Let me start like this: you are here because you’re wondering whether you’re using AI the right way. good for you. You’re probably not.
The good news, though, is that there is no universal “best way” to use AI. It depends entirely on your skill level. I know that’s disappointing.
You listened to Dario, Jensen, and half of Claude opus tweet, and now you’re convinced claude is going to transform you into the next Bjarne Stroustrup before lunch, It won’t.
The uncomfortable truth is that AI can only amplify what’s already there.
If you’re a beginner, AI will mostly give you beginner-level code. The funny part? You probably won’t notice. To you, AI looks like a senior engineer because you’re comparing it against your current skill level, everything looks advanced when you’re new.
People buy an Opus subscription and suddenly start talking like they’re about to revolutionize software engineering. That’s like getting a driving permit and now you think the only difference between you and Lewis Hamilton is networking and connections he has.
Take a simple example.
A beginner asks AI: Think like jeff bezos and build me an e-commerce application where users can place orders, check out and pay me make it unhackable and have too much security and make it faster. AI happily generates 4,000 lines of code. The beginner celebrates, LinkedIn gets another “I built Amazon in a weekend” post, Life is good.
Meanwhile, a senior engineer asks something completely different. Create a confirmCheckout() function that takes a cartId and userId validates that the cart still belongs to the user, recalculates the total from the database, creates an order inside a transaction, reserves inventory for each item, and publishes an order.created event to Kafka after the transaction succeeds.
Notice the difference?
The senior engineer isn’t better because they tell AI to think as senior engineer at amazon. They’re better because they know what exists. You cannot ask for something you don’t know exists.
Nobody wakes up one morning and says: You know what this application needs? Distributed event processing. If they’ve never heard of distributed event processing.
I always laugh everytime i see someone posting a 20 page prompt to transform the whole industry and its all about:
Be careful.
Double check everything.
Think harder.
Think even harder.
Think like a senior engineer.
Think like a principal engineer.
Think like the primegen at netflix.
Think like Linus Torvalds in his prime.
debug like tsoding
At this point you are praying and hoping for the best.
The best engineers I’ve seen don’t use AI by dumping an entire application into it and hoping for the best. They use it the same way they would use a junior developer. One task at a time authentication, validation, Database schema, Dockerfiles, Background workers, API handlers, One piece after another. The AI isn’t building the system, the engineer is.
For example, I don’t need to tell AI: Read my entire codebase, analyze every file, understand my business logic, discover my purpose in life, and then create the perfect application. I can simply say: Create a function using Drizzle that fetches users, joins products, joins reviews, and returns the result or Create a Kafka publisher for an item_added_to_cart event. See small tasks clear objective, easy to review, easy to debug, easy to understand. And most importantly, I still remember what was built because i was actually involved.
One thing nobody wants to admit is that AI doesn’t magically remove the need for skills. It actually exposes the lack of them. When something breaks, the person who understands databases fixes it, the person who understands distributed systems fixes it. the person who understands networking fixes it. The person who only knows how to paste prompts into a chatbot starts another chat and types: “Why isn’t this working?”.
Then I see people looking at senior engineers shipping incredible products with AI and asking: “Am I using AI wrong?”. Not necessarily. They just know things you don’t know yet. Thats it. There is no secret prompt, There is no hidden AGENTS.md ritual, There is no sacred YAML file blessed by the AI gods. They’re simply operating with more knowledge.
And before somebody quotes Dario, AI is not a replacement for learning. its a multiplier, A multiplier applied to zero is still zero. The fastest way to improve your AI output isn’t finding a better model. It’s becoming a better engineer. Learn database, learn networking, learn how large systems actually work. Then watch how dramatically your AI results improve.
Stop treating every AI prediction from tech executives as gospel. Just remember the people telling you AI will replace every engineer on Earth are often the same people raising billions of dollars to sell AI. I m not saying they are wrong, But if a car salesman tells me I need a car, I at least check whether he’s standing next to a dealership first.
Top comments (14)
This is a spot-on perspective. I really like how you highlight that AI is a multiplier, not a replacement, and that senior engineers use it like a skilled junior developer, breaking work into small, verifiable tasks rather than dumping an entire system into the model.
The distinction between knowing what exists versus asking AI to generate abstractions you haven’t internalized is critical. AI amplifies your existing knowledge, so the key to better results is becoming a better engineer—understanding databases, distributed systems, networking, and large-scale architecture.
I’d love to collaborate or exchange insights on structured AI-assisted workflows, particularly for distributed systems or event-driven architectures. Sharing strategies for task-level prompt design, verification loops, and incremental AI-driven development could help teams maximize productivity while maintaining reliability.
Would you be interested in exploring a joint discussion or pilot workflow around this approach?
thank you so much, i would love to explore a joint discussion around the aproach
I am too. would u share any one?
Can't build what you don't know. That's why fundamentals goes harder today than ever before. Learn it, build it yourself, understand the tradeoffs. After you've built enough to the point that you actually know what you are doing, that would be a good indicator to integrate AI in your workflow for "that" specific task. Work smart but don't cut corners.
Everyone things he has done his best due to a fact that nobody knows the fundementals anymore nobody care about the characteristics of a good software anymore they just prompt without any fundamental of why that thing is there ,they dont even know what the output should look like so every output is great to them
The senior/junior gap isn't about knowing more prompts — it's about knowing when the AI is confidently wrong. Seniors spend their context budget on verification, not generation.
Most people are relying on AI generating the code to an extinct nobody has the identity anymore. every codes and every architecture are starting to look the same ,same issues ,same performance and people are not making improvment to software anymore due to this AI depedency.
The senior difference is usually where the AI is placed in the workflow. Less experienced engineers often ask it to produce the answer; senior engineers use it to widen options, attack assumptions, draft disposable code, or explain unfamiliar surface area.
The strongest use is still paired with taste and constraints: what problem are we solving, what can break, what should not be abstracted yet, and how will we know the result is actually better?
exactly people should always focus on the last question "how will we know the result is actually the best we can get" and nowadays people has outsourced all those works to their AI it does the thinking and the review for them
Exactly. The dangerous pattern is letting the same system both generate the work and grade the work without any outside evidence. I still like AI in the loop, but the final check needs something concrete: tests, logs, user behavior, a diff, or a human acceptance criterion that was written before the answer existed.
One of the biggest misconceptions about AI is that it replaces engineering knowledge. In practice, it often does the opposite: it reveals exactly where knowledge is missing.
A junior engineer sees generated code and asks, “Does it run?”
A senior engineer asks, “What assumptions does this code make about transactions, consistency, failure recovery, concurrency, security, and observability?”
The difference is not prompting skill. The difference is having a mental model of how systems behave when things go wrong.
AI can generate implementation. It cannot generate experience. And experience is what tells you where to look when the implementation inevitably fails.
No one care about software anymore , reviewing codes and putting effort to having better codes is vanishing slowly by slowly nowadays everyone things they are seniors as long as they paid their subscription. people should understand that seniors are made from reviewing codes and understanding why things are happening
Experience is still the ultimate debugging tool. AI can accelerate implementation and even accelerate learning, but it cannot compress years of production failures into a prompt. The engineer who has seen race conditions, data corruption, outages, scaling bottlenecks and security incidents will always evaluate AI-generated code differently from someone seeing the pattern for the first time.
Agree it amplifies rather than levels you up. The output is only as good as the judgement pointed at it, and the senior edge is mostly having a standard to hold the result against and actually reading it before it lands. A beginner can't tell a plausible diff from a correct one, so the amplification just runs wherever the model went, fast.