I want to highlight some unique things that Grok Code can do that no other model seems able to handle besides being widely considered one of the weakest models overall.
Writing Books
Right now, Grok Code is the only model where you can ask for a full book to be written. You can prepare a chapter list and it will generate the entire book for you. And the best part: you can try it for free while it lasts.
Try this in Cursor, it’s excellent. I’m considering using Cursor to write books, instead of coding from now on, and sell them on Amazon.
Behind the scenes:
Grok Code is arguably the least lazy model out there. It just keeps working without pause and rarely interrupts with clarifications like “Did you mean this?” or “Shall I continue?”
A Life Lesson in Disguise
There’s also something deeper you can learn from this model. Grok Code shows us that a product can be both terrible and relentless at the same time. Even when backed by a multibillion-dollar company, the output can still be disappointing. Yet people hesitate to admit it directly, instead praising its speed when in reality, all it’s doing quickly is ruining the project faster.
Grok Code is arguably the least lazy model out there. It just keeps working without pause and rarely interrupts with clarifications like “Did you mean this?” or “Shall I continue?”
Uh, weird, i’ve never seen it happens in my session.
It might be related to the way i interact with agents: they have a very short leash.
extensive PRD
updated roadmap on file to keep track of changes
ask to develop one feature at the time
testing
commit to repo
Ask to develop one feature at the time (and so on)
Maybe, trying to be as Batman, preparation is important.
I can definitely see the value in your process. It is thorough, well-documented, and reduces the risk of major mistakes. However, sometimes it feels too long and heavy. For a small change, I do not want to go through fifty steps of documents, tests, and commits.
The whole point of using AI for me is to save time. With a model like GPT-5, I find that in 99 percent of the cases it does exactly what I asked for, and there is almost no need to fix or revisit the code afterward. That lets me move much faster.
In contrast, with a process that requires so many layers, you not only need to check whether the output is correct, but also whether it broke something else along the way. For me, that defeats the purpose of using AI as a speed booster.
i understand your point and it’s mostly the case for me when i need to do quick bugfixes.
Just keep in mind that within 1 month you (and the LLM) will not have memory of WHY some changes were done to the code, therefore when you need it again to fix an issue, the LLM could make a series of wrong assumptions (wasting tokens) in order to solve the problem.
Look mate, they are marvellous at writing documentation: ask them to write annotations on disk and you will see improvements…
Sometimes Grok code does great while fast, sometimes it does terrible and does it really fast (grateful for restore checkpoints). If you review the changes, which is faster than writing them, then its hard to have these AI’s really inject bad code. Of course most people aren’t reviewing the code, so only God knows what slop is being added.
a PRD document in markdown, regarding the project. For example if want to make a specific application, like a fastapi python app for backend and nextjs for frontend, i write a proper document detailing the characteristic i want to achieve (most often i get help from big models like gpt5 or gemini 2.5 pro).
I use some grounding rules for every project. For example the Zen of Python PEP 20 – The Zen of Python | peps.python.org is a very good starting point (in my opinion). I also add some common sense rule related to version control (git) and common sense about coding: always comment when you create a file/function/class, commit often and with comments, when you’re tackling an important feature create a branch, and so on. I started my coding adventure on a ZX Spectrum so i comment EVERYTHING even if i’m the only person that will read those lines in the future.
My first try with grok-code-fast-1 was good today using it under specific circumstances.
One of the React Material UI component properties was deprecated so I had many files I need to make adjustments to. I used gpt-5-high to research with the MUI MCP but told it not to make any changes, just tell me the new way of doing things. I then tagged all the files in and switched to grok-code-fast-1 and it made all the changes correctly very quickly
My take with Solidity, a language that LLMs still don’t manage well.
Normally, I do things manually since Solidity has to be extremely safe (aka bug free), but from time to time, I use gpt-5 to write scenarios, tests, theories .
Well, grok-code-fast-1 was able to do the same that I was doing with gpt-5-high in waaaaay less time, and did it good. I’d still won’t use this code in prod/mainnet, but I can see the improvement of LLMs in such niche languages like Solidity.
Can’t wait to test grok-code-fast-1 in my normal Python code!
