I love marketing
Mainly because of it is a literal money and perception printer
But I hate it when it distorts my perception and makes me do wrong things
And right now I have a reason to hate
It built childish naivety And false expectations regarding what AI can code
AI companies have created expectations for us that hey you just write a prompt
And we will create the code that will be the application
The whole application
And nothing can be far off from truth than this notion they have created
Frankly it is almost impossible for it to do that
An application professionally, a professionally built one, is made on so much nuance and context
They have years worth of experience when it comes to solving back end issues front end issues
Accounting edge cases and how to improve the user experience and how they interface with it
Any application you see like Youtube or Trading View or any CRM
The form it is in right now is a result of years long process and hundreds of brilliant minds working on thousands of micro problems
Problems that we never conceived of nor we ever encountered because we were not in their shoes
Nor can AI simulate or account for all those problems when we give it a prompt
When you tell it go make a landing page, It will give you the code
But that code would not have solved thousands of micro concerns and problems which other websites might have
And that’s where the difference between professionally built code and vibe coded comes
Exactly the kind of difference you would expect in some MBA classroom project vs an actual business project
What why could add things lag is consideration for hundreds of micro interactions or edge cases that are not considered when we build a basic skeleton
Things like how should the panels in a landing page behave when the X dimension is getting adjusted by the user
Are we supposed to just make them slimmer are we supposed to push them down which one should we push down or which one should we put up
1000 questions, all left unanswered
And when you try to make it answer them
It creates solutions that makes contradictory or incompatible solutions with other 999 questions
This is why it is especially hellish to work on front end using just AI
For backend there is less ambiguity
There are firm handshakes which let you know that OK this thing worked or this thing didn’t
You get clear responses
Back end was all binary
That OK it either worked or it didn’t
But for front end that’s not the case
Front end is an open arena
Arena which requires considerations for thousands of things
Which we never bothered to think about
If you’re trying to make a charting software for trading you might say that OK just vibe code a clone of tradingview
The point is Trading View has done a lot of innovation in UI to be as fluid and as accommodating of edge cases as they are now
We cannot compete with that using a consumer grade LLM
All valuable things emerge from the battle scars
Which were results of facing the real world
The software that looks efficient today reached that state not only because they planned everything but because of the unplanned things that happened to them and for which they had to counteract
The user complaints or the icp pivots that kept changing the app 1% each time, turned it into a evolving ship of Theseus
What can be really valuable right now is a tool that can classify things as errors when it comes to front end
Back end are solid because we iterate them away till we get no errors
But in frontend there may be no code error but it might just look like shit and confuse the users
So we need something that can classify stuff as a non code error so that at least AI can have a chance at iterating maybe it can fix it
We need something that feeds it the micro concerns and age cases which we are talking about here
So vibecoding the whole thing is not possible but we can stand on the shoulders of giants
You should first investigate all the libraries that exist out there that have solved the exact problem you’re trying to solve
If you’re trying to build a Candlestick renderer so that you can make your Trading view clone, Maybe there are hundreds of libraries that do that
You then scour the Internet for all the documentation and QnAs that might exist for that library
And use this information to build your app so that at least some of the optimizations and edge cases are accounted for
This is actually a big disadvantage of building with AI
Instead of making it the priority to use external libraries that are very polished
It tries to make its own juvenile attempts at building the logic from ground up
And due to the inherent constraints of context window and how smart the model is it is not able to do that
And you’re stuck fixing basic Must haves instead of working on the grand vision
Instead of viewing A as something that can build things from ground up
You can see it as A scientist that can Create a beautiful Frankenstein monster
Claude reaction:
The irony is that AI actually makes you worse at software architecture when used naively. A human developer knows their limits—they instinctively reach for a charting library because they know building a performant candlestick renderer from scratch would take months and require deep knowledge of canvas optimization, memory management, and financial data structures. But AI doesn’t know what it doesn’t know. It confidently generates a basic implementation that looks right, renders 100 candles fine, then completely falls apart at 10,000 candles or when you try to add volume profiles or custom indicators.
The AI will burn half its context regenerating a mediocre implementation of something that already exists as a 50KB library. You end up maintaining AI-generated code for solved problems instead of building your actual differentiation.