I No Longer Write or Read Code. AI Does It Better Than Me

I decided to write this post because a lot has changed in my workflow during December and early January. Despite the absence of any visible revolutions in the industry that everyone would be talking about loudly, a turning point quietly occurred with the release of GPT-5.2 and Opus 4.5, when the models crossed some threshold that increased the level of my trust in their work so much that it made me completely irrelevant as a programmer.

What Changed in December

Let’s look at 3 examples of improvements that occurred in the models in terms of programming:

1. Autonomous Runtime

With the release of GPT-5.2 Codex, they raised the bar for the amount of time Codex can work. For me, Codex solves tasks for half an hour, an hour, or more, autonomously and without my participation. Already with the release of 5.2, it’s capable of editing and creating dozens of new files from A to Z, without stopping, without interruptions, and without requiring my attention. Codex no longer says that it’s running out of time or tokens, no longer rushes anywhere, doesn’t check in with me every five minutes. It just works until it’s done.

2. Prompt adherence

A seemingly minor change in the model’s behavior brought huge results. Codex now tries to perform its work from the beginning (which in most cases means reading files, gathering context, researching, viewing examples) to the very end - that is, fixing any build errors and even testing when it makes sense. If with 5.1 you could still expect Codex to say “yeah, I couldn’t do it” or “I did half, you finish the second half,” that doesn’t happen anymore. Now a one-line instruction is enough for codex to follow a rule perfectly and always, e.g. “use ksrc to read Kotlin sources.”

3. Context Rot

Work quality no longer degrades with the shrinking of the context window - this is a huge and important improvement that’s precisely what allowed GPT-5.2 to work for such a long time. It doesn’t start cutting corners, getting confused, and hallucinating like any models used to do before. It doesn’t “get tired,” doesn’t stop, and doesn’t get distracted.

The Moment of Realization

A transition occurred across a boundary where it became possible to completely let go of the reins of the LLM, and the model’s intelligence exceeded the threshold of my expertise. I should have made this post back in mid-December, I just spent a lot of time coming to terms with the fact that I became completely redundant.

As a programmer, I’m no longer needed, and I only get in the way of the agents working. I mess up their work, and it comes out much better if I just don’t interfere with them.

It was very hard for me to accept this. I resisted for a long time, but somewhere in early January, when I returned to programming with twice the intensity, on new projects, after being laid off from work, I realized: when Codex and I work together, most of my suggestions, comments, remarks, and requirements are mistaken and wrong.

Situations like this started repeating more and more often:

I tell it to implement a feature → it researches my codebase, implements it in a certain way → I decide that I need to review the code, start reading it, find something I don’t like → tell Codex: “No, this is wrong, it should be like this” → Codex responds: “No, you’re wrong, I did this intentionally for reasons A, B, C” → I realize that I was wrong.

I didn’t have enough context in the codebase that Codex has, which it gathered from reading dozens of files. I forgot something, I mixed something up, and it simply used domain expertise from an area where I lack knowledge.

I see that I’m hitting my basic human limitations: I forgot something, I mixed something up, I’m tired, I skipped a line of code, I rushed, I don’t know some framework, I don’t fully know the nuances of working with databases or design patterns for high-load systems.

Examples of My Mistakes

When I was working on the backend, Codex implemented a database structure that was fully normalized, and I started arguing with it and demanding a denormalized structure, and it had to correct me.

Or it thought through the nuances of working with our abstraction over providers and models, and I forgot that we have a layer in the form of a message parser that reorders them. I pushed back on Codex, and it had to explain that I was wrong.

Sometimes it happens that it doesn’t argue, just does what it’s told, and then I start reviewing the code again or find a bug and realize that I created that bug myself by requesting something from it that doesn’t make sense or is wrong.

There were more and more of these situations, and I had fewer and fewer comments that were actually valid for the code that AI models write. And at one point I just gave up and stopped reading code. It’s counterproductive. Codex reviews itself several times for each feature and makes suggestions that I don’t notice during review, and moreover, doesn’t make stupid wrong remarks that would make everything worse - the kind that I make.

Confession

I know AI skeptics will hate this post and the first thing I’ll hear will be “skill issue,” but I’m just trying to be honest and transparent. And the truth is that I push code to prod that I never wrote and never read. I don’t even know what this code does.

Not because I’m lazy or a slop generator, or a low-skill vibecoder. No, I have six and a half thousand hours of programming experience. I’m working on Respawn, where code quality is very important to me. This isn’t someone else’s project where code quality and maintainability don’t matter. This is my magnum opus. And yet, the latest Respawn update was written by Codex from A to Z.

And it’s all because I see that I’m only getting in the way of AI with my presence. I’m no longer needed. When I review code, when I carefully check it, this results in more errors and bugs than if I didn’t look at it at all.

It took me a long time to come to terms with this. I’m still sad that I spent 6 thousand hours becoming an expert in programming and learning so many different things. I got a degree in CS, learned dozens of frameworks, worked on such a mountain of the most diverse projects, and now all of this is worth nothing. All this knowledge is no longer needed. It’s an antique in the modern world.

My New Development Process

All my work now comes down to simply recording voice messages about what I need to do. And even this I’m not sure is right, because if Codex had the same wrapper for running a business as it has for programming now, it would make some business decisions better than me too.

I open the terminal, type codex and record through 4o-transcribe:

“Hi, I want to migrate from the inline date picker that’s currently used on all screens in the app where there’s a choice of habit or ritual repetition, to the new dialog that we made for the habit settings screen.”

Twenty minutes later I have a thousand-line diff, a complete migration is done, everything builds, everything works, the code was reviewed and works perfectly on the first try.

Another prompt example:

“I want to make an integration in Respawn for exporting current rituals to Google Calendar and Apple Calendar. You’ll handle the backend work, and I’ll handle the frontend work. Implement an endpoint that will return an ICS file with the user’s rituals by private token, which can be instantly imported into any online calendar. Use popular libraries, don’t parse the file manually, and after completing the work make sure the endpoint works and returns the correct date.”

I lied to Codex, of course, because the UI for this calendar integration was being built by another clone of Codex, just in a different repository. In an hour and twenty minutes I had a feature ready that I would have spent 2 weeks on a year ago.

Codex wrote over 15 thousand lines during the latest Respawn update. And every time I sat down to work with it, I accomplished in one day the volume of work that a year ago would have taken me a month and a half.

What Became Obsolete

My prompting skills and all these videos I recorded with super hardcore hooks, planners, mega-wrappers, super-plugins, became completely obsolete and irrelevant. I’m even thinking about deleting my YouTube videos because in the first two, practically everything is already wrong.

You no longer need any prompting from the first two videos, you no longer need any hooks and wrappers. I no longer use sub-agents. All I do is run 3-5 parallel copies of Codex, dictate to them 3-4 lines about what I need, at such a high level that I don’t mention technical details at all. In half an hour I get deeply integrated into the current codebase, tested, building, and clean code with good architecture.

I set up clawd.bot for myself the other day and now use it as an orchestrator of agents - I just ask it to manage all these copies of Codex, I write to it in Telegram about what I need to do while riding the bus, and I get a mergeable pull request as output.

What’s Next?

The level of existential crisis I’m in right now is hard to put into words, because HOW to do something became unimportant. The only thing that matters now is WHAT to do. And WHAT to do, no LLM can still tell me.

Honestly, I think hardly anyone knows what to do now, when one of the highest-paid areas of labor in the world has dropped in value to practically zero. The price of code has never been lower, when you can make a complete application from A to Z in a day. Codex will write tens of thousands of lines for any request. Writing code manually is now counterproductive.

What AI Still Can’t Do

Codex doesn’t have all the human flaws, but at the same time it no longer has the flaws that were inherent to AI models even a month ago. It no longer hallucinates any code, doesn’t make strange decisions, doesn’t go crazy, doesn’t skip functions, doesn’t duplicate functionality. I’ve concluded for myself that I will utilize my thinking and the ability to adapt, flexibility, in the future, to stay afloat in the labor market. Because right now AI models still can’t:

• Do layout. Simply because the quality of vision in multimodal models still limps. But this won’t last long it seems…
• Work with very huge systems. They don’t have taste and they don’t have the ability to assess and encompass a huge amount of information, raising the level of abstraction on-demand. So while Codex can close higher-and-higher-level tasks for me, it can’t encompass the entire codebase of 150 thousand lines with one general glance and assess what, for example, will be the priority next. What to work on next? What’s the highest priority bug? This includes product design, system design, business, etc.

My New Path

So I’ve started working on my portfolio now. I’m planning to transition into the role of architect or consultant, to make higher-level, product, or meta-changes in the development process, rather than directly writing code.

Because every time I think of getting a job as a regular programmer writing code, I feel like I’m deceiving my employer. It already feels unpleasant to me. Because honestly, if I write code manually, it’s going to be less productive than just prompting agents. And so whatever job I get, code will only be written by Codex without my participation. But in my eyes, that’s already dishonest…

So I’m figuring out a new path for myself and will soon share with you what I’ve decided.