When I started developing SAFi over a year ago, coding harnesses either did not exist or I was not aware of them.
So I built the first version of SAFi completely manually. I created it file by file in sublime text and used tools like ChatGPT and Gemini to help with the code.
The process was slow and painful. I had to upload a specific file with a highly detailed prompt, the AI would write a section, and I would download it, test it, and repeat the process over and over again. The entire program was about nine big files.
I had to do a lot of debugging myself because it got to the point where not a single chatbot could hold the entire system in memory. Editing one piece broke another.
At the beginning of this year, I started using Claude Code. It was incredible to watch Claude go through my files and summarize my project in less than five seconds.
That was a game changer.
But the first rule I set up for Claude was to not modify the five faculties that made up the core system without my permission.
There was always something about Claude that I did not like.
It has this overconfidence, pushing back on things that are not its business, only to chicken out after you press it.
Some people see that as AI having character, personality, and even intent. But knowing it is only a stochastic parrot and sometimes a diarrhea mouth, I know it is just something it has been trained on by whoever develops it.
So about a week ago, I went out to search what else was out there for coding and found OpenCode.
I installed OpenCode in my terminal and found out it comes with a free AI model called "Big Pickle." It is totally free, so I started using Big Pickle for a bit.
This site was running on WordPress. WordPress is a great platform but requires maintenance and is vulnerable to cyber attacks.
So I used Big Pickle to convert it to a static site, but kept the WordPress database to retain some functionalities like drafting, editing, and scheduling posts with a SAFi agent I built.
I ended up with a headless WordPress setup.
The amazing thing is that Big Pickle did it all without much of a problem, and totally free.
I also like that Big Pickle does not have an attitude like Claude does. It just does things. I have to be more descriptive sometimes, but I don't mind it.
Eventually, I ran out of Big Pickle's free usage and got the $10 plan called Go, give you a lot more usage and comes with a huge selection of AI models to choose from, but honestly, big pickle is enough for me.
I don't think I am going back to using the cocky Claude again. I don't see the need for it. GLM models are amazingly good for writing, which is the only place where I think frontier models still have an edge.