Cindy Zhu.
← all free guides
Productivity Claude

9 Claude Code token-saving moves

Hey, it's Cindy ๐ŸŒฑ You commented TOKENS, so here's the full playbook for cutting your Claude Code bill. The bonus: almost every move also makes Claude smarter, because a bloated context costs more AND thinks worse. So trimming it saves money and sharpens answers at the same time.
the one idea

First, why Claude Code costs what it costs

One idea explains every move below, so it's worth 20 seconds. Every time you send a message, Claude Code quietly re-sends your entire context window in the background: the system prompt, every tool and MCP definition, your CLAUDE.md, and the whole conversation so far. You pay for all of it, on every single turn. So a long, cluttered chat taxes each new message, and Claude also reasons worse when the window is stuffed. The whole game is keeping that window lean.

๐Ÿ’ก There are two kinds of move below: habits you do while you work (clearing, compacting, planning) and setup you do once (CLAUDE.md, models, MCP). The habits save the most, day to day.

screenshot this

โœ… The checklist

Screenshot this and tick them off. The first three are the daily habits that save the most.

  • โ˜  1. /clear between unrelated tasks
  • โ˜  2. /compact when a phase ends but you still need the thread
  • โ˜  3. Plan mode (Shift+Tab) before any real work
  • โ˜  4. /context to see what's eating your window
  • โ˜  5. /usage to watch your real spend
  • โ˜  6. Keep CLAUDE.md under ~200 lines, push the rest into skills
  • โ˜  7. Send noisy work to subagents (tests, logs, doc fetching)
  • โ˜  8. Match the model to the job with /model
  • โ˜  9. Turn off unused MCP servers and write specific prompts

one at a time

The 9 moves, explained

1. Clear the deck with /clear

What it does: wipes the conversation so your next message starts from a clean window. Why it saves: the moment you switch to a new task, the old thread becomes dead weight that still gets re-sent (and paid for) on every message. Clearing it drops that cost to zero. How: type /clear as soon as you finish one task and start an unrelated one. It's a full wipe with no summary, so when you still need the history, use /compact instead.

2. Shrink without losing the thread: /compact

What it does: replaces the long conversation with a short summary you can keep working from. Why it saves: you keep the important decisions but drop the token-heavy back-and-forth. How: run /compact at a natural break, and steer it if you like: /compact focus on the API code and the decisions we made. Claude Code auto-compacts when you're nearly full, but doing it yourself at a clean stopping point gives a much tidier result.

3. Plan before you build (Shift+Tab)

What it does: turns on plan mode, where Claude lays out the approach before touching anything. Why it saves: the most expensive tokens are the ones spent going down the wrong path and redoing the work. A 200-token plan you approve first prevents a 20,000-token wrong turn. How: press Shift+Tab to cycle into plan mode before any real task, read the plan, then let it run.

4. See what's eating your window: /context

What it does: shows a live breakdown of everything in your context right now, the system prompt, tool and MCP definitions, your memory files, and the conversation. Why it helps: you can't trim what you can't see. If a stack of MCP tools or a giant CLAUDE.md is hogging the window, this is where you catch it. How: run /context whenever things feel sluggish or pricey.

5. Watch the meter: /usage

What it does: shows your usage against your plan's limits. Why it helps: no bill surprises, and you can see when you're burning fast and ease off. How: run /usage once a day. On Pro or Max it shows your plan limits; if you're on pay-as-you-go API billing, /cost shows the dollar cost of the current session instead.

6. Keep CLAUDE.md lean (under ~200 lines)

What it does: your CLAUDE.md loads in full at the start of every session, so its length is a tax you pay constantly. Why it saves: a bloated memory file eats your window before you've typed a word, and Anthropic's own guidance is to keep it tight. How: trim CLAUDE.md down to your always-true rules and facts, and move any multi-step procedure into a skill, which only loads when you actually call it.

7. Offload noisy jobs to subagents

What it does: hands a heavy, verbose task (running the test suite, parsing logs, fetching docs) to a subagent that has its own separate context. Why it saves: the wall of output stays in the subagent's window, not yours, and only a clean summary comes back, so your main chat stays small and cheap. How: ask Claude to "use a subagent to run the tests and just report what failed," or set up a named subagent for jobs you repeat.

8. Match the model to the job: /model

What it does: switches which model is answering. Why it saves: the bigger the model, the more each token costs, and most work does not need the biggest one. How: keep Sonnet as your everyday default, drop to Haiku for simple, repetitive stuff, and only reach for Opus on genuinely hard reasoning. Switch anytime with /model.

9. Cut dead MCP servers, and be specific

What it does: every connected MCP server loads its tool definitions into your context on every turn, whether you use it or not. Why it saves: turning off the ones you're not using this session frees real space, and specific prompts stop Claude wandering (and wandering is expensive). How: run /mcp to see and disable unused servers, and write "add validation to the login function in auth.ts," not "improve my codebase."


keep this handy

๐Ÿ“‹ Command cheat sheet

CommandWhat it doesUse it when
/clearWipes context, fresh startSwitching to a new task
/compactSummarizes, keeps the threadA phase ends, you still need it
/contextShows what's in your windowThings feel slow or pricey
/usageYour spend vs plan limitsOnce a day
/modelSwitch modelTask is simpler or harder than usual
/mcpManage connectorsTurning off unused servers
Shift+TabPlan modeBefore any real work

skip this one

โœ… One thing you don't need to do

You don't need to "turn on caching." In Claude Code, prompt caching is automatic and already makes repeated context much cheaper. So skip that rabbit hole and put your energy into the moves above.

start here

๐ŸŽ“ Do these 3 first

If you only build three habits, make them /clear between tasks, /compact at the end of a phase, and plan mode before real work. Those three alone cut the most, because they attack the biggest cost: a long, messy context that gets re-sent on every message. Then glance at /usage once a day so nothing sneaks up on you.


the receipts

๐Ÿ”Ž Proof

Claude Code, manage costs (Anthropic official)

p.s. this lean-setup thinking, a tight CLAUDE.md, skills, subagents, the right model for the job, is exactly the kind of system we build together inside The AI System Club. The waitlist gets in first.

Want more like this? Follow @cindiezhu for daily AI you can actually use.