context-engineering

Manage context windows by applying just-in-time data loading, subagent orchestration, and memory tooling with documented patterns and guardrails. This preserves tokens and prevents hitting context limits during extended agent sessions.

Just-in-time context loading reads and loads data on demand instead of upfront, preserving tokens for critical tasks. Use it in long-running agent sessions where upfront data dumps would exhaust context budgets prematurely.

Design subagent prompts to spawn summarizing subagents that return concise findings and file references rather than full data dumps. This pattern prevents context overflow and maintains efficient context budgeting across orchestration workflows.

Maintain cross-session continuity using memory tools like external session state, progress logs, and sidecar notes. These memory tools preserve continuity without consuming critical context window space in long-running agent sessions.

Keep context metadata lean by following CLAUDE.md guidance to ensure metadata remains navigable. Lean context metadata prevents unnecessary token consumption and supports efficient context budgeting in long-running agent sessions.

Avoid upfront data loading when long-running agent sessions risk hitting context limits. Instead, apply just-in-time context loading patterns to read data on demand, preserving context budget for critical tasks and subagent orchestration.