Context Window Management

Context Window Management

Context is finite. Manage it or lose coherence.

Symptoms of Context Pressure

• Tool outputs getting truncated or compacted
• Forgetting earlier parts of the conversation
• Repeating work already done
• Losing track of the current plan

Strategies

1. Externalize State

Don't hold state in context — write it to files:

• Progress tracking → memory/YYYY-MM-DD.md
• Plans → workspace files
• Intermediate results → temp files

2. Minimize Tool Output

• Use head, tail, grep instead of cat for large files
• Use read with offset/limit for targeted file access
• Pipe commands: command | head -20 instead of full output
• Use --quiet or --brief flags where available

3. Delegate to Sub-Agents

Large independent subtasks should be sub-agents, not inline work. Each sub-agent gets a fresh context window.

4. Checkpoint and Restart

If context is getting heavy:

1. Write a handoff document (see task-handoff)
2. Summarize current state to memory
3. Suggest the user start a new session to continue

5. Avoid Context Bloat

• Don't re-read files you've already processed
• Don't paste large outputs into messages — summarize
• Clean up: once a tool output is processed, the information should be in your working state, not re-fetched

Rule of Thumb

If you're doing more than ~15 tool calls on a single task, consider whether a sub-agent or checkpoint would be more efficient.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gb3h
• Source: gb3h/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.