Breaking down code into smaller files (<200 lines of code), and then ONLY feeding the relevant files into the LLM helps a lot to improve the quality of code output and saves tokens.