I’ve been trying out Claude to help with some pet Fortran projects and have found it very useful. As an example it has written a test program for my rbtree.f90 implementation. I’ve attached both so you can see the sort of code it writes.
test_rbtree.f90 (10.2 KB)
rbtree.f90 (25.4 KB)
I second having very good experienced with Claude, I have been using it extensively over the last couple of months. I use a version of this for my projects. I used it to help me debug and setup a solid chunk of my quantum chemistry project
Codex GPT 5.3 with OpenCode and nvim has been excellent. No skills, no tools, just my agents.md file and a project synopsis to read before each session. I wrote first 25,000+ lines of code without AI so the structure is solid. Now I let it write all the little subprograms and helper utilities and it follows my established style. The code it writes looks identical to mine including comment style. It’s like having the perfect employee that works for around $6 an hour. I work in motion control, robotics. Fortran is naturally the best choice. 
Could you please share how-to-do for those who are inexperienced with Claude? Thank you
Would you be using it to develop new things from scratch or work on existing codebases? if it is the latter, how large are these?
@BFelsher I am curious why you say that “Fortran is naturally the best choice” given that you work in motion control and robotics where Python and C++ dominate and also most Control/Optimal Control and RL libraries are developed in these languages?
I ask because in the past I wanted to explore the use of Fortran in these fields but got discouraged due to the paucity of libraries/support and visualistion options.
@general_rishkin I write post-processor software for converting point-vector data to machine kinematics. I have all the math routines and libraries that I could ever need/want locally. I find FORTRAN to be the most readable and easy to understand. I have no need for anything fancy - it’s all just linear algebra, matrices, arrays. It runs EXTREMELY fast. I have yet to see a post-processor that runs as fast as a FORTRAN post. There was a time when mainframe computers were shared that the airforce ONLY allowed post-processors on them if written in FORTRAN. This is pretty much a tradition now in my field. The source code for ALL the major post builders are written in FORTRAN, and for the most part still in f77. I know 2 guys that use C. Python is horrible…code is very unwieldy, and the posts run sloooowwww…not only that, but it is a royal pain….my FORTRAN posts will run on ANY OS, even a phone, without ANY changes to the source code. Plus…well…I love fortran, and if I have yet to see anything I’d need/want to do that it cannot accomplish. I do very little work with characters and strings. 99% numerical computing. I did write a post in python before, and I have written them using BASIC. At one gov’t company I worked for they wouldn’t give me access to fortran, so to make my work easier, I even wrote one using auto-it. Fortran is by far the best solution. FORTRAN FOR LIFE.
Claude is good but I prefer Codex. A new project to facilitate the use of Claude for Fortran projects (I have not tried it) is
Is it the same “Codex” that appears in the left menu of my ChatGPT Plus subscription main window? On clicking that I get a new window only offering to “Join the Codex app waiting list”.
Yesterday I started using GPT5.5 pro. It’s very good. Compacts more often for some reason (40-50% instead of 80%) which used to cause issues with memory/context loss, but I have not noticed any issues with GPT5.5 pro doing this- no degredation. I don’t use the app, so not sure about your question. I do all my development in the terminal using opencode, git, and nvim.
For those who are putting to much faith (in my opinion) in LLM coding agents might want to reconsider. Just saw this article on Tom’s Hardware. A Claude powered agent deleted a company’s entire database along with all the backups.
Scary shit, and a year ago, I used AI for a few weeks and it did the same thing to me. Fortunately, I was just testing it to the limit to test for “obedience”. In my case, it actually did it maliciously, and after MUCH pushing, admitted to it. I didn’t use AI again until a few weeks ago. I have learned to be EXTREMELY careful, not trust it, and severely limit its abilities. Key is to NOT GET LAZY, constantly have my git diff open and NEVER commit without a thourough test. I do not allow AI to commit. I keep my own separate backups. YES, it can be very dangerous. Much has improved though. I have had zero obedience issues, and zero hallucinations. It has really come a long way. I don’t use Claude though. Claude is much too “rah-rah-rah” cheerleader, trying to build up the programmer into thinking his code is the “second coming”. CODEX has been excellent. I wrote around 14,000 lines of code (previously stated 25,000 was an error in line count, onefetch was counting documentation and reference files) without ANY AI. For the most part it reuses MUCH of my code. It has now written about 5,000 lines of code, and I am even hard pressed to see which code IT wrote versus mine. It has done an excellent job of following my very tight guidelines and style. I have to put some of the blame on what happened to that company on THEM. They obviously put FAR TOO MUCH faith in AI, and bypassed the human. That is laziness…and complacency…