Learn / Reduce token usage
How to reduce Claude Code token usage
Seven concrete techniques to make your Claude Code allowance go further — from things you do by hand to things worth automating. For each: what it does, why it works, and how to actually do it.
Coding agents spend tokens fast, and most of that spend is avoidable waste rather than real work — the same file read twice, a thousand-line log dumped into context, a session that never gets reset. None of these techniques raise your Anthropic allowance; they just stop you burning it on nothing. Four of the seven can be automated so you don't have to remember them mid-flow. Those are marked.
01Compact and reset sessions sooner
Run /compact to summarize a long session, and start a fresh session when you switch tasks instead of carrying one all day.
Every turn resends the whole conversation. A bloated session pays for that bloat on every remaining message, so shrinking it early compounds across the rest of your work.
Type /compact at a natural break; open a new session per distinct task. Claude Code also auto-compacts as the window fills, but that's reactive and it drops detail you might have wanted.
02Read narrowly, not whole files
Point the agent at the specific function, symbol, or line range you care about instead of loading entire large files.
A read pulls the whole file into context and keeps paying for it on later turns. Most of a big file is irrelevant to the change you're making.
Ask for a specific range ("read lines 40–80 of X"), or search for the symbol first and read only around the hit. Prefer grep-to-locate, then a targeted read.
03Trim oversized tool output
Cut noisy command, search, and fetch results down to the part that matters before they reach the model.
The entire result of a tool call lands in context — a giant test dump or verbose build log costs the same whether you read ten lines of it or none, and it lingers in the resent history.
Pipe noisy commands through head, tail, or grep; ask for quieter flags; scope searches tightly. Tare automates this — it trims oversized output to a recoverable on-disk summary automatically.
04Block redundant re-reads
Stop the agent pulling the same unchanged file into context a second and third time in one session.
Each re-read is billed in full even though the file hasn't changed and is usually still in the window. Across a busy session this repeats silently over many files.
By hand this is nearly impossible to catch — you can't see it happening. Tare automates this — it recognises a file already in context and blocks the duplicate read before it costs a token.
05Delegate heavy work to subagents
Hand big, self-contained searches, reviews, or research to a subagent that returns only the conclusion.
The subagent does its bulky reading in its own context. The intermediate work never lands in your main session, so it never keeps costing you across later turns.
Use Claude Code's subagents for "go find/read/review X and report back." Keep the main thread for decisions, not raw exploration.
06Ease off as the window fills
Get more conservative about what you load as a session approaches its context ceiling, rather than pushing until it forces a compaction.
The fuller the window, the more each turn costs and the closer you are to a destructive auto-compact that drops detail. Backing off early keeps sessions sharp deeper in.
Manually, watch the context indicator and slow down. Tare automates this — it tightens its trimming automatically as context pressure rises.
07Keep receipts so you can tell what worked
Measure what your changes actually save, instead of guessing. You can't improve what you can't see.
Token waste is invisible in the moment. Without a record, you never learn which habits paid off and which were placebo — so the discipline slips.
Free monitors like ccusage read your local logs and report consumption over time. Tare automates this — it writes an append-only ledger of every intervention with measured savings, viewable with tare today.
The reason most of these techniques fade is that you never see them pay off. A number you can point to changes that. Tare's whole design is receipts-first: it doesn't ask you to trust a marketing figure — it logs what it actually removed, measured against your real sessions, and keeps that ledger on your disk where you can audit it.
Four of the seven — trim output, block re-reads, ease off under pressure, and keep receipts — are exactly the parts you'll forget to do by hand. That's what Tare automates. The other three (compact, read narrowly, delegate) stay yours to drive, and this list works with or without any paid tool.
Automate the four you'll forget
Tare handles the mechanical token-saving work inside Claude Code and shows you a measured receipt. Full-power free for 7 days — watch it save 2× its price before you pay $49.