This skill addresses context drift that occurs during extended coding sessions when an AI agent loses track of strategic decisions as chat history grows. It targets Claude Code specifically and operates through four coordinated mechanisms.

First, it separates tactical chat noise from strategic context, saving the latter to filesystem-based persistent storage under ~/.claude/projects/ so that critical decisions survive beyond any single session window.

Second, it uses TaskCreate to maintain a durable, visible record of work-in-progress that persists across sessions. This replaces informal notes or manual handoff documents that tend to go stale quickly.

Third, it implements a Pause/Resume protocol that updates both CLAUDE.md and project memory at the point of interruption, so a resumed session can reconstruct the exact state at which work stopped.

Fourth, it triggers Plan Mode at decision points where major code changes are under consideration, forcing explicit trade-off analysis before execution proceeds.

The skill is specifically designed for situations where a single context window is insufficient — multi-day engineering work, large refactors, or any project where strategic intent must be preserved across many interactions. It does not introduce external databases or APIs; persistence is handled entirely through Claude Code’s native filesystem primitives and CLAUDE.md conventions.

Use cases

Preserve architectural decisions across a multi-day refactor without rewriting handoff notes each session

Track in-flight tasks through interruptions using TaskCreate so resumed sessions know exactly what was in progress

Prevent an agent from reverting to earlier, discarded approaches as context window fills

Enforce explicit trade-off review before the agent makes significant structural code changes

Resume a coding session the next morning with project memory fully synchronized to the previous session's endpoint

When to use it

Engineering work that spans multiple sessions or days and involves evolving strategic decisions

Large codebases where context window saturation causes the agent to lose track of earlier design choices

Projects where multiple pause/resume cycles are expected and continuity of intent matters

Workflows already using Claude Code that need structured memory without external tooling

When not to use it

Short, self-contained tasks that fit within a single context window with no session breaks

Environments not using Claude Code, since the skill depends on Claude Code's native primitives

Projects where CLAUDE.md and ~/.claude/projects/ filesystem conventions are not available or permitted

Teams needing multi-user shared memory, as this skill targets single-agent filesystem persistence

Managing Long Coding Sessions With Claude

Use cases

When to use it

When not to use it