From Vibe Coding to Context Coding: My AI Journey

I tried vibe coding for 2 weeks. Nearly lost my mind. Don't get me wrong, watching Cursor auto-complete entire functions feels like magic. But here's what nobody talks about After a few days, I had zero idea what was happening in my own codebase. AI was making architectural decisions I didn't understand. Writing functions I couldn't debug. Building features that "worked" but I couldn't explain how. The breaking point? When a simple bug took me 3 hours to fix because I didn't know which AI-generated code was doing what. I realized vibe coding isn't development. It's just expensive copy-pasting. Now I do something different. I call it "context coding with a sprinkle of vibe." First I plan everything. User flows, tech stack, component structure. Then I feed that context to AI and let it build. The difference is night and day. AI becomes a supercharged teammate instead of a mysterious black box. I can debug faster. Add features cleaner. Actually understand what I built. This realization became the entire foundation for what we're building at Syncally. Anyone else felt completely lost in their own AI-generated code?

This is another 'vibe coding is the shit' post, so if that isn't your vibe, scroll away now, you've been warned. I shipped my first code change in years this week on Relevance AI's platform, thanks to Cursor. Not a big change (honestly, a few lines of code), but enough of a change to make a difference. I saw an opportunity to make a change that would direct our customers into a new support channel (psst -- have you joined the Relevance Community yet?), made that change with the help of Cursor (and Matthew Byers and Daniel Palmer) and shipped it. Vibe coding isn't going to replace software engineering teams, but it is going to allow them to focus on work that matters. At Relevance AI, our designers, Product Managers, even Solution Engineers, have been using tools like Cursor and Claude Code to make quick changes to our product, and get them reviewed by our engineering team. It's been invaluable - if you want to learn more about how design in particular has been using it, give Freya Stockman a follow! This is an initiative every support team should be jumping onto too. Gone are the days where support teams would file a small feature request that would sit in the backlog for years, or report an issue with copy in a product that wouldn't get actioned for months. Now, when customers report an issue, support teams can try to fix it with vibe coding too. If you're a support engineer looking to upskill - download Cursor. Try to get access to your product's codebase, and instead of just flagging a feature request, or copy change, find a small one to work on and see if you can create it with vibe coding, then share it with your R&D team. If you're leading strategy / managing your support team - lean into this. This is one of the ways support teams can show value in an AI first world, by not just presenting a problem, but also quickly shipping the solution too. Is vibe coding infallible and going to replace engineering teams? Absolutely not. But, it is a tool that your support teams should adopt? Absolutely yes.

Ok you vibe coding pros - here’s a total life saver. One of the BIGGEST PAINS IN THE 🍑 is AI constantly hard coding colours, styles and so on. And it’s very hard to catch them all. But you can force AI to fix them before it is able to commit the code to GitHub. Welcome to the wonderful world of hooks. What’s a hook I hear you cry? Basically, it’s a trigger. One action triggers another. In this example: “Before you commit, you must fix all hard coded values. If you don’t, NO COMMIT.” Ask AI to do this: “Use Husky to add a pre-commit hook to ensure all hard coded values are changed to constants as defined in [wherever your types are defined]” A pre-commit file will be generated which kicks in before any commit. Mine has a bunch of other hooks, like ensuring code and database schema are aligned. Hooks will change your life. If you’re a sad vibe coding obsessive like me. The rest of the world is more sensible.

“There’s a new kind of coding I call #VibeCoding, where you fully give in to the vibes… and forget that the code even exists.” — Karpathy He later clarified: vibe coding ≠ AI-assisted coding he professionally cares about. The latter means you properly read, review, change, test, and understand. (Simon Willison also summarized this nicely. reference in comments) Takeaway: don’t conflate playful speed runs with responsible engineering. Here’s how to avoid bad practices in vibe coding: 1. Pure Vibe Coding: Use it for prototypes, and weekend hacks, not production features. 2. Don’t “no-read, just-ship.” Always review code you’ll keep. 3. Ground the model with fresh docs and your entire codebase; don’t trust hallucinated APIs, variables, etc.. Context is 👑 4. Never paste secrets; start with .env from day one. 5. Add a quick smoke test for every integration. Actually TDDing was made more feasible thanks to AI. 6. Set billing caps/limits; avoid surprise charges or unwanted usage. 7. Don’t vibe-code essentials like auth/payments. Use proper IDs + server checks. 8. Prepare a highly detailed prompt, or prefer a simple stack; complexity kills the vibe if no grounding and instructions were provided. 9. Write a short product brief to stop endless scope drift. 10. Do Vibe Checking/Reviewing as well.

I totally agree with every word Itamar says. Except we all know developers are lazy. If you can finish a job in 20% of the time, and deliver 80% of the value - you will. Especially when the one reviewing your code is also using AI. Let's face it - the future is about making the autonomous agents better, not holding their hands. Current coding agents lack eyes, fingers, independence, and more proactive communication skills. Once that is accomplished, they will be autonomous, we will trust them, and we will have a lot of AI driven, AI reviewed, AI tested, AI verified code in production. Like 98% of it. Hopefully, we could still set the rules and frameworks of how code is deployed, when it's deployed, which code has to go through human validation, etc. So our job will be to moderate and govern the AI

I Took a Big Risk: I Asked 30 People to Pay Me to Watch Me Build From Scratch with AI.... Yesterday we had the first session of my masterclass VIBE CODING TO PRODUCTION. The class is all about teaching participants how to vibe code like a professional developer using the Cursor IDE; taking an idea to code and deploying it live in just 2 days. But here’s the twist: I could have picked an app idea weeks before and rehearsed everything to avoid any “story that touch,” but I decided we’d pick the idea on the call during Day 1. Yes, I walked into the class with no idea what we were building until the class voted live. If you’ve ever tried real "vibe coding"—not just a landing page but a full-blown app with database, backend, frontend, user authentication etc, then you know it can get frustrating, and I promised we’d do it in under 8 hours. That’s a big risk. I only had the confidence to do this not because I blindly trust Cursor or the LLMs behind it, but because I trust my creativity, my context-engineering skills, and years of building software without AI. To make it more interesting, the participants chose to build an online booking system with calendar features. 😂 These people wan kill me, but I love it! Day 1 recap: Wrote a very detailed PRD with AI Built the project foundation Completed the landing page Implemented user authentication Now the real question: will this end well? Stay tuned.

Vibe coding is everywhere right now. The idea is simple. You let AI generate most of the code and then you guide and patch things as you go. It feels fast, fun, even magical at times. But is it really the future of building software? For seniors, it can speed up prototyping like crazy. You sketch an idea and have a demo in hours. But the same seniors admit they spend just as much time fixing what the AI wrote. Bugs, weird design choices, security holes, all still need cleanup. For juniors, I think it is even riskier. If you lean on AI too early, you stop learning the fundamentals. You risk becoming an AI babysitter instead of a real builder. Security is the other red flag. The New Stack called vibe coding a “productivity boom hiding massive governance risks.” Shipping unreviewed AI code is basically asking for hidden vulnerabilities and bad dependencies. So what is the right way to use it? - Great for prototypes and hackathons - Always with a senior in the loop - Treat it like an intern, never ship unreviewed - Guardrails like tests and code reviews are non negotiable Top tools in 2025: Lovable, GitHub Copilot X, Cursor IDE, Claude Code, and Replit AI. I like the speed of vibe coding, but I still see it as something you watch closely. Kind of like handing a smart teenager your car keys. Would you trust vibe coding for production today, or only for prototyping?

Hand-Code vs Power-Code Even with all its nuances, limitations, and hallucinations—the new age GenAI powered coding tools help me code better. I have used the tools on new as well as on brown-field projects, on prototypes and on production code. I see the value, and if you're still apprehensive, I suggest dropping those inhibitions. I took baby steps with iterative prompting and vibe coding and am currently experimenting with more features—and boy, the features are coming at a faster rate than I can handle. On one hand, I’ve seen these tools generate irrelevant code, bad code, too much code, incorrect code—and sometimes even incomprehensible code. I’ve seen them break existing code and give incorrect suggestions. On the other hand, I saw excellent suggestions, better design patterns, and better structures. I find that traditional best practices—like static code analyzers—when combined with newer paradigms like prompt files, are enough to make the whole experience joyful. With features like planning mode, I can get the tools to execute the plan step by step, under my control. The biggest advantage I see is that I can try different coding ideas and patterns really quickly. What I like most though is that I don’t need to remember all the syntax. I can focus on features and patterns. With vibe coding, it sometimes feels like the tool is reading my mind. At times I got carried away, and even typed "Sorry, I made a mistake." I still write things like "please ...", and I haven't got started with voice prompting yet. “In the long run we are all dead,” as Keynes said; but until then, why use handlooms when you have powerlooms—and why hand-code when you have power-coding?