Agree. I think LLMs are usually not "harnessed" correctly for complex, multi-step problems—hence the `raink` CLI tool: https://github.com/noperator/raink