A practical guide to structuring multi-turn conversations with Claude (or any LLM) so that you get sharper answers, spend fewer tokens, and stop fighting context drift.
Most people treat a chat window like a bottomless notepad — they dump a data set, five tasks, and a vague goal into one thread and wonder why the output feels average by message 15. This repo collects the mental models and templates I use to avoid that.
- Context Window Fundamentals — what a context window actually is, and why "it can hold 200K tokens" doesn't mean "it should."
- When to Split Conversations — a decision checklist for staying in a thread vs. starting fresh.
- The Scaffolding Pattern — a 4-phase structure for high-stakes or multi-step conversations.
- Case Studies — the pattern applied to real workflows:
- Conversation Templates — copy-paste starting points for common conversation types.
Conversation design = output quality. How you split, sequence, and frame your prompts matters as much as the prompts themselves.