></p><p>If you’re building something with OAuth, MCP, or really any complex integration:</p><p><strong>**1. Trust the logs more than your assumptions**</strong></p><p>Your mental model is probably wrong. The logs know what actually happened.</p><p><strong>**2. Build incrementally and test constantly**</strong></p><p>Get OAuth working without MCP. Get MCP working without OAuth. Then combine them.</p><p><strong>**3. Document as you go, not at the end**</strong></p><p>Future you will thank present you. And your team/community will thank you even more.</p><p><strong>**4. Use AI tools as thinking partners, not magic boxes**</strong></p><p>Claude Code didn’t solve my problems for me. It helped me think through them systematically. That’s more valuable.</p><p><strong>**5. Share your failures, not just your successes**</strong></p><p>The community learns more from “here’s what I tried that didn’t work” than “here’s my finished product.”</p><p>— -</p><p><strong>## The Bigger Picture</strong></p><p>I set out to build a bridge between AI models. What I actually built was:</p><ul><li>A deeper understanding of OAuth 2.0</li><li>Appreciation for the subtlety of security protocols</li><li>A reusable pattern for OAuth-protected MCP servers</li><li>Complete documentation that others can learn from</li><li>A story about the value of persistent debugging</li></ul><p>The bridge works now. ChatGPT and Claude can exchange messages, share files, collaborate on tasks.</p><p>But the real value isn’t the code. It’s the <em>*knowledge*</em> of how to debug complex integrations, how to read error messages as puzzles rather than roadblocks, and how to build incrementally even when you’re tempted to skip steps.</p><p>— -</p><p><strong>Closing: The Invitation</strong></p><p>I’m sharing all of this — the code, the mistakes, the lessons — because I believe we learn faster together.</p><p>My GitHub repo has three documents:</p><p><strong>SETUP-GUIDE.md</strong>: <a href="https://github.com/bkelson/-claude-chatgpt-bridge/blob/main/SETUP-GUIDE.md">https://github.com/bkelson/-claude-chatgpt-bridge/blob/main/SETUP-GUIDE.md</a></p><p>2. <strong>RESUME-SESSION.md</strong>: <a href="https://github.com/bkelson/-claude-chatgpt-bridge/blob/main/RESUME-SESSION.md">https://github.com/bkelson/-claude-chatgpt-bridge/blob/main/RESUME-SESSION.md</a></p><p>3. <strong>TALK-NOTES.md</strong> (this talk): <a href="https://github.com/bkelson/-claude-chatgpt-bridge/blob/main/TALK-NOTES.md">https://github.com/bkelson/-claude-chatgpt-bridge/blob/main/TALK-NOTES.md</a></p><p>Use whichever serves you best.</p><p>And if you build something similar and hit walls I didn’t document? Please, share those too. The next person will thank you.</p><p>Because that’s how we all get better: not by hiding our struggles, but by turning them into stepping stones for others.</p><p>— -</p><p><strong>Q&amp;A</strong></p><p><strong>**Q: “Would you recommend Auth0 for this use case?”**</strong></p><p>A: Yes, but understand it’s enterprise-grade OAuth with all the complexity that implies. For learning, it’s great because you see the full OAuth protocol. For production, it’s robust but requires careful configuration.</p><p><strong>**Q: “Why not use a simpler auth approach?”**</strong></p><p>A: I wanted to learn OAuth properly, and ChatGPT specifically requires OAuth for MCP servers. This was driven by the platform requirements, but it ended up being a great learning experience.</p><p><strong>**Q: “How long did this actually take?”**</strong></p><p>A: About 12–15 hours total over two days. If I did it again with current knowledge? Maybe 1–2 hours. That’s the value of documented learning.</p><p><strong>**Q: “What would you build next?”**</strong></p><p>A: Role-based tools (Claude as “red team”, ChatGPT as “synthesizer”), structured debate protocols, confidence scoring. The bridge is built — now comes the interesting orchestration layer.</p><p><strong>**Q: “Is this production-ready?”**</strong></p><p>A: For personal/learning use, absolutely. For business-critical use, you’d want: permanent URLs, JWT signature verification, rate limiting, monitoring, and proper secrets management. But it’s a solid foundation.</p><p>— -</p><p><strong>**Final thought:**</strong></p><p>Building in public means admitting you don’t know everything. That’s uncomfortable.</p><p>But it also means you learn faster, help others avoid your mistakes, and build credibility through authenticity rather than perfection.</p><p>I’ll take that trade every time.</p><p>— -</p><p><em>🧠 All ideas presented here begin in my own mind. Their expression and refinement, however, are the product of an unapologetic collaboration with synthetic intelligence. 🤖</em></p><p>✉️ Discussions, critiques, ideas and other such banter always welcome: <a href="/service/mailto:brian@kelson.ai">brian@kelson.ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0cfca91753d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>