← Claude Code, start to finish
Claude Code

Managing Context and Cost in Claude Code

29 Jul 2026·6 min read

In the last two posts we covered the building blocks that give Claude Code real power — subagents and hooks. Now for the skill that quietly decides your speed and your bill: managing context. Every fact here is checked against the official docs.

What the context window actually is

Context is everything Claude can see in this session: the system prompt, your CLAUDE.md, auto memory, the files it has read, and the whole conversation so far. It is finite — a fixed budget of tokens. Run /context any time for a live breakdown of what is loaded.

Here is the part people miss: Claude Code sends your full conversation with every message. So a one-line question in a session that has been open all day does not cost one line — it pays for the whole history again. That is why a long session gets slower and more expensive even when you are barely typing.

/clear, /compact, and /context

# wipe the context when a task is finished — costs nothing
/clear

# summarize a long session but keep the important parts
/compact

# see exactly what is loaded right now
/context
Three commands do most of the work.

Use /clear when you switch to unrelated work — old conversation crowds out the files you need next and taxes every message. Use /compact when one long task starts to drag; you can even focus it, like /compact focus on the auth fix. One honest caveat: /compact reads the whole conversation to summarize it, so it is itself a large request, while /clear is free. And you do not have to remember — Claude Code auto-compacts as you approach the limit.

What survives compactionA fair worry: does compaction throw away my rules? Good news — your project-root CLAUDE.md is re-read from disk and re-injected after compaction, and so is auto memory. The system prompt is never in the message history, so it stays untouched. What you can lose are path-scoped rules and nested CLAUDE.md files; they reload the next time Claude reads a matching file. So keep any must-always rule in your project-root CLAUDE.md.

Why cost follows context

  • Prompt caching reuses repeated content like the system prompt and your CLAUDE.md, so it is not reprocessed every time. On a subscription the cache lasts about an hour.
  • Match the model to the job — Sonnet handles most coding and costs less than Opus. Switch with /model.
  • Tune thinking. Extended thinking helps hard reasoning but is billed as output tokens; for simple tasks lower the effort with /effort.
  • Delegate big reads to a subagent — the verbose file reads stay in its context and only a short summary returns to yours.

Treat context as your main budget. Keep it small on purpose: /clear between tasks, /compact when one drags, /context to check, /usage to see your spend, and the right model for the job. Do that and both speed and cost take care of themselves.

Next post: MCP — connecting Claude to your outside tools and data.

Managing Context and Cost in Claude Code · Ali Huzaifa