Raw LLM Responses

4 YOE. We been adapting AI assisted coding in our team. Most improvements I've heard are like 10-30%. Software still needs \- Copious amount on coordination and communication \- You are making an API? You have to communicate to know exactly what they want. Design a solution that leverages the tradeoffs for the best solution but also modularity & stability in case things will need to change. \- Research. Research. Research. \- Testing. Testing. Testing. And not just tests but also manual tests. Checking the logs if everything is fine. Make tests robust so they don't be flaky, give false positives otherwise you can't trust your tests which makes them useless. \- Cleaning up the code. \*\*\*IT WORKS\*\*\* is not a good metric if I will need to walk on egg shells to modify it. \- Monitoring Logs. Some errors are fine. some are not. While this can be automated for resource use, dashboards and summaries of logs for earlier detection you still can't risk AI or automated tools to miss anything so you must have to go yourself to be assured everything is fine. I also worked during most of my University. Used LLMs for Labs. Stopped using them for proper labs where I knew I can learn something (we all know some labs are only passed by memorization) because otherwise I did not learn anything if used AI.