# Context Window Management

> >-

- **Type:** Skill
- **Install:** `agentstack add skill-gb3h-agent-skills-context-window-management`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [gb3h](https://agentstack.voostack.com/s/gb3h)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [gb3h](https://github.com/gb3h)
- **Source:** https://github.com/gb3h/agent-skills/tree/master/skills/context-window-management

## Install

```sh
agentstack add skill-gb3h-agent-skills-context-window-management
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# 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](https://github.com/gb3h)
- **Source:** [gb3h/agent-skills](https://github.com/gb3h/agent-skills)
- **License:** MIT

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

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-gb3h-agent-skills-context-window-management
- Seller: https://agentstack.voostack.com/s/gb3h
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.