How to stop paying for bloated AI coding instruction files on every message

Developers using agentic coding tools like Claude Code often accumulate large instruction files — sometimes 800 lines or more — that are loaded into context on every single message, inflating token costs continuously over a project's lifetime. The core problem is that all context is treated as equally important, even though most facts in an instructions file are only relevant occasionally. A practical solution involves splitting project memory into tiers: a minimal always-on file capped at around 250 tokens containing only universal conventions, path-scoped detail files loaded only when relevant code is touched, and an optional dynamic layer that injects facts matched to the current prompt. This tiered approach keeps baseline costs flat as a codebase grows, since new module details go into deeper tiers rather than the always-loaded file. The key discipline is enforcing a hard size ceiling on the top-tier file and ensuring each tier remains append-only and lightweight to avoid compounding overhead.