Claude Code is slow? Here is why, and how to make it fast again

You ask for a small change and Claude Code takes a minute to answer. Every turn feels heavier than the last. This is a real and documented problem, Anthropic issue #3477 is titled Claude Code is operating very slowly, and it almost always comes down to three causes. The good news is the biggest one is also the easiest to fix.

Why Claude Code gets slow

Three causes, in rough order of how often they are the real culprit:

The fix, in order

1. Compact or clear the window. Run /compact to summarize the session down to the essentials and keep going, or /clear between unrelated tasks to wipe it while keeping your project loaded. This one step fixes most slowness.
2. Keep CLAUDE.md lean. A short, specific CLAUDE.md loads every turn without weighing the window down. Cut anything the agent does not need on every single message.
3. Write specific prompts. Name the exact file and function instead of fix this bug. A precise ask means Claude plans and edits, rather than reading half your repo just to orient itself.
4. Fix the local setup. On Windows, switch from PowerShell to Git Bash to avoid the known CLI freezing, and close editor tabs you are not using so the extension stops pulling them into context.
5. Use /fast only as a stopgap. Fast Mode trades some quality for speed. It is a band-aid. The durable fix is a lean context, not a faster-but-dumber run.

How to keep it fast

The developers who never fight slowness are not on a better plan. They keep the window small on purpose: /clear between unrelated tasks, /compact before a big one, and prompts specific enough that the agent plans instead of exploring blindly. The habit that prevents the slowdown is the same one that keeps the agent accurate, because a bloated window is also a forgetful one.

That habit, keep the window lean and the agent planning instead of guessing, is what kitstarter builds in. The kit ships a PreCompact hook and context-hygiene rules that keep the window small without you thinking about it, and its ask-first behavior means the agent plans before it explores. In kitstarter's own clean-room benchmark, the asked-first kit wrote about 40% less code than vanilla, and fewer, leaner edits mean less in the context window and a faster session. That same bloat is what causes the prompt is too long error, so fixing one fixes both.

Common questions

Why is Claude Code suddenly so slow? Usually your context window has filled up. Every file Claude reads and every tool it runs stays in the session, so after a dozen or so turns it can be reprocessing more than 100,000 tokens on every message. That is slow and it burns quota. Run /compact or /clear to give the space back and it speeds up right away. Anthropic has a documented issue for this, #3477.

How do I make Claude Code faster? Compact or clear the window between tasks, keep your CLAUDE.md short, and write specific prompts that name the exact file and function so Claude plans instead of running many exploratory reads. On Windows, switch from PowerShell to Git Bash and close unused editor tabs. Fast Mode with /fast helps in a pinch, but a lean context is the durable fix.

Does a bloated context window make Claude Code slow? Yes, it is the most common cause. The model reprocesses everything in the window on every turn, so a session that has piled up 100,000-plus tokens of old files and tool output is slow on every message, not just occasionally. Keeping the window lean with /compact, /clear, and a short CLAUDE.md is the single biggest speed win.

Is Claude Code slow because of Anthropic's servers or my setup? Both happen, but your setup is the more common and the more fixable cause. Server-side slowdowns show up as widespread reports and pass on their own. If it is only you, and it gets worse the longer a session runs, it is context bloat or local lag, both of which you can fix yourself in seconds.