What problem does it solve? Long AI coding sessions degrade as the context window fills with irrelevant tokens, causing repeated mistakes, lost state after compaction, and diluted attention. This Skill provides a systematic framework for controlling what enters, stays in, and leaves the context window. ## Core Features & Use Cases - Four Operations Framework: Write (persist state to CLAUDE.md, NOTES.md, memory files), Select (precision retrieval via grep, glob, subagents), Compress (compaction strategies and triggers), and Isolate (subagents, worktrees, fresh sessions). - Context Budget Planning: Phase-based token budgets (planning <20%, implementation <50%, testing <70%) with concrete actions when thresholds are exceeded. - Decision Tables: Clear guidance on when to use /clear vs /compact vs subagents, plus anti-patterns to avoid like dumping whole files or running excessive MCP servers. - Use Case: During a large refactoring task, write architecture decisions to NOTES.md, delegate test runs to subagents, compact at 50% context usage, and recover full state after compaction by re-reading the scratchpad. ## Quick Start Ask the AI to apply context engineering principles to plan a long coding session with a scratchpad file, subagent delegation, and compaction checkpoints.
