Watched this Matt Pocock workshop last week and it nailed something I've been wrestling with: https://www.youtube.com/watch?v=-QFHIoCo-Ko His core idea — LLMs have a smart zone (~100K tokens) where they work brilliantly, and a dumb zone beyond that where they hallucinate and forget decisions. The fix isn't more memory. It's methodology: PRD → grill-me → kanban of vertical slices → each slice runs in a clean session that fits the smart zone. Pocock built it for traditional devs. I'm building n8n + Claude automations for B2B clients, and his framework doesn't translate 1:1 — different stack, different unit of work. What I want to know: what's your methodology for building real projects with Claude, end-to-end? How do you slice the work, carry state across sessions, and keep the model sharp on long builds? Anything battle-tested under real client work — I'd love to hear it.