Like many here, I kept running into Claude usage limits when building anything non-trivial. I was working with a job search automation pipeline (based on the Career-Ops project), and the naive flow was burning ~16k tokens per application — comple…