新版舊版
Claude Code Token Savings
The 7 setups I really use every day, so Claude saves you tokens and does more
黃志弘 Leo
Published 2026-06-07 · Updated 2026-06-09 · 1 readers
After a year and a bit with Claude, my biggest shift was going from opening a browser to ask questions, to letting it do the work in the background.
What made the difference was moving to Claude Code. The same job now costs fewer tokens, and it can do a lot more. Much of my content, data work and customer support runs on it now.
This page collects what I actually use, for anyone who wants to work the same way.
What made the difference was moving to Claude Code. The same job now costs fewer tokens, and it can do a lot more. Much of my content, data work and customer support runs on it now.
This page collects what I actually use, for anyone who wants to work the same way.
Chat version vs Claude Code
Chat version (browser / phone)
- Open a page, type a question
- No memory, you re-explain the background every time
- Can't touch your files or your tools
- Can't run tasks on its own
- Good for: occasional use
Claude Code (terminal)
- Runs in the terminal, can carry out tasks
- Has long-term memory (CLAUDE.md)
- Connects to your tools (MCP)
- Can spawn subagents and run on a schedule
- Good for: when you want it to really work for you
Not installed yet? Start with the Claude Code install guide, you can install it without writing a line of code.
7 setups I really use every day
1Give it a long-term memory (CLAUDE.md)
Put a CLAUDE.md in your project root with who you are, what you're working on, and your rules. Claude Code reads it on every start, so you never re-explain the background.
Token savings: That big block of background you retype every day is gone. For how to write one, see the full CLAUDE.md guide.
2Save repeated rules as a Skill
Write your fixed tone, formatting rules and workflows once, and save them as .claude/skills/skill-name/SKILL.md. Put a short description at the top saying when to use it, and Claude loads it automatically on related tasks, so you stop pasting them in again and again.
mkdir -p .claude/skills/brand-voice
# Start .claude/skills/brand-voice/SKILL.md with these three lines, then add your tone rules below:
# ---
# description: Brand tone to apply when writing posts and copy
# ---Token savings: You stop retyping the same instructions, which saves a block of input every time.
3Run subagents in parallel
Split a complex task across several subagents working at the same time. Each one has its own memory space and reports back only the conclusion, not the whole process.
Fewer tokens and faster: The main conversation doesn't get buried under detail. Every early morning I run three subagents in parallel to sort data, write drafts and analyse.
4Write results to files, keep only the key points in the chat
Long data, breakdowns and reports: tell it to write them to a file. With a big file, tell it to read only the part it needs. The cleaner the conversation, the cheaper every later message.
Save the result to report.md, don't paste all of it back into the chat
Read only lines 200–260 of this file, don't load the whole thingToken savings: Your context doesn't get more expensive and more confused as it grows.
5Connect your tools with MCP
MCP lets Claude Code go straight into your apps (mine are LINE, my database, my posting tools). One sentence and it really runs, with no copy-pasting or window switching.
/mcp
# Type this inside a session to open the menu, pick the tool you want, then authorise once in the browser as prompted
# You can also run this outside the session first: claude mcp add <name> ...It can do more: Without tools it's a consultant. With tools it's an employee.
6Schedule it to run while you sleep (the honest version)
Set up the repetitive work once and let it run every day. The way that actually works: use your computer's scheduler (launchd on a Mac, or cron) to load a script. As long as the machine is on, it fires on time while you sleep. Ask Claude to write the technical part.
Ask Claude: write me a script that runs OOO automatically at 10:00 every morning,
set it up with launchd, confirm it loaded, then send me the result when it's donelaunchd needs the computer on: Your own computer's scheduler (launchd / cron) only fires while the computer is on. If you want it to keep running with your laptop closed, use Claude Code's cloud Routines instead (Pro and above, still a research preview as of October 2026), created with /schedule in the CLI. The batch that went out to a few hundred people at 10:00 this morning was launchd, with my machine left running.
Saves both time and tokens: Set it up once and it runs itself every day, with nobody watching it.
7It isn't a god (the part I most want to say)
It makes mistakes and it says wrong things. Anything that matters, I verify myself and decide myself. Treat it as a teammate that never sleeps, not something to worship. Keep that straight and you'll last longer with it, and its errors won't burn you.
The mindset: Using it well is worth more than using it a lot.
Just starting? Set up these common ones first
You don't need all of it at once. Get these three right and you're already using it more deeply than most people:
① One CLAUDE.md
Write down who you are, what you're working on, and this week's focus. Highest return of any single step.
② Two or three slash commands you use most
For example /script (write a script), /post (carousel copy), /reply (answer a customer). Save the prompt you retype ten times a week.
③ An anti-style banned list
List the words and sentence shapes that make writing sound like AI and tell it to avoid them (for example "it's not A, it's B", "empower", "seamless"). The output reads human straight away.
Skip for now: installing a pile of plugins and hooking up every tool
One tool you really use beats twenty you installed and never touched. Get comfortable with one before adding the next.