The context window is the fundamental constraint — everything else follows

Claude’s 200k token context window is significantly less than 200k usable before performance degrades. This single constraint explains every effective AI coding practice: subagents isolate research into separate context, /clear resets between unrelated tasks, specs files persist knowledge across context resets, and one-objective-per-conversation keeps focus sharp.

The concept of “context engineering” — a level beyond prompt engineering — captures this well. Where Context is the product, not the model argues that context is the product differentiator, here the claim is more operational: context is the bottleneck you’re always managing. Treat it like RAM — precious, limited, and requiring active garbage collection. This is the basis for Treat an agent as an operating system, not a stateless function — the full OS mental model (RAM as context, hard drive as persistent memory, garbage collection as decay) turns context management from ad-hoc to systematic. The economics compound the constraint: without Prompt caching makes long context economically viable, every conversation turn resends the full context at full price, making large context windows prohibitively expensive in practice.

The practical implication: common failure patterns (kitchen-sink sessions, correction loops, over-specified CLAUDE.md, trust-then-verify gaps, infinite exploration) are all context management failures. Fixing any of them means respecting the fundamental constraint. This is why Declarative beats imperative when working with agents matters operationally — verbose step-by-step instructions waste tokens that could carry actual working context. The most direct architectural response is One session per contract beats long-running agent sessions — using task contracts as session boundaries so cross-contract context bloat never accumulates in the first place.