Context Engineering

The shift from prompt engineering to context engineering — why an agent's attention budget is a finite resource and how the three pillars (prompt, memory, context lifecycle) compose

The static skeleton — finding the right altitude, structuring instructions the model can actually parse, using examples as the primary steering tool, and ordering prompt blocks so caching helps

What survives outside the window — a taxonomy of memory types, when to write vs. when to recall, structured notes as durable planning state, and how memory invalidates

The operation that replaces window content with a more compact form — spectrum, triggers, preservation, custom-instruction tuning, design extensions, measurement, and a cross-framework reference

The runtime discipline — attention budgets, the compaction spectrum, just-in-time context loading, progressive disclosure, and sub-agents as a context-engineering tool

A close reading of Anthropic's ~340-line Claude design-agent system prompt — naming every design move worth learning from, grouped into seven design domains

A 10-step method distilled from the case study, followed by the elevation from single-prompt design to composed architectures — same invariants, different assembly