Vibe coding is a catchy buzzword. What it means is AI-assisted software creation, from simple assistants to autonomous AI agents modifying the codebase on your behalf. They do the work, and you own the results, get the credit or blame.
Intro
I haven’t been in hard-core production coding for a decade. Frameworks appeared, processes changed, and what was the dream back then became the norm today. As a manager, I stayed at a distance, not getting practical experience. But the pressure to engage piled up. I slept on frontend frameworks longer than a sleeping beauty.
When I woke up, instead of horses and carriages, the cars were flying around.
Yet with the advent of AI-assisted coding, I started a new project on my own. I needed to buckle up and ride that horse again. That was around the time when OpenAI just published ChatGPT so I had my superb knowledgeable assistant to help me out exploring further. I built a decent and functional web application written in Vue and Tailwind CSS, the tech stack I understood from a high level, but never used before. And that brought in external funding and four teammates.
The point is, we can be brave and ambitious, as we have super-hero help next to our fingertips.
👉 We’re all cyborgs now!
Prototyping With AI – Build the Right Thing
AI tools are so fast and capable, and to skip the prototyping is a colossal waste of opportunity. Invest some time experimenting with why, for whom, and what to build, prior to jumping to execution. And return to it often.
Start with what you can handle. Developers may jump right into IDEs such as VS Code, IntelliJ or Cursor, whereas non-technical folks may turn to Lovable or Replit.
But then, do not delegate it all. If you’re looking for a worker while you dip your toes in sand, you’ll end up with toes in sand and no work done.
👉 You can’t delegate what your startup needs from you: the passion, drive and perseverance.
On deciding what you will be building, allow yourself to be bold and ambitious. Quite often, I wouldn’t dare to think what’s possible and would censor myself. Don’t fall into the same trap. AI agents are incredibly fast and have seen more of the world than anyone else on the planet, so ask it to come up with several very different solutions, and involve it in the ideation process.
At this moment, our prototype for the latest product has 5 versions of activity display, 3 tabs for displaying agenda, just to name a few experiments. Just add “give me 3 versions that I can choose from”, and use the LLMs’ cognitive load capabilities.
But also, get your fellow teammate in front of the screen. Vibe code together, but don’t get entangled in details. Along with that, don’t throw away pen and paper, whiteboard or a cup of coffee just yet.
All in all
- Use AI tools to explore the problem space and try different solutions
- Don’t delegate or fake passion and drive. Having AI suggest 100 use-cases is nice, but you may end up overwhelmed and discouraged instead of passionate and full of energy
- Be bold and brave
From Prototype to Product – You’re a Manager Now
It’s time to start building the “real” software: manageable, maintainable, bug-free, and reliable. One that your future biological and digital coworkers will be able to modify.
With current-day tools, we all have our super-smart, confident and enthusiastic minions on our side. Well, they also behave like ones, and we need to manage them. They hallucinate, lie, sometimes are plain stupid, get us wrong, and the list goes on. Or they follow our instructions, well, the way we wrote them. Would you trust a monkey with a loaded machine gun?
👉 So, congratulations! You’re a manager now.
There are multiple aspects you, as a good manager, and an architect, and senior engineer, must decide upon. Tech stack, modularity, processes, testing. Books and books are written on each of those topics, so let’s just mention them.
Tech Stack
Decide on a tech stack that is well-known, battle-tested, and that the LLMs had huge volumes to learn from. If you want a well-functioning team of autonomous AI agents typing your code, make sure they have competencies and tools. That means well-known technology with plenty of examples and best practices available online.
Keep Things Small
Technically speaking, keep well within the context window, and not only for technical reasons. Vibe coding, as of now, can only go so far. We can’t yet throw a gigantic task on our agents and expect working stuff. Work in small bits, and iteratively. Commit code often, in small-enough bits of code that you can digest and check.
Modularity
Besides keeping the sessions small enough, ensure modularity. Architect the reasonably sized modules, with well-defined responsibilities and boundaries, well documented. The modularity is a key to productivity and reliability. But be wise, don’t slice too thin and end up with a distributed monolith.
Testing
Now, you’ll be driving at 10x speed. And you don’t have 10x reflexes. You need to automate whatever you can, and start with testing automation. TDD is great to systematically articulate the product requirements and ensure it’s working as intended.
Testing, as frequent and thorough as possible, is even more important now with minions at the keyboard. And watch out, they may solve the failing tests by deleting them!
Be vigilant with code commits, and don’t fall for confident claims. My agent on one occasion insisted that all the tests are now passing, when I found it not executing them at all. I realized because the page wouldn’t load, just stuck dead and with my agent’s claiming “everything is working, refactored, and cleaned”. Well, not on my computer.
A Few Other Things to Consider
- Embed within the project as much as reasonable: documentation, agent rules, tests, etc.
- Articulate and share agent’s rules, including what’s acceptable, what’s not, and instructions to run tests on every change.
- Pay attention to branching strategy to avoid collisions and improve clarity.
- Watch out for your own cognitive load as you’ll be validating the PRs.
- Save and share templates (of prompts and any other temporary stuff you need) and work with your human coworkers on improving how you work.
All in all, treat your AI coworkers as biological ones.
Ask them to explain, review, rework. Make them move in small-enough steps that you can follow through and catch their mistakes.
One more thing to do with AI-assisted coding is what it means for your coworkers. How do you hire? What do you look for? Personally, I want a curious person that is open to new technologies, looks at a broader picture, as we’ll focus more on WHAT on top of HOW, and a very responsible person that will ensure the minions don’t open fire. They say “nobody hires juniors anymore”, but I think we’re looking at a bigger change than seniority.
To Conclude
AI tools are valuable in every aspect of our professional and private lives. They have great mental capacity, bravery, and they don’t give up where we would mentally give up.
Use those to your advantage. Don’t skip something that would drain you. It won’t drain the AI system. Use it with care and curiosity, but not blindly.
All in all, a few key points to take home:
- Treat your AI as a more confident, capable, knowledgeable, and persistent coworker.
- But also treat them as you would monkeys with machine guns.
- For software delivery, lay the groundwork by selecting an appropriate tech stack and architecture, and ensuring boundaries and standards.
- You own all the results — good or bad — it’s all your responsibility.
What is your experience? I’d be glad to connect and learn from yours too.
