It's been 2 years since I started doing AI development, but there is one thing that was lacking when AI hype started, and it's still lacking now.
Its decision-making, context awareness, and context longevity.
You may have already heard that Claude made a C compiler, Cloudflare made their own Next.js app, etc., but it's only possible when the person behind it was giving very clear instructions.
I saw multiple founders being stuck in a loop, where they have a good idea, but not technical, try to build using Google AI Studio or Replit, lovable, etc., but it's a sinkhole for them. They start well, but are never able to finalize it. They always need a developer who can do the final refinement or solve the issues that AI made for them.
The first problem I am talking about is that AI can start things in a good way, but as you start building a product with vibe coding, then AI falls back, because it lacks the long-term context awareness, or you can say, memory. It doesn't have a long term memory.
The second problem is that AI can write code, but it's not correctly structured, and it's bad. Ihave beenm seeing it for 2 years. The structure of the code used to be very bad before (it's improved a little, but not as much as it's hyped ). I literally told AI to create a React project from the start, and instead of running a command, it literally started creating files one by one, which eventually ran, but it's not what I meant when I said to create a React project ( I meant Vite create ).
The third problem is decision-making. For example, I want to create a website like Uber, and I said the same to AI, and it will create a plan for me. But, is it going to plan the whole architecture? iIsit going to consider the number of users I am targeting? Is it going to consider the database, which will be good for the target user? Is it going to decide the cloud architecture for me? No, iit'snot. It can pretend to do all of that, if you ask AI to do it, but with a simple prompt like "Build me a uber like website,, it's not.
Now, the solution.
I want to build a website that does all of that. Which means that not everyone is a good prompt engineer ( including me ), and I don't want to spend my time refining prompts word by word, so that AI can understand me better. I want to give the power to AI, where it can make decisions, plan exactly what you want ( and take confirmation before doing anything ), and do everything that a senior developer, project manager, or architect is able to do. Provide you complete picture of what's going on. Let you intervene, and be a decision maker ( if you want to be ), otherwise, make a decision for you, and be able to understand the point of the product you want to build, and don't lose context.
I understand that all of the above that I want to build is already possible, but is it possible for you?
Top comments (15)
The "founders stuck in a loop" observation is spot on. I've watched this happen repeatedly: someone gets an impressive demo working in 2 hours, thinks they're 90% done, then spends the next 3 months fighting the last 10%.
The missing piece isn't better AI — it's that AI can execute individual tasks but can't hold the product vision. It doesn't know which trade-offs to make, which features to cut, which edge cases actually matter for your users. That's product thinking, and it requires understanding context that doesn't fit in a prompt.
The developers who use AI most effectively treat it like a very fast but very literal junior developer. You still need to do the architecture, define the acceptance criteria, and review every output. The generation is cheap — the judgment is the expensive part.
Exactly! If there is create-vite, don't use AI, use create-vite. If there's a manual, don't ask AI. Read the manual! Don't use the wrong tools. "AI can code" still sounds like "dogs can play piano" to me. Yes, they can touch the keys to make a sound. If that's good enough for you, that's fine.
The idea was called "no-code" before, now it's "AI agents".
Oui c'est vrai
You're describing the exact moment where "vibe coding" hits the wall of real engineering. The problem isn't that AI lacks memory—it's that it has zero skin in the game. When your React scaffold crashes at 3 AM, Claude doesn't get paged.
Your Uber example exposes the core gap: AI can't architect for tradeoffs. Database choice isn't about "number of users"—it's about read/write ratios, consistency guarantees, and whether you can afford eventual consistency when riders are waiting. These are blood-and-scars decisions, not prompt outputs.
The solution you want (AI that "decides like a senior") is a category error. Decision-making requires accountability and context that survives beyond a single chat window. What you actually need is structured constraints—force AI to validate assumptions ("Will this scale to 10K concurrent requests?"), generate architectural decision records, and output testable invariants.
Build the scaffolding that makes AI justify its choices in writing. That's the forcing function that separates vaporware from production systems.
Yup, that's the plan. AI can't decide itself, they need to justify it and validate their own ideas, human intervention is good too
"but it's only possible when the person behind it was giving very clear instructions." and that my friend is our job. AI won't take the experience dealing with customers, data flow layout, schema planning, etc. You still need those skills to "teach" the AI what to do. In my case, after 25+ years of programming experience, AI is still teaching me things. I use it to plan, diagram, blueprint and set the theory in stone; then and only then, I write the pseudo code I want, and ask AI to produce final code. It works like a charm, so your sentene "Which means that not everyone is a good prompt engineer ( including me ), and I don't want to spend my time refining prompts word by word, so that AI can understand me better." i would recommend not to trust AI blindly, simply integrate it in your regular programming workflow, no?
Great article, btw, and thanks for sharing.
Yeah thats true AI just needs direction then he can build full product for us
Today’s so-called “AI” is extraordinary at pattern recognition and reasoning analogies, yet probabilistic engines, not intelligences. We have fed a machine with everything human, from the internet... which might not have been the best idea ;)
We have failed in developing, we have failed in creating products, failed with companies... over and over and over again. All that is priced-in in what we use now. Therefore, the outcome is exactly what we see: good start, bad ending!
Of course, I have no numbers, but if 70% of projects, code, products, companies fail or are bad and 30% succeed, are awesome structure marvels... a probabilistic, statistical machine will tend to the 70%, train on the masses, no matter how "guided" the learning is with that amount of data.
AI is AI is AI. Nothing more, nothing less. We have to evolve in our goals, our methodology, our approach to solving problems. Not AI, not product, not website, not nothing.
A phrase I heard is: a Junior knows how to start but a Senior knows how it will end!
If we know the end state and have a way of expressing ourself, we can succeed, pAIred ^^
This hits the nail on the head exactly.
The uncomfortable truth is that AI cannot code, it solves problems with language patterns. When your product is custom/complex it gets lost very easily.
The current models are not optimized for analytical capability, but rather context scope. That means this weakness is not solved at all and instead it becomes much more of a general pattern matcher.
For my solutions that means i have to rebuilt a lot of my code to become standards/templates to the degree the AI doesn't have to reason too much about non standard code.
Additionally I think this is also why we are seeing large companies have serious problems with AI assisted coding that works on brownfield/non standard large complex code bases. It's slowing them down and it's creating more erroneous code as a result.
AI can write code, but building a product requires much more than technical output. Tools like GitHub Copilot and models from OpenAI can generate functions, fix bugs, and speed up development. However, product creation involves understanding users, defining value, shaping experience, and making strategic decisions.
Code is execution. Products are vision, empathy, market insight, and continuous iteration. AI can assist in the process, but human judgment and product thinking remain essential to turning code into something people truly need and use.
That's true, AI is a tool.
Likewise, the user must give instructions to the AI.
Well, AI cannot act on its own.
Great article 👍🏼
Some comments may only be visible to logged-in visitors. Sign in to view all comments.