Files
claude-foundations/improvements/session-and-context-management.md
Paul O'Reilly a4967df815 Initial commit: Claude Code foundations and improvements research
Conventions, community best practices research (Sept 2025 - March 2026),
and prioritized improvement backlog for Claude Code workflows.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-12 23:02:54 +13:00

2.1 KiB

Session & Context Management

Why This Matters

Context is the most precious resource in a Claude Code session. The 200k token window seems large but fills fast — especially with MCP tools, long files, and multi-step tasks. Community consensus: managing context deliberately is a 2-3x productivity multiplier.

Current State

  • No explicit context management practices documented
  • Auto-memory system is configured (built-in)
  • No compaction hooks

Key Practices

Session Hygiene

  • Use /clear between unrelated tasks. A fresh session costs ~20k tokens (10% of budget), leaving ~180k for work. Polluted context from failed approaches actively degrades output quality.
  • After two failed corrections, /clear and start fresh. Context is now polluted with failed approaches. Write a better initial prompt instead.
  • One task per session. The "kitchen sink session" (start task A, ask about B, return to A) is a documented anti-pattern.

Context Budgeting

  • Use /context to audit token usage. MCP tools consume context even when unused. Browser automation tools alone can eat 8-30% of available context.
  • Remove unused MCP servers from settings when not actively needed.
  • Compact proactively at 60-70% usage — don't wait for auto-compact.

Extended Thinking

  • Only ultrathink activates extended thinking as of Claude Code v2.0.0. Previous keywords like "think" or "think hard" no longer work.
  • Use ultrathink for complex reasoning tasks, architectural decisions, and debugging subtle issues.

Compaction Strategy

  • Default reserves 32k tokens (22.5% of 200k) for auto-compact buffer.
  • Setting CLAUDE_CODE_MAX_OUTPUT_TOKENS to 64k increases the buffer to ~40%.
  • PreCompact hook (see hooks-and-automation.md) saves transcript before compaction.
  • Pair with auto-memory: gradual learning via memory, emergency snapshots via PreCompact hooks.

Handoff Between Sessions

  • Before ending a long session, create a handoff document: goals, progress, blockers, next steps.
  • A /handoff or /catchup skill (see skills-and-progressive-disclosure.md) automates this.
  • "Context is like milk — keep it fresh and condensed."