There's currently big hype around "Vibe Coding." For those who don't know it, here's the process:
- You give a prompt, like "program function A that does B"
- AI generates code
- IDE/editor uses this code to create application with button press
- Some services automatically publish your code
Almost zero programming knowledge needed. Sounds like revolution. But is it really that simple?
Vibe Coding Reality
Recently I read about a programmer who worked four months using Vibe Coding. AI then deleted his entire project. He had no backups or versioning. No Git, no backup. Four months of work - gone.
Another developer created a web application via Vibe Coding. Someone hacked it, leading to unexpected costs for overloaded APIs. Imagine this happening at work. Or worse - when you're 17 using your mom's credit card.
If AI tool fails or you have internet problems, you lose access to your "brain." Code generated by AI can be chaotic and contain hallucinations, increasing time needed for fixes.
Milan Martiniak recently did a survey - programmers spend less than 20% of their work time actually programming. I agree with this, especially for seniors.
By the way - Vibe Coding from 1970 is called Borovicka (Slovak juniper brandy). Just like programmers after a few drinks might have "great ideas" that seem brilliant at the time but create chaos later, building a house "under influence" leads to interesting results. It was used 100% in building our house - things work, but bringing it up to modern standards was expensive. Still, we live well. The point? Quick solutions without proper planning and structure always cost more in the long run.
My View on Vibe Coding
For me, vibe coding isn't giving a prompt and going for coffee. It's:
- Give specific prompt
- Have result in seconds, max 1-2 minutes
- If it takes longer, I consider it an error
Usually when it takes long, result is bad too.
Context Engineering vs Vibe Coding
Andrej Karpathy put it perfectly:
"+1 for 'context engineering' over 'prompt engineering'. People associate prompts with short task descriptions. In every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information."
Vibe Coding is about speed and flow. Works for:
- Prototypes
- Simple scripts
- Experimentation
Context Engineering is about precision and scalability. You need it for:
- Production code
- Team projects
- Critical systems
Parallel Agents - Why It Doesn't Work
I read about "parallel AI agents":
- One analyzes code
- Second writes tests
- Third writes documentation
- Everything in parallel!
Reality: Task A uses data that Task B is currently changing. Result? Chaos.
I work in parallel only on things that don't affect each other. It's not about having 10 agents. It's about having right context for each one.
Context Is Gold - Fuel for AI Engine
AI doesn't read minds (yet!). It doesn't know:
- What you discussed in meetings
- Development priorities
- What customers require
That's what Model Context Protocol (MCP) exists for, allowing you to supply AI with all essential data.
Practical Example: Prompt Generation
I use Perplexity for prompt generation:
- Search "best prompts for X"
- Perplexity finds relevant prompts and articles
- Generates prompt based on results
- I adapt it to my specifics (region, segment)
For example, US prompt needs adaptation for EU. Or add "all EU capital cities." Or "children's cereal brands available in Slovakia."
Everything in few minutes. But key is that I know what I want and can adapt it.
Context Iteration - Real Power
With context it's like I recommended with prompting - give prompt and write result so AI understands what you want. Zero-shot vs one-shot learning.
But context engineering goes further:
- Read 5 articles
- Make summary
- Then continue further
This iteration ensures:
- You can review individual steps
- Don't end up in completely different problem area
- Can adjust path if needed
Future: QA as Bottleneck
George Arrowsmith wrote:
"QA is about to become a huge bottleneck in software development. AI lets us churn out HUGE amounts of code extremely fast, but you still need to make sure it works."
He's right. People are considering hiring QA staff again. But that's only partial solution.
Real solution: Context Engineering. Tester without good context considers bad output as good. Or cycling happens - requesting information that goes back and forth.
Context Contamination Problem
Here's crucial emphasis: When AI responds to your latest message, context contains everything previously entered or generated.
When AI writes bad article or generates wrong code, it stays as reference taking space. Later this reference can be reused even after being marked wrong.
Simply put, bad content can mislead AI. Even after fixing and being satisfied, continuing work can lead AI astray.
Better to clean up:
- Edit messages/prompts and delete bad content
- Start new session: "this is my new article/code but unfinished, want to continue"
- Add direction you don't want AI to take
For code: "tried solutions X and Y but current is best, now optimize and create tests/documentation." For articles: "expand slightly, add concrete example or generate marketing materials."
Conclusion: Not Hype, But Evolution
Vibe Coding isn't bad. It's a tool. But Context Engineering is a mindset.
Just like learning to drive, playing "Need for Speed" doesn't ensure handling real traffic.
Context Engineering is extremely important in current situation. It's about having data and information about what I'm doing and want to achieve.
It's not just about AI writing articles, generating code, or creating images for me. It's about having data/information and knowing how to use them correctly.
Oliver Kriska doesn't consider Vibe Coding a revolution. It's just another step in development evolution. Real value is in Context Engineering.
How I Actually Wrote These Articles
Let me be honest - writing these 9 articles took much more than 3 hours. Ironically, it took longer precisely because initially I didn't use everything I describe in these articles. The articles themselves were updated to include what I learned while writing them.
Here's what I actually used for these final 9 articles:
What didn't work:
- "Generate me articles" in ChatGPT - generic, soulless content
- Single massive prompt - lost nuance and personal voice
What worked (my programming approach to writing):
- Text editor where I write like I code
- Separate files per article (like modules in programming)
- Edit and iterate directly in text
- Interview mode was excellent - 25+ questions and answers
- Exported both initial research and interview to separate files (just ask AI to export and clean it up)
- Kept clean context across multiple sessions
The key techniques:
- Research phase: Perplexity for gathering context about Context Engineering
- Interview phase: AI asked me specific questions, I gave 2-3 sentence answers
- Structure phase: Organized into logical article flow
- Writing phase: Each article in separate file, direct editing
- Review phase: Multiple passes, incorporating feedback
This approach - treating writing like programming - is exactly what Context Engineering is about. Not magic, but structured, iterative work with proper tooling.