Context Budget Strategies

Spend Wisely Instead of Wasting

You now know how to tidy up your desk. But real pros go one step further: They plan in advance how to allocate their context budget -- before they type the first line.

Subagents: The Intern with Their Own Desk

Imagine you have an intern. You can send them off to research something. They go to their own desk, dig through documents, and come back with a short result. Your desk stays clean.

That is exactly how Subagents work in Claude Code. When you ask Claude to research or analyze something, it can start a separate agent that works in its own context. Only the result lands in your main context.

> Find all places in the project where we use user authentication,
  and summarize which patterns we are using.

Claude often recognizes on its own when a subagent makes sense. But you can also explicitly instruct it:

> Use a subagent to search through the entire API documentation
  and summarize only the relevant endpoints for user management.

One Task Per Session

The simplest and most effective strategy: One task, one session, one tidy desk.

Not:

• Fix bug, then build feature, then write docs -- all in one session

Instead:

1. Fix bug --> /clear
2. Build feature --> /clear
3. Write docs --> done

Each task gets a fresh context. It sounds obvious, but the temptation is great to just keep going. "Just this one more thing..." -- and suddenly the context is at 90%.

MCP Tool Search: Deferred Loading

If you have MCP servers configured (for example, for GitHub, Linear, or Jira), their tool definitions load into the context at startup. This can cost enormously:

• GitHub MCP without optimization: ~77,000 tokens just for tool definitions
• GitHub MCP with Deferred Loading: ~8,700 tokens

The difference? Almost 70,000 tokens saved -- that is more than a third of the entire context window.

You configure this in your .claude/settings.json:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "toolConfiguration": {
        "defaultMode": "deferred"
      }
    }
  }
}

Checkpoints and Rewind: Save Points

Do you know save points in video games? Before you fight the boss, you save. If you die, you load the save point and try again.

Claude Code has a similar concept: Checkpoints. Each step in the conversation automatically creates a point you can jump back to.

• Esc+Esc jumps to the last checkpoint
• You can also jump back multiple times to go further into the past

This is especially useful when Claude has pursued a wrong approach. Instead of leaving the error in the context and building on it, you jump back and give a better instruction.

Session Continuation: /continue and /resume

Sometimes work gets interrupted -- your terminal crashes, you have to go to a meeting, or the computer restarts. No problem:

• claude --continue (or /continue in chat) -- continues the last conversation in the current project
• claude --resume (or /resume) -- shows a list of all previous sessions for you to choose from

The context of the previous session is restored. You pick up exactly where you left off.

# Continue last session directly
claude --continue

# Choose from list
claude --resume

The Golden Rules

1. New task = /clear -- Always
2. Big research = Subagent -- Your context stays clean
3. MCP tools = Deferred Loading -- Save thousands of tokens at startup
4. Wrong path = Rewind -- Instead of leaving the error in the context
5. Session interrupted = /continue -- Nothing is lost