chrisarg

After a very long hiatus due to the triplet of work-vacation-work, we return to Part 3 of my AI assisted coding of a Perl interface to a foreign library.
I will not repeat the post on my Github pages Github pages, or the documentation of my MetaCPAN package Bit::Set which features a "vibecoding" section.
However, I would like to share the take home points from my exercise:

• I found the agentic bots to not be very helpful, as they entered these long repetitive and useless reflections without being able to fix the problems they identified when the build of `Bit::Set` failed.
• The "Ask" mode chatbots could generate lots of code, but with subtle mistakes.
• Success with porting test suites from one language to the other was highly variable, ranging from near perfect to outright refusal to execute a difficult task.
• On the other hand, the chatbot was excellent as an auto-complete, often helping me finish the structure of the POD and putting together the scaffold to fill things in.

Are the chatbots worth the investment? I'd say that as an amateur programmer the chatbot was helpful to get me out of the writer's block, but did not really save much time for me.
I can not imagine that a professional who knows what they are doing would be assisted much by the AI chatbots, which is what the METR study said:
overall, experienced developers experienced a 19% drop in productivity with AI assistance.