While we don’t do game development, my company has investigated using gen ai for code and we found that it doesn’t reliably assist us in anything other than boilerplate code. I personally found that it hallucinates APIs on the regular which made it a massive waste for me as I basically had to go back and rewrite most of the code myself. Using gen ai for code reminded me of every time I’ve worked at a company that outsourced code; we rarely ever got what we asked for and by the time we got something usable it still wouldn’t be up to our standards, which generally resulted in scrapping what was delivered and having to rewrite the whole thing internally.
Using AI for code requires a high level of specificity, software houses that are getting the most out of it are not using it so much inline or vibe coding because as you say that leads to hallucinations or having to repeat yourself.
The common approach now is to use multiple agents with different responsibilities, the human effectively becomes a team leader and uses one chatbot to orchestrate a plan being very specific about the outcomes and the requirements and breaking the plan down into phases,
Another agent will be used to write the code, they will have set rules files to keep them on guard rails, they will check their work as they go and deliver a pull request at the end of it for the human to check and approve
The next agent will be there to write tests and check the work of the second agent.
It’s a lot more work than writing a single prompt and expecting a good outcome or close approximation, but it is getting better results now.
While we don’t do game development, my company has investigated using gen ai for code and we found that it doesn’t reliably assist us in anything other than boilerplate code. I personally found that it hallucinates APIs on the regular which made it a massive waste for me as I basically had to go back and rewrite most of the code myself. Using gen ai for code reminded me of every time I’ve worked at a company that outsourced code; we rarely ever got what we asked for and by the time we got something usable it still wouldn’t be up to our standards, which generally resulted in scrapping what was delivered and having to rewrite the whole thing internally.
Using AI for code requires a high level of specificity, software houses that are getting the most out of it are not using it so much inline or vibe coding because as you say that leads to hallucinations or having to repeat yourself.
The common approach now is to use multiple agents with different responsibilities, the human effectively becomes a team leader and uses one chatbot to orchestrate a plan being very specific about the outcomes and the requirements and breaking the plan down into phases,
Another agent will be used to write the code, they will have set rules files to keep them on guard rails, they will check their work as they go and deliver a pull request at the end of it for the human to check and approve
The next agent will be there to write tests and check the work of the second agent.
It’s a lot more work than writing a single prompt and expecting a good outcome or close approximation, but it is getting better results now.