pukaworks

I've been building a multi-agent pipeline on top of Claude Code — 7 agents chained together (scout, researcher, writer, editor, evaluator, publisher). One surprising lesson: the Editor agent approved 100% of articles while the independent Evaluator rejected 45%. LLMs are bad at being critical of LLM output. We ended up enforcing title quality with regex instead of prompts because the Writer kept ignoring banned-word rules.