Claude Code's token overhead is mas

Claude Code is a beast for terminal-based refactoring, but the token burn is absolutely terrifying if you don't constrain its context. I've been swapping between Cursor and Claude Code for a few days, and while the "agentic" feel of the CLI is addictive, the way it reads files and pipes them back into the prompt can drain your budget (or hit rate limits) in minutes.

The biggest issue is how it handles the "Context Window" during iterative debugging. If you let it blindly ls and cat through a medium-sized project, it starts stuffing massive chunks of boilerplate into the prompt. Since it's using Claude 3.5 Sonnet, you're paying for those input tokens every single turn.

To keep the overhead manageable, I've stopped letting it "explore" on its own. Instead, I feed it specific file paths immediately. If I need a change across three files, I explicitly tell it:

# Instead of "Fix the auth bug", I use:
claude "Fix the auth bug in src/auth/provider.ts and src/auth/types.ts. Do not read the rest of the /src folder."

Another massive gotcha is the .claudecodeignore file. Most people forget this exists, but if you have a dist folder, a massive node_modules cache, or large .json data dumps in your repo, Claude Code might try to index or read them if it thinks they are relevant. I’ve added everything that isn't raw source code to the ignore list to prevent "token leaks."

My current .claudecodeignore setup:

• node_modules/ (Obvious, but essential)
• dist/ and build/ (Prevents it from reading compiled JS)
• package-lock.json (Absolute token killer; it never needs to read the whole lockfile)
• .git/ (Avoids indexing git internals)

.log (Keeps huge log files out of the prompt)

If you're doing a large refactor, don't use the interactive shell for the whole process. I've found it's way cheaper to use the one-shot command mode for specific tasks. When you stay in the interactive session, the conversation history grows, and every subsequent request sends that entire history back to the API.

One productivity trick that actually saves tokens: use a "Context Summary" file. I keep a docs/context.md that outlines the architecture of my app. When I start a session, I tell Claude to read that file first. It prevents the AI from spending 5-10 turns (and thousands of tokens) trying to "discover" how my state management works by reading every single slice file.

# Project Context
- State: Zustand in /src/store
- API: Axios wrappers in /src/api
- Routing: React Router v6

The tradeoff is clear: Claude Code's ability to execute shell commands and actually verify its own fixes via npm test is a superpower, but the cost is the overhead. If you aren't disciplined about what it reads, you're basically paying a "convenience tax" on every single line of code it generates.