<?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 Suhail Roushan on Medium]]></title>
        <description><![CDATA[Stories by Suhail Roushan on Medium]]></description>
        <link>https://medium.com/@suhailroushan?source=rss-4d10b081bc95------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*7WxJ7F-x81fDu_rsfXtVDA.jpeg</url>
            <title>Stories by Suhail Roushan on Medium</title>
            <link>https://medium.com/@suhailroushan?source=rss-4d10b081bc95------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 20:59:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@suhailroushan/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[My Medium Draft Had No Images. So I Wrote Them in HTML.]]></title>
            <link>https://suhailroushan.medium.com/my-medium-draft-had-no-images-so-i-wrote-them-in-html-4859a4712a81?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/4859a4712a81</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[blogging]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[playwrights]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 13:30:02 GMT</pubDate>
            <atom:updated>2026-11-16T13:30:02.758Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Three images, one Playwright script, and a deprecated API that still accepts uploads. Here’s the pipeline, and the three things that broke along the way.</em></p><figure><img alt="An HTML file in a code editor on the left turns into a finished blog hero image on the right." src="https://cdn-images-1.medium.com/proxy/1*0scKv2T7c4HaLe_VANOmng.png" /></figure><h4>The draft looked finished until I opened it</h4><p>I’d just pushed a draft to Medium for an article about <a href="https://shipx.live/">ShipX</a>, the product launch platform I’m building. The text was done. The title was done. The tags were set.</p><p>Then I looked at it the way a reader would: a long column of text, no hero image, no screenshot of the product the article was about, nothing to break up the sections.</p><p>Medium uses the first image in your story as the preview card. With no images, my article would show up in feeds as a grey title and nothing else.</p><p>I needed three things:</p><ol><li><strong>A preview image</strong> that would work as the story’s hero and its card in the feed.</li><li><strong>A product image</strong> showing what ShipX actually looks like.</li><li><strong>A diagram</strong> for the section about why launches go quiet.</li></ol><p>I didn’t open a design tool. I worked in Claude Code, in the same terminal session where I’d written and uploaded the draft, and every image ended up as an HTML file rendered by a headless browser.</p><p>Here’s how that pipeline works, and where it broke.</p><h4>Why HTML is a surprisingly good image format</h4><h4>Simple explanation</h4><p>A blog image is mostly layout: a background, some text, maybe a screenshot in a frame. Web pages are layout too. If you can build it as a web page, a browser can take a picture of it.</p><p>So instead of drawing the image, you write a small page at exactly the size you want, and let a headless browser screenshot it.</p><h4>Technical explanation</h4><p>The setup is three pieces:</p><ul><li><strong>One HTML file per image</strong>, sized with a fixed width and height on body (1400×788 for the hero, 1600×1000 for the product shot, 1600×700 for the diagram).</li><li><strong>A shared stylesheet</strong> for repeated pieces, like a fake browser window frame.</li><li><strong>Playwright</strong> opening each file with device_scale_factor=2, so a 1400×788 layout comes out as a sharp 2800×1576 PNG.</li></ul><p>The payoff is that images become editable text. Change a headline, re-run one command, get a new PNG. Fonts, spacing, and colors stay consistent because they come from the same CSS.</p><figure><img alt="Five steps: screenshot with Playwright, extract logo and color with Pillow, template in HTML, render with Playwright, upload with the Medium API." src="https://cdn-images-1.medium.com/proxy/1*O1ScXRM1Ycu7Mnw--g2ZHw.png" /></figure><p>| | Design tool | HTML + headless browser | | — -| — -| — -| | Change the headline | Open the file, edit, export | Edit one line, re-render | | Keep three images consistent | Manual | Shared CSS | | Version control | Binary files | Plain text diffs | | Real product screenshot inside | Paste and position | &lt;img&gt; tag |</p><h4>Step 1: Take a real screenshot of the product</h4><p>The product image had to be real. I didn’t want a mockup of ShipX; I wanted ShipX.</p><pre>from playwright.sync_api import sync_playwright<br><br>with sync_playwright() as p:<br>    browser = p.chromium.launch()<br>    page = browser.new_page(<br>        viewport={&quot;width&quot;: 1440, &quot;height&quot;: 900},<br>        device_scale_factor=2,<br>    )<br>    page.goto(&quot;https://shipx.live&quot;, wait_until=&quot;networkidle&quot;)<br>    page.screenshot(path=&quot;raw-home.png&quot;)<br>    browser.close()</pre><p>That produced a 2880×1800 screenshot of the homepage, with that week’s launches ranked by upvotes.</p><p>It also produced the first problem.</p><h4>What broke: the cookie banner</h4><p>The screenshot had a cookie consent box sitting on top of the product list: “Help us improve ShipX,” with two buttons.</p><p>Of course it did. A headless browser is a brand-new visitor every time, so it sees every first-visit prompt a real visitor would.</p><p>The fix was to click it before taking the picture:</p><pre>page.get_by_role(&quot;button&quot;, name=&quot;Essential only&quot;).click()<br>page.wait_for_timeout(800)<br>page.screenshot(path=&quot;raw-home.png&quot;)</pre><p>Using get_by_role with the button&#39;s visible label is more robust than a CSS selector here, because it targets what the user sees rather than how the page happens to be built.</p><p><strong>Lesson:</strong> look at your first screenshot before building anything on top of it. Banners, chat widgets, and sign-in prompts all show up for a fresh browser.</p><figure><img alt="The first screenshot with a cookie banner over the product list, beside the clean screenshot after clicking &#39;Essential only&#39;." src="https://cdn-images-1.medium.com/proxy/1*lyqsPLeUUQefdxVlI1KZAw.png" /></figure><h4>Step 2: Get the logo and brand color without asking for assets</h4><p>I wanted the hero to use ShipX’s real logo and accent color. The obvious approach was to screenshot the logo element directly.</p><h4>What broke: the logo “wasn’t visible”</h4><p>Playwright’s element screenshot waited 30 seconds and then timed out with:</p><pre>element is not visible</pre><p>The selector matched an element that existed in the page, but not the one being shown. Rather than hunt for the right selector, I already had a perfectly good 2× screenshot with the logo in it. So I cropped it from there.</p><p>The trick is to find the logo’s bounding box automatically by looking for dark pixels in the header area:</p><pre>from PIL import Image<br><br>im = Image.open(&quot;raw-home.png&quot;).convert(&quot;RGB&quot;)<br>dark = [<br>    (x, y)<br>    for x in range(640, 740)<br>    for y in range(30, 120)<br>    if sum(im.getpixel((x, y))) &lt; 150<br>]<br>xs, ys = zip(*dark)<br>im.crop((min(xs), min(ys), max(xs) + 1, max(ys) + 1)).save(&quot;logo-mark.png&quot;)</pre><p>For the accent color, I counted the bright green pixels inside that same box and took the most common one: rgb(208, 250, 72), or #D0FA48. That became the highlight color across all three images.</p><p><strong>Lesson:</strong> a high-resolution screenshot is an asset library. Logos, colors, and UI fragments are all in there.</p><figure><img alt="The ShipX header with a dashed box around the logo, leading to the cropped logo file and the extracted color #D0FA48." src="https://cdn-images-1.medium.com/proxy/1*dX-n63l_L569MJPzYYs7PQ.png" /></figure><h4>Step 3: Build each image as a page</h4><p>With a screenshot, a logo, and a color, the images themselves were small HTML files.</p><p>The shared piece was a fake browser window, so the product screenshot looks like a site rather than a floating rectangle:</p><pre>.browser {<br>  background: #fff;<br>  border-radius: 18px;<br>  overflow: hidden;<br>  box-shadow: 0 40px 80px -20px rgba(0,0,0,.45);<br>}<br>.bar {<br>  height: 44px;<br>  background: #f3f3f1;<br>  display: flex;<br>  align-items: center;<br>  gap: 8px;<br>  padding: 0 18px;<br>}</pre><p>Three colored dots and a little shipx.live address bar sit inside .bar. Then the screenshot is just an &lt;img&gt; underneath.</p><p><strong>The hero</strong> is a dark 1400×788 card: the logo, a two-line headline with the second half in #D0FA48, a one-sentence subtitle, and the browser frame on the right, tilted with a CSS 3D transform:</p><pre>.right {<br>  transform: perspective(1800px) rotateY(-14deg) rotateX(4deg);<br>  transform-origin: left center;<br>}</pre><p><strong>The product image</strong> is the same frame, flat and centered on a soft background.</p><p><strong>The diagram</strong> is four cards in a row (See it → Understand it → Try it → Return), each with the question it answers and what to do if the answer is no. It’s a flexbox layout, nothing more.</p><p>Rendering all three is one loop:</p><pre>jobs = [(&quot;hero&quot;, 1400, 788), (&quot;product&quot;, 1600, 1000), (&quot;diagram&quot;, 1600, 700)]<br><br>for name, w, h in jobs:<br>    page = browser.new_page(viewport={&quot;width&quot;: w, &quot;height&quot;: h}, device_scale_factor=2)<br>    page.goto(Path(f&quot;src/{name}.html&quot;).resolve().as_uri())<br>    page.screenshot(path=f&quot;{name}.png&quot;)<br>    page.close()</pre><p>When the first version of the diagram had empty space at the bottom, the fix was changing 760px to 700px and re-running the loop. That&#39;s the whole editing workflow.</p><figure><img alt="The hero image&#39;s HTML source and render loop beside the rendered hero image." src="https://cdn-images-1.medium.com/proxy/1*lPaZStD6McvrQcUtXiWGbw.png" /></figure><h4>Step 4: Get the images into Medium</h4><p>Medium’s official API is deprecated, and it no longer issues new integration tokens. But I had an existing token, and two endpoints still did what I needed.</p><p>Uploading an image is a multipart POST:</p><pre>import requests<br><br>requests.post(<br>    &quot;https://api.medium.com/v1/images&quot;,<br>    headers={&quot;Authorization&quot;: f&quot;Bearer {token}&quot;},<br>    files={&quot;image&quot;: (&quot;hero.png&quot;, open(&quot;hero.png&quot;, &quot;rb&quot;), &quot;image/png&quot;)},<br>)</pre><p>Each upload returned a URL on cdn-images-1.medium.com. Those URLs went straight into the article body as HTML, which lets you include captions:</p><pre>&lt;figure&gt;<br>  &lt;img src=&quot;https://cdn-images-1.medium.com/...&quot; alt=&quot;...&quot;&gt;<br>  &lt;figcaption&gt;A quiet launch can hide several different problems.&lt;/figcaption&gt;<br>&lt;/figure&gt;</pre><p>I converted the article’s Markdown to HTML, inserted each &lt;figure&gt; after the right paragraph, and created the post with &quot;contentFormat&quot;: &quot;html&quot; and &quot;publishStatus&quot;: &quot;draft&quot;.</p><h4>What broke: you can’t edit a post through the API</h4><p>My text-only draft already existed. I wanted to add images to it.</p><p>The API can’t do that. It has endpoints to create posts, but none to edit or delete them. So the only way to get images in was to <strong>create a second draft</strong> with the images included, then delete the first one by hand in Medium’s editor.</p><p>That’s a small annoyance, but it changes how you should build this pipeline: <strong>finish the images before the first upload</strong>, not after.</p><p><strong>Lesson:</strong> with Medium’s API, treat each post creation as final. Draft locally, then create once.</p><figure><img alt="Two columns: what Medium&#39;s old API can do (profile, image upload, create posts, list publications) and can&#39;t do (edit, delete, read stats, search)." src="https://cdn-images-1.medium.com/proxy/1*knj0UWSLcB3s_g-BKxJ1-Q.png" /></figure><h4>What I’d tell someone trying this</h4><p><strong>Start from a real screenshot.</strong> It gives you the product image, the logo, and the brand color in one capture.</p><p><strong>Keep the HTML sources next to the article.</strong> Mine live in images/shipx/src/. When the headline changes, the hero changes with one command.</p><p><strong>Render at 2×.</strong> Medium displays images large, and a 1× render looks soft on most modern screens.</p><p><strong>Check every first render with your own eyes.</strong> The cookie banner and the empty space under the diagram were obvious the moment I looked, and invisible until I did.</p><p><strong>Know what the API won’t do.</strong> Upload images and create drafts, yes. Edit, delete, or read your stats, no.</p><p>This approach has limits. It’s great for cards, frames, diagrams, and screenshots. It won’t paint you an illustration, and if the product’s design changes, you need a new screenshot, not just a re-render.</p><figure><img alt="The three finished images: a dark hero card, a framed ShipX homepage, and a four-step launch diagram." src="https://cdn-images-1.medium.com/proxy/1*McszOsfaSxHR-xFT9ZMgsw.png" /></figure><h4>The real problem was never the images</h4><p>The images were never really the hard part. The hard part was the assumption that images belong in a different tool, made by a different process, after the writing is done.</p><p>Once they were HTML files sitting next to the article, they behaved like the rest of the post: editable, versioned, and one command away from updated.</p><p>If your drafts keep stalling at “I’ll add images later,” try writing the first one as a web page. You might not go back.</p><p>If you try this, I’d like to see what you render. And if you’ve launched something recently, you can share it on <a href="https://shipx.live/">ShipX</a>.</p><p><em>Drafted with AI assistance (Claude), working from my own code, configs and notes.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4859a4712a81" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Let Claude Post to ShipX’s Instagram. Then a Reel Went Out With the Wrong Date.]]></title>
            <link>https://suhailroushan.medium.com/i-let-claude-post-to-shipxs-instagram-then-a-reel-went-out-with-the-wrong-date-928f6fa27f9c?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/928f6fa27f9c</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[instagram]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 13:30:03 GMT</pubDate>
            <atom:updated>2026-11-11T13:30:03.889Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A Reel went out saying “Sunday, 19 Oct” when Sunday was the 18th. Instagram’s API can’t delete posts, so here are the rules my agent follows now.</em></p><figure><img alt="Headline card: &quot;I Let Claude Post to ShipX&#39;s Instagram. Then a Reel Went Out With the Wrong Date.&quot;" src="https://cdn-images-1.medium.com/proxy/1*06iclJeeU24jW92atN0YLA.png" /></figure><p>The Reel looked great.</p><p>Dark background. Lime accent. Clean type. A launch-week reminder for <a href="https://shipx.live/">ShipX</a>, the product launch platform I’m building.</p><p>Burned into the video, in big letters, was the date: <strong>Sunday, 19 Oct.</strong></p><p>Sunday was the 18th.</p><p>It was already live on Instagram. And here’s the thing I learned that day: <strong>you can’t delete an Instagram post through the API.</strong> You can’t edit the caption either. You open the app on your phone and delete it by hand, like everyone else.</p><p>My AI agent could publish to Instagram in seconds. It couldn’t take anything back.</p><p>That mismatch changed how I set up the whole thing.</p><h4>What I built</h4><p>I run ShipX’s social accounts (@shipx_live) from Claude Code. The setup has three parts:</p><ul><li><strong>An Instagram MCP server</strong>, for reading the account: media, comments, DMs and insights, plus replying to comments.</li><li><strong>ig-publish</strong>, a small CLI that posts Reels, stories and images from local files.</li><li><strong>threads-publish</strong>, the same idea for Threads.</li></ul><p>Tokens live in the macOS Keychain. The scripts read them at runtime and never print them.</p><p>Claude decides <em>what</em> to post using a skill file: ShipX’s voice, its colours, its hashtags, and a short list of rules. The CLIs do the posting.</p><p>On paper, that’s a great setup. In practice, I learned three lessons the hard way.</p><h4>Lesson 1: An agent doesn’t know what day it is</h4><p>A weekday is exactly the kind of thing that gets assumed instead of checked. And when an agent is the one writing dates into media, that matters.</p><p>A language model is very good at sounding sure about a date and surprisingly bad at being right about one. And a date burned into a video isn’t like a typo in a draft. You can’t fix it after publishing.</p><p>So the first rule in the skill is now:</p><blockquote><strong>Compute every date and weekday with a tool.</strong> Never assume today’s weekday.</blockquote><p>Something like python3 -c &quot;import datetime; ...&quot; before writing any date into media. It&#39;s a single command, and it guards against the most embarrassing mistake the account has made.</p><h4>Lesson 2: Draft first, publish separately</h4><p>Instagram’s API publishes in two steps. First, you create a <em>media container</em>: you upload the video and Meta processes it. Then you publish that container.</p><p>I turned that into a rule. ig-publish has a --draft flag that does step one and stops:</p><pre>ig-publish reel reel.mp4 --caption-file caption.txt --draft<br># → prints a container ID, valid for 24 hours<br><br>ig-publish publish &lt;container-id&gt;</pre><p>The draft step checks that Meta actually accepts the file, so format problems show up before anything goes public. Then Claude shows me the caption and a preview frame.</p><p><strong>Nothing gets published until I say go.</strong> The only exception is if I’ve already approved that exact content.</p><p>It sounds slow. It isn’t. It’s one message: “looks good, post it.”</p><h4>Lesson 3: Delete one app and every token dies</h4><p>This one surprised me most.</p><p>Instagram’s API tokens are issued by a Meta “app” you create in Meta’s developer dashboard. I had two of them. On 18 October, one of them was deleted.</p><p>Every token it had issued died with it. The next publish failed with:</p><pre>Error validating application. Application has been deleted.</pre><p>The fix was to generate a new token from the remaining app and store it again. The lesson went into the skill file, so the next session doesn’t have to rediscover it:</p><blockquote>If an app behind a token is deleted, every token from it dies at once.</blockquote><h4>Smaller things that will trip you up</h4><ul><li><strong>Instagram images must be JPEGs at a public HTTPS URL.</strong> PNGs are rejected, and you can’t upload a local image directly. I added a host-image script that uploads to Cloudinary as a JPEG and prints the URL.</li><li><strong>Videos are different.</strong> Reels and stories can go up with Meta’s resumable upload straight from a local file, so they don’t need hosting.</li><li><strong>Links in Instagram captions aren’t clickable.</strong> And “link in bio” is a lie if your profile has no link. So the first command in every session is ig-publish whoami, which shows the profile link, token expiry and publish quota.</li><li><strong>Threads is more forgiving.</strong> threads-publish can delete a post. Instagram&#39;s API can&#39;t.</li></ul><h4>The rule underneath all the rules</h4><p>When I started, I thought the hard part of automating social media would be the publishing.</p><p>It wasn’t. Publishing is a few API calls.</p><p>The hard part is that <strong>publishing is public and permanent</strong>, while agents are fast and confident. The more an action can’t be undone, the more checks I want before it.</p><p>So my setup now does the opposite of “full automation”:</p><ul><li>The agent drafts, checks dates with a tool, uploads, and validates.</li><li>I approve.</li><li>Then it publishes.</li></ul><p>It’s still much faster than doing it by hand. It just isn’t faster than my judgement.</p><p><em>I’m building ShipX in public, with the mistakes included. I’m also writing up the WhatsApp channel I use to send Claude tasks from my phone, and how it asks permission before touching anything. Follow if you want to see it.</em></p><p><em>Drafted with AI assistance (Claude), working from my own code, configs and notes.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=928f6fa27f9c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Counted the Commits on Every Product I Built This Year. None of Them Measured the Hard Part.]]></title>
            <link>https://suhailroushan.medium.com/i-counted-the-commits-on-every-product-i-built-this-year-none-of-them-measured-the-hard-part-934793fbc59a?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/934793fbc59a</guid>
            <category><![CDATA[side-project]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[indiehackers]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 13:30:03 GMT</pubDate>
            <atom:updated>2026-11-07T13:30:03.044Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Eight products, 301 commits, one 12-day sprint with 122 of them. The only real usage numbers I could find didn’t come from any of my websites.</em></p><figure><img alt="Headline card: &quot;I Counted the Commits on Every Product I Built This Year. None of Them Measured the Hard Part.&quot;" src="https://cdn-images-1.medium.com/proxy/1*-_v5hna5F_rP28UkszCqaA.png" /></figure><p>I wanted an honest answer to a simple question:</p><p><strong>Where did my time actually go this year?</strong></p><p>So I went through every product folder on my laptop and counted. Git commits, first commit, last commit.</p><p>It was the easiest data I’ve ever collected. Every product had a perfect record of how hard I’d worked on it.</p><p>Then I tried to add a second column: <em>how many people used it.</em></p><p>For most of them, I couldn’t fill it in.</p><h4>The list</h4><p>Here’s everything I started in 2026, from git history:</p><p>| Product | What it is | Commits | Active | | — -| — -| — -| — -| | biryani.live | Crowd-sourced tracker for Hyderabad’s best biryani spots | 53 | Feb → Sep | | gitai | AI-powered Git CLI: commit messages, PR descriptions, security scans | 18 | 4 days in March | | getqrmenu | QR menu product with a landing page and pricing | 3 | 1 week in March/April | | ShipCheck | CLI that audits a website for performance, SEO, accessibility and security | 20 | 1 day in September | | problems.live | Public directory of real problems worth solving | 122 | 12 days in September | | Vibe Check for X | Chrome extension that scores your X drafts before you post | 1 | 1 day in September | | TimeAsAService | Marketplace where people auction an hour of their time, starting at $1 | 2 | 1 day in September | | ShipX | Weekly product launch platform for indie makers | 82 | October → now |</p><p>Look at September.</p><p>In one month I started five products. One of them, problems.live, got 122 commits in 12 days. That’s roughly ten commits a day, every day, for almost two weeks.</p><p>I was <em>very</em> productive.</p><h4>The second column</h4><p>Now here’s the column I wanted to add next to it: users, signups, visits, anything.</p><p>For the websites, the column stayed empty.</p><p>Not because I knew the answer was zero. Because whatever the answer was, nothing in front of me could tell me. The git log had been recording every evening I spent on them. Nothing had been recording whether anyone showed up.</p><h4>The only numbers I could find</h4><p>Two of my projects did have real usage numbers. Neither was a website.</p><ul><li><strong>gitai</strong>, my AI Git CLI: <strong>1,011 npm downloads</strong> in the past year.</li><li><strong>killport-npm</strong>, a tiny CLI I published back in 2023 that kills whatever process is holding a port: <strong>610 npm downloads</strong> in the past year.</li></ul><p>(npm download counts include bots, CI servers and mirrors, so they’re an upper bound on real people, not a user count.)</p><p>Here’s what got me about killport-npm. I made it three years ago. It’s been downloaded hundreds of times this year anyway.</p><h4>What the npm packages had that my websites didn’t</h4><p>They weren’t better products.</p><p>They were <strong>inside a place where people were already looking.</strong></p><p>When a developer types “kill port” into npm or a search engine, a small CLI with a clear name can show up. The registry is a distribution channel. It brings people to you whether you’re paying attention or not.</p><p>My websites didn’t have that. biryani.live, problems.live, TimeAsAService: each was a new URL in a world with no reason to visit it. The commits made them <em>exist</em>. Nothing made them <em>findable</em>.</p><p>I spent my effort on the part I could measure, commit by commit.</p><p>The hard part had no commit history at all.</p><h4>Why builders over-invest in building</h4><p>I don’t think I’m unusual here. Building has a lot going for it:</p><ul><li><strong>It feels like progress.</strong> Every commit is a small win.</li><li><strong>It’s measurable.</strong> The git log never lies to you.</li><li><strong>It’s safe.</strong> Nobody can reject a commit.</li></ul><p>Distribution is the opposite. It’s slow, hard to measure, and every attempt is a chance for people to ignore you.</p><p>So when you have a free evening, it’s always easier to open the editor than to post about what you built.</p><h4>What I’m doing differently with ShipX</h4><p>ShipX is the first product where I’ve treated distribution as the work, not the afterthought. Partly because the product <em>is</em> about distribution: it gives makers with small audiences a fair place to launch.</p><p>But I also changed how I spend my own time:</p><ol><li><strong>Distribution gets scheduled time, like coding does.</strong> These Medium posts are part of that.</li><li><strong>I put the product where people already look.</strong> Search, launch platforms, communities, not just my own feed.</li><li><strong>Every launch comes with a plan for after launch day.</strong></li><li><strong>I track one usage number from day one</strong>, before the first commit feels like an achievement.</li></ol><h4>The question I ask now</h4><p>Before I start anything new, I ask:</p><p><strong>“Where will the first 100 people come from?”</strong></p><p>If I can’t answer that, I don’t need more commits. I need a different plan.</p><p>Your git history will always show how hard you worked.</p><p>It can’t show whether anyone noticed.</p><p><em>This is part of </em><strong><em>Building ShipX</em></strong><em>, where I document the numbers, mistakes and experiments of growing a product in public. Previous: </em><a href="https://suhailroushan.medium.com/i-built-a-product-then-realized-getting-users-was-the-hard-part-dfd60001dcd1"><em>I Built a Product. Then Realized Getting Users Was the Hard Part.</em></a><em> Follow if you want to see whether doing distribution first actually works.</em></p><p><em>Drafted with AI assistance (Claude), working from my own code, configs and notes.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=934793fbc59a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My CLAUDE.md Is Only 20 Lines. The Rest of My Claude Code Setup Isn’t.]]></title>
            <link>https://suhailroushan.medium.com/my-claude-md-is-only-20-lines-the-rest-of-my-claude-code-setup-isnt-5ab09981d42a?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ab09981d42a</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 13:30:03 GMT</pubDate>
            <atom:updated>2026-11-04T13:30:03.254Z</atom:updated>
            <content:encoded><![CDATA[<p><em>10 hook events, 33 skills, 11 small scripts, and a Keychain full of tokens Claude can use but never see. Here’s the whole setup and what each piece is for.</em></p><figure><img alt="Headline card: &quot;My CLAUDE.md Is Only 20 Lines. The Rest of My Claude Code Setup Isn&#39;t.&quot;" src="https://cdn-images-1.medium.com/proxy/1*J-umLvzKPCxi9FKVqfcpDA.png" /></figure><p>A story about a single CLAUDE.md file went viral on Medium recently. Its subtitle mentioned 91,000 GitHub stars.</p><p>It made me look at mine.</p><p>It’s 20 lines long.</p><p>It doesn’t explain my coding style. It doesn’t list my favourite libraries. It says three things: how to reach me, where my GitHub token is, and how to post to ShipX’s Instagram.</p><p>That surprised me. I use Claude Code for almost everything I build, including <a href="https://shipx.live/">ShipX</a>, my product launch platform. I expected my instructions file to be huge.</p><p>It turns out the instructions aren’t where the setup lives. <strong>The setup lives in what Claude can <em>do</em>.</strong></p><p>Here’s the whole thing.</p><h4>Layer 1: CLAUDE.md, the short version</h4><p>My global CLAUDE.md has three sections:</p><ol><li><strong>Contacting me.</strong> A hook emails me Claude’s final message every time it finishes, so the final message must make sense on its own. For something urgent mid-task, there’s a script to email me directly.</li><li><strong>GitHub token.</strong> Where it is, which account it’s for, what scopes it has, when it expires, and the command to read it.</li><li><strong>Instagram and Threads.</strong> Which skill to use, where the tokens are, and one hard rule: get my go-ahead before publishing anything.</li></ol><p>Each section ends with the same kind of line:</p><blockquote>Never print it, copy it into files, or put it in commits.</blockquote><p>That’s the pattern. CLAUDE.md doesn’t hold knowledge. It holds <strong>pointers and boundaries</strong>: where things are, and what not to do with them.</p><h4>Layer 2: Secrets live in the Keychain, not in files</h4><p>Every secret Claude uses is in the macOS Keychain: the Gmail app password, the GitHub token, the Instagram and Threads tokens, Cloudinary keys.</p><p>Scripts read them at runtime:</p><pre>security find-generic-password -s github-gist-token -w</pre><p>There’s no .env file in any of these tools. There&#39;s nothing to commit by accident, and nothing in a file that some other tool might upload.</p><p>Claude <em>uses</em> the tokens. It never needs to see them.</p><h4>Layer 3: Small scripts in ~/.claude/bin</h4><p>There are 11 scripts in that folder. Each one does a single job. The main ones:</p><ul><li>email-me: email me a subject and body</li><li>email-on-stop: the hook that emails me Claude&#39;s final message</li><li>ig-publish / threads-publish: post to ShipX&#39;s Instagram and Threads</li><li>host-image / host-video: upload media to get a public URL</li><li>instagram-mcp, medium-mcp: launchers for MCP servers</li><li>instagram-renew-token, instagram-mcp-store-token: token maintenance</li></ul><p>Why scripts and not MCP servers for everything? Because a script is something I can also run myself, test from the terminal, and read in one sitting. If Claude can call it with Bash, it doesn’t need a protocol.</p><p>I use MCP servers when the tool needs to hold a connection or offer many related actions, like reading Instagram comments, DMs and insights.</p><h4>Layer 4: Hooks on every event</h4><p>Claude Code can run a command when something happens. I have hooks on 10 events: session start and end, every prompt, before and after every tool call, permission requests, notifications, stops, failed stops and subagent stops.</p><p>Most of them call one small program that updates my terminal’s status, so I can see at a glance what each session is doing.</p><p>One of them, on Stop, emails me. (I wrote about that hook separately: <a href="https://medium.com/@suhailroushan/i-stopped-watching-my-terminal-claude-code-emails-me-when-its-done-21370cbb552d">I Stopped Watching My Terminal. Claude Code Emails Me When It&#39;s Done.</a>)</p><p>Hooks are the part of the setup Claude doesn’t control. The harness runs them, every time. <strong>If something must always happen, it belongs in a hook, not in an instruction Claude might forget.</strong></p><h4>Layer 5: Skills for jobs with rules</h4><p>I have 33 entries in my skills folder. Twelve are Remotion video skills, and a group of them are HyperFrames video skills. Others cover Supermemory, a web agent, and my Instagram posting workflow.</p><p>A skill is a file of instructions Claude loads only when the task needs it. That’s the difference from CLAUDE.md: CLAUDE.md is read every session, and a skill only when it’s relevant.</p><p>The Instagram skill is the best example of why this matters. It holds:</p><ul><li>ShipX’s voice (“Ship it. Launch it. Get discovered.”, no “revolutionary”, always “upvote”)</li><li>Brand colours and fonts</li><li>The approval rule</li><li>Lessons from real mistakes: compute every date with a tool, and what happens when a Meta app gets deleted</li></ul><p>That’s knowledge I’d never want loaded while Claude is fixing a database query. As a skill, it shows up exactly when it’s needed.</p><h4>Layer 6: Channels, so I’m not tied to the terminal</h4><p>The last layer is reach. Claude can:</p><ul><li><strong>Email me</strong> when it’s done (hook).</li><li><strong>Receive tasks over WhatsApp</strong> through an open-source Claude Code channel plugin, and ask me to approve tool use from my phone.</li><li><strong>Post</strong> to Instagram, Threads and Medium drafts, always behind an approval step.</li></ul><h4>Two accounts, same structure</h4><p>I run two configurations side by side, ~/.claude and ~/.claude-work, each with its own CLAUDE.md, skills and scripts. The email hook labels which account a message came from, so I always know which session to reopen.</p><h4>What I’d copy if I were starting today</h4><ol><li><strong>Keep CLAUDE.md short.</strong> Pointers and boundaries, not an essay.</li><li><strong>Put secrets in a keychain</strong> and give Claude the command to read them, never the value.</li><li><strong>Write tiny scripts</strong> before reaching for MCP.</li><li><strong>Use hooks for anything that must always happen.</strong></li><li><strong>Turn repeated workflows into skills</strong>, and write your mistakes into them.</li><li><strong>Give the agent a way to reach you</strong>, so you can stop watching it.</li></ol><h4>The lesson I didn’t expect</h4><p>I thought a good Claude Code setup meant better instructions.</p><p>Mine got better when I stopped <em>telling</em> Claude things and started <em>giving</em> it things: a script to call, a hook that always fires, a skill that loads at the right moment, a rule that keeps it from publishing without me.</p><p>The viral CLAUDE.md was about what to say to your AI.</p><p>My 20 lines are mostly about what it’s allowed to touch.</p><p><em>I’m building ShipX almost entirely with this setup, and I’m writing up each piece as I go. Follow along if you want the details of each piece.</em></p><p><em>Drafted with AI assistance (Claude), working from my own code, configs and notes.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ab09981d42a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Build With AI Every Day. I Still Think It Needs to Be Stopped.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://suhailroushan.medium.com/i-build-with-ai-every-day-i-still-think-it-needs-to-be-stopped-05290a1cfa68?source=rss-4d10b081bc95------2"><img src="https://cdn-images-1.medium.com/max/600/1*kNIr5JofziIvhN-sJP4TXg.png" width="600"></a></p><p class="medium-feed-link"><a href="https://suhailroushan.medium.com/i-build-with-ai-every-day-i-still-think-it-needs-to-be-stopped-05290a1cfa68?source=rss-4d10b081bc95------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://suhailroushan.medium.com/i-build-with-ai-every-day-i-still-think-it-needs-to-be-stopped-05290a1cfa68?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/05290a1cfa68</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[ai-safety]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 09:52:12 GMT</pubDate>
            <atom:updated>2026-11-04T09:52:12.793Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Stopped Watching My Terminal. Claude Code Emails Me When It’s Done.]]></title>
            <link>https://suhailroushan.medium.com/i-stopped-watching-my-terminal-claude-code-emails-me-when-its-done-21370cbb552d?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/21370cbb552d</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Sat, 31 Oct 2026 13:30:02 GMT</pubDate>
            <atom:updated>2026-10-31T13:30:02.927Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A Stop hook, about 110 lines of Python, and one line in CLAUDE.md that changed how Claude writes its final message.</em></p><figure><img alt="Headline card: &quot;I Stopped Watching My Terminal. Claude Code Emails Me When It&#39;s Done.&quot;" src="https://cdn-images-1.medium.com/proxy/1*ShEx1dh6JntvNwk6IsG5yw.png" /></figure><p>I used to babysit Claude Code.</p><p>I’d give it a task, switch to another window, and then check back every few minutes.</p><p>Still working.</p><p>Still working.</p><p>Done. Who knows how long ago. Waiting for me to answer a question.</p><p>The agent wasn’t the bottleneck. <strong>I was.</strong> Every task finished with Claude waiting for me to notice.</p><p>So I made it tell me. Now, every time Claude Code finishes a turn, its final message lands in my inbox.</p><p>Setting it up took one hook and one small script. The part I didn’t expect was what it did to the way Claude writes.</p><h4>What a Stop hook is</h4><p>Claude Code lets you run your own commands when certain events happen: a session starting, a tool being called, a permission being requested.</p><p>The one I needed is <strong>Stop</strong>. It fires whenever Claude finishes responding, whether it&#39;s done, stuck, or asking you something.</p><p>That’s exactly the moment I want to know about.</p><p>Here’s the hook in my ~/.claude/settings.json:</p><pre>&quot;Stop&quot;: [<br>  {<br>    &quot;hooks&quot;: [<br>      {<br>        &quot;type&quot;: &quot;command&quot;,<br>        &quot;command&quot;: &quot;/Users/me/.claude/bin/email-on-stop&quot;,<br>        &quot;timeout&quot;: 60,<br>        &quot;async&quot;: true<br>      }<br>    ]<br>  }<br>]</pre><p>Two settings matter here.</p><p><strong>async: true</strong> means Claude doesn&#39;t wait for the email to send. Sending mail over SMTP can take a few seconds, and I don&#39;t want that delay on every turn.</p><p><strong>timeout: 60</strong> stops a hung mail server from hanging around forever.</p><h4>The script</h4><p>When the hook fires, Claude Code sends it a JSON payload on stdin. That payload includes the session ID, the working folder, and, usefully, Claude’s last message.</p><p>The script does four things:</p><ol><li>Reads the JSON.</li><li>Takes Claude’s final message.</li><li>Builds a subject line from the project folder and the message’s first line.</li><li>Sends it to me.</li></ol><p>The core is short:</p><pre>data = json.load(sys.stdin)<br><br>reply = (data.get(&quot;last_assistant_message&quot;) or &quot;&quot;).strip()<br>if not reply and data.get(&quot;transcript_path&quot;):<br>    reply = last_reply_from_transcript(data[&quot;transcript_path&quot;])<br><br>project = os.path.basename(data.get(&quot;cwd&quot;, &quot;&quot;).rstrip(&quot;/&quot;))<br>first_line = next((l.strip(&quot; #*`&gt;-&quot;) for l in reply.splitlines() if l.strip()), &quot;&quot;)<br>subject = f&quot;Done in {project}: {first_line[:67]}&quot;</pre><p>The fallback matters. If the last message isn’t in the payload, the script walks the session transcript backwards and collects the text Claude wrote after its last tool call.</p><p>The subject line matters more than I expected. “Done in medium: I rebuilt your Medium skill…” tells me which project finished and what happened without opening anything.</p><p>I also cap the body at 6,000 characters. Some final messages are long, and the full version is always in the session.</p><h4>Sending the email without leaking a password</h4><p>The sending part is a separate tiny script, email-me, that I can also call by hand. It uses Gmail&#39;s SMTP server with an app password.</p><p>The rule I care about: <strong>the password never touches a file.</strong> It lives in the macOS Keychain, and the script reads it at send time:</p><pre>password = subprocess.run(<br>    [&quot;security&quot;, &quot;find-generic-password&quot;,<br>     &quot;-a&quot;, ADDRESS, &quot;-s&quot;, &quot;claude-gmail-smtp&quot;, &quot;-w&quot;],<br>    capture_output=True, text=True, check=True,<br>).stdout.strip()</pre><p>No .env file to commit by accident. No secret in shell history.</p><p>Then it’s standard library only: smtplib.SMTP_SSL(&quot;smtp.gmail.com&quot;, 465), log in, send.</p><h4>The part I didn’t expect</h4><p>Once the emails started arriving, I noticed a problem.</p><p>Claude’s final messages were written for someone <em>watching the terminal</em>. They pointed back at things: “see above”, “as shown in the diff”.</p><p>In an email, “see above” points at nothing.</p><p>So I added this to my global CLAUDE.md:</p><blockquote>A Stop hook emails me your final message automatically every time you finish a turn… So make your final message stand on its own: what happened, what you need from me, and any links.</blockquote><p>That one instruction changed the shape of every reply.</p><p>Final messages now read like a short report. What was done. What failed. What I need to decide. Where the file is. I can act on them from my phone without opening a laptop.</p><p>I also told Claude <strong>not</strong> to email me itself at the end of a task, because the hook already does. Otherwise I’d get two emails for every task.</p><p>It’s still allowed to email me mid-task with the same email-me script if something is urgent.</p><h4>Two accounts, one inbox</h4><p>I run two Claude Code configurations, one under ~/.claude and one under ~/.claude-work. Both use the same hook.</p><p>The script checks CLAUDE_CONFIG_DIR and adds the account to the footer of the email, together with the folder and session ID. When something goes wrong, I know exactly which session to reopen.</p><h4>If you want to build this</h4><ul><li><strong>Use the </strong><strong>Stop event</strong>, with async: true.</li><li><strong>Read the final message</strong> from last_assistant_message, and fall back to the transcript.</li><li><strong>Put the project name in the subject.</strong> It’s the one thing you’ll scan in your inbox.</li><li><strong>Keep secrets in a keychain or secret manager</strong>, never in the script.</li><li><strong>Tell your agent its final message is going to be emailed.</strong> That’s the step that turns a notification into something you can actually use.</li></ul><h4>What this really fixed</h4><p>I thought I had a notification problem.</p><p>I actually had a <em>handoff</em> problem. The agent was finishing work, and the result was sitting in a place I wasn’t looking, written for a moment I’d already left.</p><p>The email solved the first half. One line in CLAUDE.md solved the second.</p><p>The best automation didn’t make Claude faster. It made the gap between “Claude is done” and “I know what happened” disappear.</p><p><em>I’m building </em><a href="https://shipx.live/"><em>ShipX</em></a><em> mostly with Claude Code, and I write up the setup pieces that actually stick. Next: the rest of my Claude Code setup, layer by layer. Follow if you want that one.</em></p><p><em>Drafted with AI assistance (Claude), working from my own code, configs and notes.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21370cbb552d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Built a Launch Platform for People With No Audience. Its First Sale Was $9.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://suhailroushan.medium.com/i-built-a-launch-platform-for-people-with-no-audience-its-first-sale-was-9-c3454fa4ed61?source=rss-4d10b081bc95------2"><img src="https://cdn-images-1.medium.com/max/847/0*_NDK0WchS3obcDfU.png" width="847"></a></p><p class="medium-feed-snippet">My Launch Platform for People With No Audience Made Its First Sale: $9</p><p class="medium-feed-link"><a href="https://suhailroushan.medium.com/i-built-a-launch-platform-for-people-with-no-audience-its-first-sale-was-9-c3454fa4ed61?source=rss-4d10b081bc95------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://suhailroushan.medium.com/i-built-a-launch-platform-for-people-with-no-audience-its-first-sale-was-9-c3454fa4ed61?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/c3454fa4ed61</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[build-in-public]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[saas]]></category>
            <category><![CDATA[indiehackers]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 16:34:47 GMT</pubDate>
            <atom:updated>2026-10-24T16:34:47.823Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Studied the Medium Stories That Took Off in My Feed. The Winners Admitted Something.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://suhailroushan.medium.com/i-studied-the-medium-stories-that-took-off-in-my-feed-the-winners-admitted-something-1be939b954e0?source=rss-4d10b081bc95------2"><img src="https://cdn-images-1.medium.com/max/600/1*16KpWOm7hzsAX5B09Kjp3A.png" width="600"></a></p><p class="medium-feed-link"><a href="https://suhailroushan.medium.com/i-studied-the-medium-stories-that-took-off-in-my-feed-the-winners-admitted-something-1be939b954e0?source=rss-4d10b081bc95------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://suhailroushan.medium.com/i-studied-the-medium-stories-that-took-off-in-my-feed-the-winners-admitted-something-1be939b954e0?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/1be939b954e0</guid>
            <category><![CDATA[content-marketing]]></category>
            <category><![CDATA[writing-tips]]></category>
            <category><![CDATA[blogging]]></category>
            <category><![CDATA[medium]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 11:58:29 GMT</pubDate>
            <atom:updated>2026-10-22T11:58:29.007Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Built a Product. Then Realized Getting Users Was the Hard Part.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://suhailroushan.medium.com/i-built-a-product-then-realized-getting-users-was-the-hard-part-dfd60001dcd1?source=rss-4d10b081bc95------2"><img src="https://cdn-images-1.medium.com/max/600/1*PJeWu7U8xXIwqp4V-KpvXQ.png" width="600"></a></p><p class="medium-feed-snippet">A quiet launch doesn&apos;t mean nobody wants your product. Here&apos;s what shipping taught me about finding first users, and why I built ShipX</p><p class="medium-feed-link"><a href="https://suhailroushan.medium.com/i-built-a-product-then-realized-getting-users-was-the-hard-part-dfd60001dcd1?source=rss-4d10b081bc95------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://suhailroushan.medium.com/i-built-a-product-then-realized-getting-users-was-the-hard-part-dfd60001dcd1?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/dfd60001dcd1</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[indiehackers]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[product-development]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 10:57:07 GMT</pubDate>
            <atom:updated>2026-10-18T10:57:07.856Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[In 2019, Everyone Thought Elon Musk Was Joking About a Flying Tesla. He Wasn’t.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://suhailroushan.medium.com/in-2019-everyone-thought-elon-musk-was-joking-about-a-flying-tesla-he-wasnt-ae73fd2f6fd6?source=rss-4d10b081bc95------2"><img src="https://cdn-images-1.medium.com/max/1920/0*mrfSJF17-XbFVdHs.jpg" width="1920"></a></p><p class="medium-feed-snippet">In 2019, Musk told MKBHD he wasn&apos;t joking about a hovering Tesla. In 2026, it left the ground. Here&apos;s how it works</p><p class="medium-feed-link"><a href="https://suhailroushan.medium.com/in-2019-everyone-thought-elon-musk-was-joking-about-a-flying-tesla-he-wasnt-ae73fd2f6fd6?source=rss-4d10b081bc95------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://suhailroushan.medium.com/in-2019-everyone-thought-elon-musk-was-joking-about-a-flying-tesla-he-wasnt-ae73fd2f6fd6?source=rss-4d10b081bc95------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae73fd2f6fd6</guid>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[elon-musk]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tesla]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Suhail Roushan]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 18:47:14 GMT</pubDate>
            <atom:updated>2026-10-16T18:47:14.114Z</atom:updated>
        </item>
    </channel>
</rss>