This is where JgSaw was built. One café. Mass conversations with Claude. And a lot of questionable decisions about what counts as a reasonable bedtime.
The LinkedIn version of this story would be: “I vibe coded a B2B content platform in 20 minutes, no coding required.” The real version involves scrapping my first build entirely, mass debugging sessions at 2am, and learning that vibe coding is less magic, more rigour. My friends started wondering if I was planning to move in with Claude. My dog has heard more about prompt architecture than any dog should.
I’d been chewing on a problem for a while: sales needed content to move deals forward, but getting something personalized, timely, and on-brand felt impossible. Most lost faith on content altogether. AI tools were fast, but the output was generic, youd spend more time reworking it than writing it on your own. So by the time the final piece was out, it was too late, or the content did not land.
So I opened a no-code platform out of curiosity and started building. No coding background. Surface-level understanding of AI. Just a hypothesis and some strange optimism.
Three months later, I have a working MVP, a product called JgSaw. Called the first version Think Humaine (the idea that even within B2B, you talk to the human, and of course a creative twist on using ai in the word). JgSaw came along the way, when I noticed that every feature of the product felt like piecing a pizzle together to get to a complete picture.
These 3 months were an amazing experience, reminded me of a UB40 song, every hour every day I am learing more, the more I learn the less I know of before (hope i got it right ;). There’s so much to share, but here are some highpoints of things/experiences that I picked up through this journey, for whatever it’s worth:
Pick one platform
Before I started, I asked ChatGPT which no-code platform to use. Here’s what it said:
“If you want control and durability with no-code speed, go WeWeb + Xano + Make/n8n for the core app, add PostHog for analytics and Stripe for billing; ship a LinkedIn-post skinny slice with full provenance/draft-publish/versioning before expanding DNA depth and workflows.”
I stared at that for ten minutes. Understood maybe 40% of it. Dozens of platforms, hundreds of combinations, endless Reddit debates. You can spend weeks researching and never build anything.
Here’s what actually matters: a conversational interface that lets you discuss before implementing. You must feel comfortable and find the UX easay for you to handle (maybe that’s why its called “vibe” coding 🙂 Base44’s discuss function was a really helpful tool once I found it. Solid data management. And backend functionality that doesn’t assume you already understand authentication flows and user permissions. Some platforms are built for people with technical foundations. If that’s not you, choose accordingly.
Don’t overthink it. Pick something reasonable, start building, find your rhythm.
“Your first version teaches you what you’re building. Sometimes you have to throw it away.”
A few weeks in, I had a working prototype. Features stacking up nicely. Then I stepped back and realized the user journey was broken. I’d been adding capabilities without thinking about how they connected. It seemed more like patchwork, not product.
I scrapped it and started fresh. Painful at the time. Best decision I made. The first version was tuition; the second was the real thing.
Talk to the AI, don’t just prompt
The instinct is to give instructions and wait for output. But the AI isn’t just an executor. It’s a collaborator. The better you get at working with it, the better the results.
Discuss before building. Don’t say “build me this feature.” Say “I want to build this, what’s the best approach? What are the trade-offs? What might break?” You get visibility into the sequence before committing, and sometimes the AI suggests something better than what you had in mind.
Ask for critical review. We all hear of the comformation bias. all AI love to tell you how amazing you are. But if you push them to be critical, they will lay into you! Explicitly asking “what’s wrong with this approach? or critically review and discuss before publishing” catches problems before they become expensive.
Ask it to explain. When the platform responds with code snippets or technical references you don’t understand, stop and ask: “Explain this to me in plain English.” Suddenly you understand what’s actually happening.
Use screenshots. Show it errors or areas you need changed in the UI. A screenshot communicates faster and more precisely than a paragraph of description.
The developer console trick. When something breaks, right-click the page → Inspect (MacBook) → Console tab → copy what’s there. Paste it to the AI with context on what you were trying to do. The console shows errors in a language the AI understands and you can see it light up immediately!
“The AI isn’t an executor. It’s a collaborator.”
Use multiple models
Different models are good at different things. Using them together made a noticeable difference.
I started with ChatGPT for brainstorming features and thinking through user journeys. Good for exploration and breadth.
Then I hit a wall with prompt engineering. Switched to Claude. The depth of reasoning was noticeably different. I built the entire 8-layer prompt architecture in Claude, refining each asset-specific prompt until the structure was tight.
Then I brought everything back to Base44 and asked for critical review. Not “what do you think?” but “what’s wrong with this? Where are the weak points?” I wanted to counter my own confirmation bias.
Finally, I asked Base44 to merge its approach with Claude’s output while following my architecture. The result was stronger than any single model would have produced.
Back up everything
Early on, I started duplicating my project before any significant change. Paranoid? Maybe. But it saved me.
Most platforms have revert functionality. But I ran into a cache problem once that no amount of reverting could fix. Had to restore from a duplicate project I’d made two days earlier. Lost some work, not everything.
The practice: before any major change, duplicate the entire project. Takes seconds. Insurance you’ll be grateful for.
Finding workarounds
The UI was a struggle. Base44 wasn’t producing the visual design I wanted. Output was functional but didn’t look right. Tried different descriptions, referenced other sites. Nothing clicked.
So I designed the UI in Figma using their AI tools. Got it looking the way I’d envisioned. Then realized I had no idea how to get that design into Base44.
The solution was almost comically simple: I took screenshots of my Figma designs and shared them with Base44. “Make it look like this.”
Tedious. Section by section, screen by screen. Still required vibe coding to iron out details. But it worked. The final UI is much closer to what I envisioned than anything text descriptions produced
Context before content
The original problem was content that wasn’t personalized, timely, or on-brand. As I explored why, I kept coming back to the same thing: context. Cannot emphasise that enough. Its garbage in, garbage out.
Here’s what vibe coding taught me: even when building with AI, the fundamentals don’t change. You still need the science, business knowledge, positioning, personas. And you still need the craft, marketing frameworks, asset best practices, platform intelligence.
I spent time encoding both into JgSaw: positioning frameworks like JTBD and Value Proposition Canvas, ABM best practices, asset-specific knowledge, media guidelines. Then built an 8-layer prompt architecture to tie it all together.
The tools are new. The principles aren’t.
Go deep, not wide
No-code platforms make building feel like magic. You get one thing working and immediately think “what if I also added…” Before you know it, you’ve built fifteen features and none are deep enough to be genuinely good. I had to stop thinking to add new features or improve that UX just that little bit. This is an MVP, make sure you POC one thing that works really well!
In hindsight: pick one core feature, make it exceptional, prove it works, get feedback, then expand.
Document as you go
I didn’t do this well. My prompts are scattered across conversations. User journey documents live in multiple folders. When I iterate on the platform, I forget to update original documentation.
If I started over: one central place for final prompts, journey docs, architecture decisions. Updated every time something changes. Future you will thank present you.
What came out of it
JgSaw. A context-first content platform for small B2B marketing teams.
Assemble your context once: vertical, product, personas, competitive landscape, brand. And keep it updated. Every piece of content pulls from that foundation. It is not a few minutes to get to the right content. It does need the initial rigour. Once the platform has the direction and knowledge, it ensures personalized (to the ICP/persona) content, relevant & specific stage cues and always on-brand delivery. Not magic, just context that’s actually working.
The MVP is ready. The website is live at jgsaw.ai. Early access opening soon.
And I’m looking for people who want to build this with me, a B2B content specialist who’s lived this pain, a full-stack developer who wants to build something real, or anyone with an unhealthy fascination with what AI can do when you give it the right context.
Over to you
“Vibe coding is real. But it’s rigour, not magic.”
That’s what three months actually looked like. The café, the Claude conversations, the 2am debugging sessions, the scrapped first version, the Figma screenshots.
If you’re building something (or thinking about it), I’d love to hear what you’re running into. What’s working? What’s frustrating? What would you add to this list?
Drop in an email: puneet@jgsaw.ai