As I am still alone coding the TLALOC Ada 83 compiler, I tried to get help from Claude AI for coding in Ada 83. Does anyone use AI for Ada coding ? Ada 83 is especially suited to AI assisted programming with its simple package structure. You define the specification revise it with AI and then let the AI code the body under feedback supervision. It works extremely well.
Yo ![]()
I personally tried this for FOSS or a personal project.
I do not use it for industrial professional projects. Maybe in the future, if we can have a fine-tuned LLM well trained on Ada that can run on our computer / our company server, not on a remote LLM from another company.
On my side, I tried the following LLM:
- ChatGPT 5.1 (high reasoning): Meh… it needs more iterations than Claude for tasks
- Gemini 3: good but not the best
- Claude Sonnet/Opus: I use it for coding since 3, now with 4.5 it’s awesome
For your purpose, you have to give it a try because those LLM are trained on public code and ARM, so, it will try to use the latest language version possible (when it does not generate malformed code).
Furthermore, I exploring the uses of multi-agents systems and spec-driven-development.
When I 1st seen “spec-driven-development” I laughed a lot, because, this is what we doing with Ada… for decades ![]()
Currently, I am experimenting with the use of BMAD-METHOD on Windsurf and Claude Code. It is too heavy for small projects but is ok for larger ones. IMHO, we have to add custom agents for Ada (and SPARK) coding, but the default does the job… well.
I also quickly tried spec-kit and openspec but I am unable to do efficient things with it (the goal of LLM is to save time, not waste it).
In any way, you will have to review the code it write
because it can be too verbose, sub-optimal, or not doing what you want (rare with a spec-driven-framework).
Since search engines have become useless, I use AI to:
- Search the web
- Create summaries (give me an overview of this API with CRUD examples)
- Get debug snippets on environments I don’t know (and don’t want to know)
In short, to save a lot of time on acquiring/managing “stupid knowledge.”
I also use it in “rubber duck” mode. I explain my problem to AI, and it comes up with a solution strategy. Sometimes it’s silly (often depending on my own prompt), sometimes it gives me ideas ![]()
The only time I consulted her for Ada, she hallucinated in a very convincing way ![]()
In short, AI has become indispensable, like fire.
And the same consequences in case of misuse.
I don’t use it to code anything (yet). A lot of what I do relies on mathematics, and my impression, as well as my experience, is that AIs are still terrible with mathematics. (They don’t have actual reasoning skills, if by “reasoning” we mean logical deduction.) Besides, I enjoy the process of devising the solution and writing the code; that’s one reason I do the Advent of Code problems.
Last year I worked with an entry-level programmer whose community college programming class apparently consisted in teaching students how to query an AI to write the program for them. (“apparently” is load-bearing; I have no idea whether that’s what actually happened.) It was pretty impressive what the AI could do. But when it provided him invalid code, he was incapable of fixing it. I sat with him once to debug it and it was a 2-hour-long nightmare.
I am interested in using AI for code quality issues, and if AI-generated code produces decent enough comments that the debugging experience would have been easier, my attitude would change a bit. But only a bit.
To follow up: I sometimes do something like this with a different task at work, and I find more often that it’s much quicker to do it myself than put up with the give and take of prompt engineering. How do you feel about the efficiency of prompt engineering with AI 83 code?
(Some people say that AI’s will relieve us of the tedium of certain uninteresting tasks, but that’s only an improvement if the tedium of prompt engineering can be minimized.)
I occasionally use code generation features but I wouldn’t ever use Claude or AI with execution abilities. It can be made to delete all your files or worse by remote actors.
This is how you churn out people with no competence.