markus41/plugins/claude-code-expert/skills/context-management

Claude Code Context Management

Complete guide to managing context window, compression, and conversation flow.

Context Window Overview

Claude Code manages a conversation context window that accumulates messages, tool calls, and results throughout a session.

Context Limits

• The context window has a fixed token limit based on the model
• Claude Code automatically warns when approaching limits
• Auto-compact triggers when threshold is reached (configurable)

Configuration

// settings.json
{
  "autoCompact": true,
  "contextWindow": {
    "compactThreshold": 0.8,
    "warningThreshold": 0.9
  }
}

Context Reduction Strategies

1. /compact Command

/compact                         # General compression
/compact focus on authentication # Preserve auth-related context

What gets preserved:

• Key decisions and conclusions
• File paths and structure discoveries
• Error patterns and fixes
• Current task state

What gets summarized/dropped:

• Verbose tool outputs
• Intermediate search results
• Exploratory code reads
• Redundant conversation turns

2. /clear Command

/clear

Complete reset — use between unrelated tasks.

3. Sub-agents for Research

Offload research to sub-agents to keep main context clean:

// Instead of reading 20 files in main context:
Agent(subagent_type="Explore", prompt="Find all API endpoint definitions")

// The agent researches and returns a summary
// Main context only gets the summary, not all file contents

4. Targeted File Reads

// Bad: Read entire large file
Read(file_path="/path/to/large-file.ts")

// Good: Read specific section
Read(file_path="/path/to/large-file.ts", offset=100, limit=50)

// Good: Search first, then read specific matches
Grep(pattern="function authenticate", path="/path/to/")

5. Background Tasks

Long-running operations in background don't consume main context:

Agent(run_in_background=true, ...)
Bash(command="npm test", run_in_background=true)

Context Consumption by Tool

High Context Consumers

| Tool | Context Impact | Mitigation | |------|---------------|------------| | Read (large files) | Very High | Use offset/limit | | Bash (verbose output) | High | Pipe through head/tail | | Grep (many matches) | High | Use head_limit | | Agent (results) | Medium-High | Agent summarizes internally |

Low Context Consumers

| Tool | Context Impact | |------|---------------| | Glob | Low (just file paths) | | Write | Low (content sent, not echoed back) | | Edit | Low (just the diff) | | TodoWrite | Very Low | | AskUserQuestion | Very Low |

Conversation Flow Patterns

Focused Session

1. /clear (fresh start)
2. State the objective clearly
3. Claude researches and plans
4. Claude implements
5. Claude tests
6. Session ends

Long Session with Multiple Tasks

1. Task A work
2. /compact (preserve Task A context)
3. Task B work
4. /compact (preserve A+B context)
5. Task C work
6. ...

Research-Heavy Session

1. Spawn research agents (background)
2. Work on other tasks while agents research
3. Agents return summaries
4. Synthesize findings
5. Implement based on research

Auto-Memory

Claude Code can automatically save important context across sessions:

What Gets Saved

• Project structure discoveries
• User preferences
• Recurring patterns
• Key file paths
• Configuration details

Memory Location

• ~/.claude/projects/<project>/memory/MEMORY.md — Auto-loaded
• Additional topic files in same directory

Controlling Auto-Memory

# Disable auto-memory
export DISABLE_AUTOMEMORY=1

# Or in settings.json
{ "autoMemory": false }

Manual Memory Commands

/memory              # View current memories
/memory add <text>   # Add memory entry
/memory clear        # Clear all memories

# In conversation
"Remember that we always use pnpm, not npm"
"Forget the previous instruction about yarn"

Token Usage Tracking

/cost Command

Shows current session costs:

/cost

Output includes:

• Input tokens used
• Output tokens used
• Cache read/write tokens
• Estimated cost in USD

Reducing Costs

1. Use /compact to reduce repeated context
2. Use sub-agents for exploration
3. Use claude-haiku-4-5 for simple tasks (/model claude-haiku-4-5-20251001)
4. Use targeted reads instead of full file reads
5. Use Grep to find specific content instead of reading everything

Session Management

Continue Previous Session

# Resume last conversation
claude --continue
claude -c

# Resume specific session
claude --resume <session-id>

Session ID

# Set specific session ID
claude --session-id my-session-123

Conversation ID

# Set conversation ID within session
claude --conversation-id conv-456

Best Practices

1. Start focused — State your objective clearly upfront
2. Compact regularly — Don't wait for auto-compact
3. Use sub-agents — Keep research out of main context
4. Read selectively — Use offset/limit for large files
5. Clear between tasks — /clear for unrelated work
6. Track costs — Use /cost periodically
7. Leverage memory — Save important discoveries for future sessions
8. Background long tasks — Use run_in_background for slow operations