What are your AI dev workflows in large codebases?

I've been using Claude/Amp/Codex at both my job + side projects. It's always been really nice for early projects, but I found the tools to be significantly harder to use in large/more mature codebases.

If it's auth for a playwright MCP, general context, codebase history, ...

What are your workflows in those? What tools worked well for you? What feedback loops worked out best?