<?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 Nitheesh T Ganesh on Medium]]></title>
        <description><![CDATA[Stories by Nitheesh T Ganesh on Medium]]></description>
        <link>https://medium.com/@iamntg?source=rss-3c52e735d40c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*yRk3zunT7hx3td6WPoZ7xA.jpeg</url>
            <title>Stories by Nitheesh T Ganesh on Medium</title>
            <link>https://medium.com/@iamntg?source=rss-3c52e735d40c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 14:26:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@iamntg/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[Six decisions, and the app is theirs | A Case Study]]></title>
            <link>https://medium.com/design-bootcamp/six-decisions-and-the-app-is-theirs-a-case-study-143e38df5810?source=rss-3c52e735d40c------2</link>
            <guid isPermaLink="false">https://medium.com/p/143e38df5810</guid>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[accessibility]]></category>
            <category><![CDATA[design-systems]]></category>
            <category><![CDATA[react-native]]></category>
            <dc:creator><![CDATA[Nitheesh T Ganesh]]></dc:creator>
            <pubDate>Wed, 29 Jul 2026 09:41:08 GMT</pubDate>
            <atom:updated>2026-07-29T09:41:08.759Z</atom:updated>
            <content:encoded><![CDATA[<h3>How I designed and built the white-labeling module behind a multi-tenant platform</h3><p><strong>Try it before you read it.</strong></p><ul><li>Admin tool → <a href="https://wl-admin-ak5.pages.dev/"><strong>wl-admin-ak5.pages.dev</strong></a> — where a business owner picks their brand</li><li>Branded app → <a href="https://wl-app-b20.pages.dev/"><strong>wl-app-b20.pages.dev</strong></a> — the app their customers get</li></ul><p>Open both side by side. Change the primary colour in the admin tool, watch the phone preview update as you type, hit Publish, reload the app tab. Same theme, same components, now on the phone.</p><p>That sentence is the whole project.</p><h3>Where this fits</h3><p>This isn’t a standalone product. It’s <strong>one module inside a larger platform</strong> — the platform onboards small businesses (restaurants, salons, studios) and hands each one four surfaces: a storefront website, a customer mobile app, an owner dashboard, and an order manager for the counter.</p><p>All four have to look like <em>that</em> business. Not like the platform with a logo dropped in the corner.</p><p>So somewhere in the system there has to be a single place that owns the answer to “what does this brand look like?” — where a theme is authored, validated, published and served. Every other surface reads what it publishes. That module is what I built.</p><p>The hard part isn’t storing a hex code. It’s who’s doing the storing:</p><ul><li>The owner <strong>is not a designer.</strong> They know their sign is red. They don’t know what a hover state is.</li><li>There’s <strong>no staging environment.</strong> What they publish is what their customers open.</li><li><strong>Nobody reviews it.</strong> There’s no design QA between the colour picker and a real customer’s phone.</li></ul><p>Two things can go wrong, and every decision below is a response to one of them: the owner ships something illegible and never finds out, or the preview lies and they stop trusting the tool.</p><h3>Six inputs, and everything else derived</h3><p>The entire surface an owner can change:</p><pre>brand      businessName, logo<br>colors     primary, secondary, accent, background<br>typography one of five font pairings<br>shape      corner style — sharp | subtle | rounded | pill<br>buttons    filled | outline | soft</pre><p>That’s it. From those, the engine derives the rest:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vwUzPAY_5KkZDQnglhZ8cQ.png" /></figure><p>The owner picksThe engine derives4 hex coloursevery colour role — surfaces, borders, text, hover, pressed, disabled, and “on” colours chosen for legibility1 font pairingdisplay + body faces, weights, a full type scale, tracking1 corner styleevery radius in the app, in px1 button stylefill, border, foreground, and all its states — light <strong>and</strong> dark, both resolved and served</p><p><strong>The constrained input surface is the product.</strong> An owner who can set a hover shade is an owner who can ship an unreadable app. Every input I removed from that list is a class of support ticket that can’t happen.</p><p>Shade derivation runs in OKLCH rather than the usual HSL mixing, because HSL gives you a “10% darker” that reads as 30% darker on a yellow and 3% on a blue. When nobody eyeballs the derived shades before customers see them, predictable perceived lightness stops being a nicety.</p><p>Light and dark are a render-time decision, not something the owner sets. The device decides. And since the owner never chooses the dark surface, a <em>warning</em> about dark mode would be unactionable — so the resolver guarantees dark-mode legibility by construction instead of reporting on it.</p><h3>Contrast is a gate, not a hint</h3><p>This is the decision I’d keep if I could only keep one.</p><p>Every colour pair that can appear in the app is checked, against <strong>two different bars</strong>, because using a single bar is how a contrast checker earns a reputation for crying wolf:</p><ul><li><strong>4.5:1</strong> for anything you read (WCAG SC 1.4.3)</li><li><strong>3:1</strong> for button fills, accent chips, graphics (SC 1.4.11)</li></ul><p>A primary button fill is a graphical object, not body copy. Holding a warm amber accent to 4.5:1 would flag essentially every accent a small business would ever pick — and a warning that fires on every reasonable choice trains people to ignore the one that matters.</p><p>Four things around that engine matter more than the ratios:</p><p><strong>Failing or warning both hold the publish.</strong> A warning isn’t a hint you can scroll past. The owner either applies the fix or explicitly acknowledges the pair. There’s no route to publishing an unreadable theme <em>by accident</em>, which is the only kind that actually happens.</p><p><strong>The suggested fix is the owner’s colour, corrected.</strong> It nudges along the same hue to the nearest shade that passes — not a different colour chosen for them. And it’s guaranteed to actually clear the bar it was offered for, because a one-click fix that fixes nothing is worse than offering none.</p><p><strong>The server re-validates against the stored draft</strong>, never against what the request sent. Client-side validation is a courtesy; this one is a gate.</p><p><strong>One implementation, not two.</strong> The browser shows the result the instant a bad pair is set, and the API runs the identical function on publish. Never trust the client — but also never let the two disagree about what “readable” means.</p><p>There’s a moment in the demo where this pays off: pick a pale yellow, and the tool stops you, shows you the ratio, and offers a gold that works. You can override it — but only on purpose.</p><h3>The one rule: the preview is the app</h3><blockquote><em>If the web preview and the phone can disagree about what a theme looks like, the design is wrong.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ePQ6DiD7b6E0Ox31l2DAlw.png" /></figure><p>Most theming systems share a <strong>resolver</strong> — one function that turns brand inputs into resolved colours — and then let each platform render with its own components. That gets you two surfaces that agree about #e23d28 and disagree about everything the owner is actually looking at: where the card sits, how the corner reads at that radius, whether the active tab is legible.</p><p>So this one shares the resolver <strong>and</strong> the renderer.</p><p>The components are written against React Native. The Expo app renders them directly. The admin tool resolves react-native to react-native-web and mounts the same component tree inside the phone frame in its preview pane. The screens, the tab bar, the carousel, the icons — those <em>are</em> the app&#39;s, running in an iframe.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zj1xo1Af4WjO1wN--0NKCQ.png" /></figure><p>Which means the preview can only ever promise what a phone can actually do. That cost me things I liked: CSS grid, CSS transitions, and a variable-axis icon font that lit the active tab. React Native has none of them, so the icons became SVG paths. The design wasn’t wrong — it was describing a surface that hadn’t yet had to survive a device.</p><h3>The rest of the UX decisions</h3><p><strong>The tool’s chrome stays out of the way.</strong> Neutral greys, near-black text, exactly one muted accent for interactive states. The tenant’s brand colours are the only saturated pixels on screen — never the publish button, never a focus ring. If the tool has its own opinions in colour, the owner can’t tell which colours are theirs.</p><p><strong>The loading skeleton is grey, deliberately.</strong> A skeleton in the tenant’s colours looks like a rendered theme, and the owner will judge it as one.</p><p><strong>Components can’t be handed a colour.</strong> They take content and variant props only; the theme arrives through a provider. A &lt;Button primaryColor=&quot;#ff0000&quot;&gt; would typecheck and would be a bug — so there&#39;s simply nowhere to pass the mistake in. Same reasoning as the owner&#39;s six inputs, one level down.</p><p><strong>Publishing is append-only.</strong> Versions are immutable, and rollback republishes an older version <em>forward</em> as a new one rather than deleting anything. The summary of what changed (“Font pairing: Modern → Editorial”) is computed on the server and stored with the version, so a year-old version still explains itself.</p><p><strong>On desktop, the app draws itself inside a phone.</strong> The mobile app’s web build is the real app, so a desktop browser stretches a 390pt layout across 1440 and the two-column grid turns into two very wide rows. Nothing is broken; it just stops reading as the thing it is. So on a desktop it renders inside a drawn device, and on a phone browser it goes full-bleed with no frame. A (pointer: coarse) media query decides — width alone gets it wrong in both directions, because a tablet is wide and should still get the real app, while a desktop window dragged narrow should lose the frame.</p><h3>How I architected it</h3><p>Three shared packages, three apps:</p><pre>packages/<br>  theme       token schema, resolver, contrast engine, font registry, presets<br>  ui          the app&#39;s components — rendered by the phone and the preview alike<br>  api-client  the wire contract + typed client, shared by web and mobile<br>apps/<br>  web         the admin tool      (Vite + React + TypeScript + Tailwind)<br>  api         the service         (Fastify + Prisma + Zod + Postgres)<br>  mobile      the branded app     (Expo React Native)</pre><p>Two boundaries do most of the work:</p><p><strong>packages/theme has no DOM, Node or React Native imports.</strong> That&#39;s the property that lets the browser, the phone, and the server all run the identical validation. It&#39;s also why the package is reusable on its own — it&#39;s a schema, a resolver and a contrast engine in plain TypeScript, with no dependency on the rest of the repo.</p><p><strong>packages/api-client owns the request and response schemas, and the API imports them.</strong> One definition, validated by the server and parsed by both clients. The alternative is restating each payload twice and waiting for the two to drift — which they do, silently, in the direction of whichever side gets edited more often.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X7BAieX7JdBPwaYnN6Mgog.png" /></figure><p>Everything a <em>host</em> knows that the shared code shouldn’t — which fonts are bundled versus loaded from CSS, how wide the viewport is, whether an OS paints the status bar — is passed in by the host rather than detected inside the components. It looks like over-engineering until you notice that react-native-web reports the <em>browser window’s</em> width, and a component that measured itself would lay out for a 1440pt phone.</p><p>The whole thing runs on free tiers: Postgres on Neon, the API on Render, both front ends on Cloudflare Pages, each rebuilding from a Git push.</p><h3>Built with Claude Design and Claude Code</h3><p>I designed the admin tool in <strong>Claude Design</strong> before writing a line of application code , <a href="https://claude.ai/code/artifact/e7327603-ec45-4851-99de-87955cf08479">the original artifact is here</a> and both screens are exported into the repo as the reference the build is checked against. Keeping the export around is what makes “does this match?” a question with an answer.</p><p>It also recorded the one place the design lost an argument. The default theme’s accent was a bright gold, and the contrast engine flagged it against white. The design was right about the palette and the engine was right about that one swatch, so I moved the accent to the deepest gold that still clears the bar and left a comment explaining why. When your own validation tool overrules your own design and you keep the tool’s answer, the tool is real.</p><p>The implementation was then built with <strong>Claude Code</strong>, bottom-up in dependency order: the theme package first, then the wire contract, the API, the admin tool, the app, and finally extracting the shared component package once both hosts existed and the duplication was visible rather than hypothetical.</p><p><strong>Links</strong></p><ul><li>🧩 Repository — <a href="https://github.com/iamntg/WhitelabelingEngine">github.com/iamntg/WhitelabelingEngine</a> (MIT)</li><li>🎨 Admin tool (live) — <a href="https://wl-admin-ak5.pages.dev/">wl-admin-ak5.pages.dev</a></li><li>📱 Branded app (live) — <a href="https://wl-app-b20.pages.dev/">wl-app-b20.pages.dev</a></li><li>✏️ Original design — <a href="https://claude.ai/code/artifact/e7327603-ec45-4851-99de-87955cf08479">the Claude Design artifact</a></li></ul><p><em>The demo is a shared sandbox with no login, everyone edits the same three brands, so expect to find someone else’s colours, and feel free to leave your own. The API sleeps after 15 minutes idle, so the first load can take up to a minute.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=143e38df5810" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/six-decisions-and-the-app-is-theirs-a-case-study-143e38df5810">Six decisions, and the app is theirs | A Case Study</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vibe Coding vs. Building With Intention]]></title>
            <link>https://medium.com/@iamntg/vibe-coding-vs-building-with-intention-751f2dac1a7e?source=rss-3c52e735d40c------2</link>
            <guid isPermaLink="false">https://medium.com/p/751f2dac1a7e</guid>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[building-in-public]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Nitheesh T Ganesh]]></dc:creator>
            <pubDate>Thu, 16 Jul 2026 00:03:52 GMT</pubDate>
            <atom:updated>2026-07-16T19:03:14.914Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>How I turned a graveyard of half finished notes into a constellation, and why I refused to “vibe code” my way there.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wbwdYgnHpSBDNqxEDiQ0VA.png" /><figcaption>The home screen, where every idea is a star in the constellation</figcaption></figure><p>Google Keep is where my ideas go to die. No folder, no labels, no system. Just a heap.</p><p>There are 400 something notes buried in there, in no particular order, tagged with nothing. A hook for a video I still haven’t filmed. A plot twist for a story whose main character doesn’t have a name yet. A line of dialogue so good I was certain I’d remember the scene it belonged to (reader, I did not). Three thumbnail concepts. A villain’s motivation scribbled at 2 a.m. And, for reasons lost to history, a grocery list.</p><p>I’m a part time content creator and story writer, which is a polite way of saying I generate way more ideas than I will ever follow through on. And here’s the thing nobody warns you about capturing everything: <strong>capture was never my problem.</strong></p><p>I’m elite at capture. I can get a thought out of my head and into an app in about two seconds. What I’m catastrophically bad at is everything that happens <em>after</em>. Coming back to it. Remembering it exists at all. And the real gut punch, noticing that the idea I had in March was secretly the missing half of the idea I had in July.</p><p>My notes app wasn’t a garden. It was a graveyard. Neat little rows of tombstones, each one a thought I swore I’d revisit, each one quietly gathering dust. Honestly, half my best ideas got snapped out of existence and I didn’t even get a dramatic slow motion scene for it.</p><p>So I built the app I actually needed. It’s called <strong>Sparkles</strong>, and instead of another bottomless list, it turns your ideas into a night sky.</p><p><strong>The problem isn’t capture. It’s connection.</strong></p><p>When you dump thoughts into a linear notes app, two quiet little tragedies happen:</p><p>1. <strong>They fall off the bottom.</strong> Out of sight, out of mind. That note from three weeks ago may as well be in the Phantom Zone.</p><p>2. <strong>They never touch each other.</strong> A notes app treats every thought as a lonely island. But ideas aren’t islands. They’re a team. Their whole value is in the <em>lines between them</em>.</p><p>One idea on its own is Hawkeye showing up to an alien invasion with a bow and a bad attitude. A <em>connected</em> set of ideas is the whole Avengers actually assembling. The magic was never any single note. It was the crossover event.</p><p>A list can’t show you that the joke you wrote on Tuesday belongs in the scene you sketched last month. A folder can’t lean over and whisper, “psst, these five random notes are secretly one big series.” That connective tissue, the exact part your brain is <em>worst</em> at holding onto, is the part every notes app cheerfully throws in the bin.</p><p><strong>Who this is really for</strong></p><p>I built Sparkles for me. But the more I talked about it, the more I realized “people like me” is a much bigger club than I thought, and the membership card is just “too many tabs open, in your head and your browser.”</p><p><strong>🌙 The scattered capturers.</strong> You own three notes apps and trust exactly none of them. You screenshot things and never look at them again. You need capture to be frictionless <em>*and*</em> you need old sparks to resurface on their own, because you will not go looking for them. Be honest.</p><p><strong>✍️ Storytellers and novelists.</strong> Writers have always known ideas are spatial. That’s literally what the corkboard and index cards and red string is for. A character’s fear lives in one note, a betrayal in another, a stray line of theme in a third. Sparkles is that corkboard, except the string is real and you can actually follow it. <em>*That childhood wound connects to that third act betrayal,*</em> and now you can see it instead of hoping you remember it.</p><p><strong>🎬 Scriptwriters.</strong> A funny line. A cold open image. A theme you keep circling like a shark. In a flat list, that’s just noise. As a constellation, you can link the joke straight to the scene it was born for and watch your script’s skeleton appear before you’ve written a single page.</p><p><strong>🎥 Content creators.</strong> You hoard hooks, video concepts, a spicy comment that sent you down a rabbit hole. The gold is never one idea. It’s realizing eight of them are quietly the same series, and your entire content calendar has been hiding in plain sight this whole time.</p><p><strong>🧠 Deep thinkers and researchers.</strong> You read a quote in April, half build an argument in June, and the lightning only strikes when the two finally collide. A tool that nudges “hey, these belong together” is doing the single most valuable intellectual move there is: synthesis.</p><p><strong>⚡ ADHD minds.</strong> This one’s personal. When a thought lands, you have <em>seconds</em> before it’s gone forever, so capture has to be instant with zero “which folder, which tag, which planet” decisions. And because “out of sight” genuinely means “out of existence” for us, a <strong>visual, spatial</strong> memory beats a text list every single time. A sky you can wander is a lot kinder to a non linear brain than a scroll that keeps punishing you for forgetting.</p><p>If any of those made you nod a little too hard, congratulations, you’re the target audience.</p><p><strong>Meet Sparkles</strong></p><p>The whole idea fits in one sentence: <strong>every thought is a star, every connection is a thread, and your mind becomes a sky you can actually walk through.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FUGOr82AVWKJV8THEer5EQ.png" /><figcaption>Capture a spark in two seconds, typed or spoken</figcaption></figure><p><strong>Capture is instant.</strong> Type it, or just speak it and let it transcribe. No title, no folder, no tags, no decisions. Two seconds and you’ve lit a new star. It all works fully offline too, because a thought should never have to sit in a loading spinner while it dies.</p><p><strong>Then it becomes a constellation.</strong> Your ideas float as glowing stars on the home screen. Tap one, connect it to another, and watch a thread light up between them. Older, lonelier sparks fade gently instead of vanishing. Present, but not yelling for attention.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bQVAW5bGfii0XqGNN5HR8g.png" /><figcaption>The Idea stream, for when you want to read instead of wander</figcaption></figure><p><strong>Want the boring linear view?</strong> It’s right there. A searchable “Stream” of every spark, filterable by what’s linked or recent. Exploration and structure, sitting politely side by side.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MR3G63w3f_C4mL8GyAz-Gw.png" /><figcaption>Developing a single spark or Link them</figcaption></figure><p><strong>And when you’re ready, AI helps you see the shape.</strong> Ask it to cluster everything into themes, and the patterns you could never hold in your head just surface on their own. The AI is always <em>offered</em>, never forced. It’s a quiet button in the corner, not a needy sidekick asking if you’ve considered its ideas.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bGDV-PbsTuF_htG_jjt2vA.png" /><figcaption>Customize AI settings according to your needs</figcaption></figure><p>One promise sits under all of it: <strong>your ideas are yours.</strong> Everything lives on your device. Backups are encrypted on your phone and go only to <em>your</em> Google Drive. No server I run ever lays eyes on a single thought.</p><p><strong>The part I care about most: I built this on purpose, not on vibes</strong></p><p>Here’s the fork in the road where I could have taken the easy path, and didn’t.</p><p>There’s a way of building software right now that people fondly call <strong>”vibe coding.”</strong> You open an AI, describe what you want, accept whatever it hands back, and keep poking it by feel until something roughly works. For a weekend prototype it feels like a genuine superpower. I love it for that.</p><p>But vibe coding has a bill that always comes due later, and the bill says: <strong>there is no source of truth.</strong> Features drift. You forget <em>why</em> anything works the way it does. The AI happily bulldozes last week’s decision because nobody ever wrote that decision down. You end up with an app that runs but doesn’t <em>mean</em> anything, just a tall stack of “sure-exactly-whatever” suggestions in a trench coat.</p><p>For Sparkles, that felt like a betrayal of the whole point. This is an app about “<strong>intention”</strong>, about treating your own thoughts like they matter. I was not about to build it by accident.</p><p>So I used <strong>Spec Driven Development (SDD)</strong> instead.</p><p><strong>What that actually looks like</strong></p><p>SDD flips the order everyone’s used to. Instead of code first and meaning later, <strong>the spec is the source of truth and the code is just a projection of it.</strong> The flow goes:</p><p>&gt; <strong>constitution → spec → plan → tasks → implement</strong></p><p>- <strong>A constitution.</strong> Before a single feature, I wrote down the app’s non negotiable principles. <em>Local first, always. Your data is yours. Calm by default. AI is offered, never forced.</em> Every later decision has to bow to these. Batman has exactly one rule. My app has six, and like Batman, the whole personality falls apart the second you break one.</p><p>- <strong>A spec for every feature.</strong> The <em>what</em> and the <em>why</em>, in plain English, with acceptance criteria. No code, no tech jargon. Just: what should this do, and how will we know it’s actually done?</p><p>- <strong>A plan.</strong> The <em>how</em>. Data model, architecture, trade offs, every one of them traced back to a line in the spec.</p><p>- <strong>Tasks.</strong> The plan chopped into ordered, checkable steps.</p><p>- <strong>Then, and only then, the actual code.</strong></p><p>Here’s a real moment where this saved me. The spec for “Capture” had one ironclad line: <em>*capturing an idea must never wait on the network.*</em> That single sentence, born straight from the “local first” article in the constitution, forced an entire architecture decision. A local database as the source of truth, with sync as an optional layer bolted on top. If I’d been vibe coding, I would have wired capture to a cloud API on autopilot and quietly gutted the soul of the app. The spec caught it before I wrote one line of code.</p><p><strong>But wait, didn’t you use AI anyway?</strong></p><p>Oh, constantly. I built this with a <em>ton</em> of AI help. The difference is entirely in <em>what the AI was building from.</em></p><p>Vibe coding hands the AI a vague wish and a prayer. SDD hands it a <strong>blueprint</strong>. Give the AI a constitution and a spec to work against, and it stops being a slot machine you keep yanking and becomes an actual collaborator that knows exactly what you’re making. Better still, one you can <em>check</em>, because “done” was defined before the work ever started.</p><p>You know,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/245/1*_WtjW_5teqE1YVu609WQdA.gif" /></figure><p>“Thats my secret Cap, I always write the spec first.”</p><p>The AI didn’t replace my intention. The spec <strong>protected</strong> it.</p><p><strong>Assembling the team (the actual build)</strong></p><p>None of this got built in one heroic sitting, and definitely not with a single tool. The origin story went through phases, like any self respecting hero.</p><p>It started in <strong>Google Stitch</strong>, where I tried the first mockups and finally got the shape of the screens out of my head. From there I jumped into <strong>Google’s Antigravity</strong>, riding Gemini and a very healthy pile of free credits, to rough out the early prototypes. (One does not simply say no to free credits.)</p><p>Then I tried <strong>Claude Design</strong>, and <em>that</em> was the real game changer. It was the first moment the app stopped looking like a wireframe and started looking like something with a soul. Once the design clicked, I handed it to <strong>Claude Code</strong>, wired up through <strong>MCP</strong> and running on <strong>Claude Opus 4.8</strong>, to turn those screens into a real, working app. That was the pairing that did the heavy lifting: the design in one hand, the implementation in the other, both aimed at the same spec. And for one short, glorious window I got to build on <strong>Fable</strong>, which, and I truly do not say this lightly, was absolutely awesome.</p><p>If vibe coding is a slot machine, this felt more like the good kind of team up, where every member actually had a role and knew the plan.</p><p>One more deliberate call worth naming: the app runs on <strong>React Native with Expo</strong>, and that was no accident either. I’m one part time human, not a studio. Expo let me write the app once and ship it to iPhone, Android, and the web from a single codebase, which is basically the only way a solo builder covers every platform without losing their mind. It also fits a <strong>local first</strong> app like a glove, since an on device SQLite database sits right there in the stack, so “your ideas never leave your phone” was the default instead of a battle. And the feedback loop is almost unfair: save a file, watch it update live on a real phone in seconds. When you’re fussing over the exact glow of a star or how slowly a thread fades in, that instant loop is the whole ballgame.</p><h3><strong>Vibe coding is dead.</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*WCIBu0sXP6FuSiagkzfBUA.gif" /></figure><p>Let me just say the quiet part at full volume: <strong>vibe coding, as a way to build anything you actually care about, is dead.</strong></p><p>It had a great run. For about a year, “just prompt it and see what happens” felt like the future had arrived early. But we’ve all seen where that road dead ends now. A repo nobody understands. Bugs that respawn the instant you kill them. An AI confidently undoing a decision from last Tuesday because there was never a written decision to respect in the first place. The demos were dazzling. The maintenance was a horror movie. What looked like speed was just <strong>debt wearing a really nice UI.</strong></p><p>The pit is seductive because hour one feels <em>amazing</em>. You get a working screen in minutes and think, “why did anyone ever bother planning?” Then hour twenty rolls in. Nothing has a reason. Every tiny change threatens to break three things you forgot existed. You’re not building anymore, you’re just negotiating with a slot machine and hoping it likes you.</p><p>Here’s the mindset shift that dragged me out of it:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AwsVldnPzK4POUU2o8Si9A.png" /><figcaption>Prompts are cheap. Intentions are expensive.</figcaption></figure><p>Writing a spec feels slower for roughly a day. Then it pays you back for the entire life of the project, because every prompt after that is <em>aimed</em>. The AI is still doing ninety percent of the typing. It’s just no longer allowed to <em>guess what you meant</em>. That’s the whole game. The people still shipping a year from now won’t be the ones with the cleverest prompts. They’ll be the ones who bothered to write down what they were trying to build.</p><p>So if you take one single thing from this whole article: <strong>do not fall into the pit.</strong> Vibe your throwaway prototypes to your heart’s content, that’s exactly what they’re for. But the moment something starts to actually matter to you, stop vibing and start specifying. Intention is the one thing the machine cannot generate for you.</p><p><strong>What I’d tell you if you’re standing where I was</strong></p><p>If you’ve got your own graveyard of notes, your own Keep folder with 400 little tombstones, I promise I get it. The problem was never a shortage of ideas. You have too many, scattered too wide, with no way to see how any of them connect.</p><p>Maybe the fix isn’t another folder. Maybe it’s a sky.</p><p>And if you’re building something of your own, it’s worth asking, honestly, whether you’re building <em>by accident</em> or <em>on purpose</em>. Vibe your prototypes, sure. But the moment a project starts to <em>*mean*</em> something to you, write down what it means first. Let the intention lead. The code will follow.</p><p>I built a night sky for my ideas. Turns out I needed one for how I build things too.</p><p><strong>Try it, and go read the specs yourself</strong></p><p>Sparkles is <strong>open source.</strong> The whole thing lives right here:</p><h4><strong>👉 </strong><a href="https://github.com/iamntg/Sparkles"><strong>Sparkles</strong></a><strong> [github.com/iamntg/Sparkles]</strong></h4><p>Clone it, get it running locally in about a minute (the README genuinely walks you through it), and go turn a few of your own scattered thoughts into stars. And if you’re even a little curious what “spec driven” looks like in the wild, open up [`docs/specs/`](https://github.com/iamntg/Sparkles/tree/main/docs/specs). Every feature’s constitution, spec, plan, and tasks are sitting there in the open. That folder <em>is</em> the app’s intention, written down where you can read it.</p><p>If you give it a spin, I genuinely want to hear from you:</p><p>- ⭐ <strong>Star the repo</strong> if the idea speaks to you. It helps other scattered minds find their way here.</p><p>- 🐛 <strong>Open an issue</strong> with what broke, what confused you, or what you wish it did.</p><p>- 🌌 <strong>Tell me how your ideas connected.</strong> The surprising links are the entire point.</p><p>— -</p><p><em>Sparkles is a local first, offline friendly thinking space, built with Expo, React Native, and an almost unreasonable number of deliberately written specs. If “turn my scattered thoughts into a constellation” spoke to you, I’d love to hear how your ideas connect.</em></p><p><em>✦ Thanks for reading. If this resonated, a clap or a comment helps more people with graveyards of notes find it, and hey, go [give Sparkles a try](</em>https://github.com/iamntg/Sparkles<em>).</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=751f2dac1a7e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>