✂️ Cut your Claude Code bill: the 8-fix guide

Claude Code is the most powerful way to build with AI — and the easiest place to burn usage, because repo-sized context is expensive by nature. These eight fixes, in order of impact, routinely cut bills by more than half without cutting capability.

The 8 fixes

1. Clear between tasks. The /clear command resets context. Finished the login fix? Clear before the styling task. Carrying a dead task's context into the next one is the #1 silent cost.
2. Compact long sessions. When a session must continue, /compact summarizes the conversation down and keeps going — the handoff-brief trick, built in.
3. Point, don't dump. Name the files that matter ("the bug is in checkout.js") instead of letting exploration read half the repo. You know your codebase; telling it is free.
4. Keep CLAUDE.md lean. Your project's CLAUDE.md loads every session. It should be a tight brief — commands, conventions, architecture in outline — not documentation. Audit it monthly; every stale paragraph is a tax on every message.
5. Plan mode before build mode. Agreeing on the plan first (plan mode exists for this) means one build pass instead of three. Rework is the expensive kind of token.
6. Match the model to the task. Claude Code lets you pick models: routine edits and boilerplate don't need the flagship. Save the big model + ultrathink for architecture and hard bugs.
7. Scope your subagents. Subagents (agent teams) each carry their own context. Powerful for real reviews; wasteful as a default. Use them where wrong-and-shipped costs more than the tokens.
8. Fix the environment, not the symptom. If Claude keeps re-discovering the same quirk ("tests need the env flag"), put it in CLAUDE.md once. Every re-discovery is paid exploration.

The mindset shift

Cheap sessions and good sessions are the same sessions: small context, clear target, one task at a time. The bill isn't a tax on using Claude Code — it's a meter on unfocused context. Focus is free.