A lot of viral Claude cheat codes are useless. These are the 4 I actually use every day to get better answers while burning way fewer tokens, so you stop running out halfway through.
If you use Claude Code, "tokens" are basically your fuel. Every message, every file Claude reads, every tool it runs uses some up, and when you run low, things get slower and pricier. Most people have no idea where theirs are going, so they just burn through them and hope.
These four hacks are the ones I actually use, not the viral junk that doesn't move the needle. They get you better answers for less fuel, which is the whole game.
One Word First: Context
Your context window is Claude's short-term memory, everything it is currently holding in mind for your task. Tokens are how that memory gets measured. When people say a chat got "slow" or "dumb," it is usually because the context filled up with junk. Three of these four hacks are really about keeping that memory clean.
Stop Guessing
Type /context
In Claude Code, type /context and hit enter. It pulls up a clear breakdown of exactly what is eating your context right now, split by category: system prompt, tools, skills, memory files, and your messages. Most people have never looked at this, so they have no idea their token usage is being eaten by, say, ten tools they never use or a giant file loaded at the start.
Once you can see it, you can fix it. Turn off connectors or tools you are not using in this project. Trim a bloated memory file. Notice that one huge document is hogging half your window. You cannot fix bloat you cannot see, and this is the one command that shows it to you.
When You're Around Half Full
/compact (And Tell It What To Keep)
When your context is getting full, around the halfway point is a good habit, type /compact. It takes your whole conversation, summarizes it, and starts fresh with that summary loaded in, so you keep going without losing the important decisions. You shed the dead weight of every back-and-forth message but keep the conclusions.
The pro move most people miss: you can tell it what to protect. Type /compact keep the final database schema and the brand color codes exactly as written, and it will guard those details word for word while it trims the rest.
Switching Tasks Entirely
/clear For A Clean Slate
When you are done with one thing and moving to something totally unrelated, do not just keep typing in the same chat. Type /clear. It wipes the context back to empty so your new task starts clean, with none of the old baggage using up your window. Use /compact when you want to keep going on the same task, and /clear when you are starting something new.
Brains Where It Counts
Top Model For Thinking, Haiku For Grunt Work
Not every task needs the most powerful model. Run a top model like Opus (or the brand-new Fable 5, released June 9, 2026) as your main agent for the work that actually needs brains: big decisions, planning the approach, reviewing code. That is where the smart model earns its keep.
Then have your subagents, the little helpers Claude sends off to do busywork like reading files, scraping pages, or searching, run on Haiku, the fast cheap model. The grunt work does not need genius, it needs speed. You set a subagent's model right in its definition file with a single line: model: haiku.
Be Honest About This One
The savings here come from putting Haiku on the busywork, since it costs a fraction of the bigger models. Fable 5 is actually the premium option (it runs about double the price of Opus), so reach for it when you want maximum brainpower on the main task, not to save money. The money-saver is matching cheap tasks to the cheap model.
Don't Sit And Watch It Burn
Interrupt Fast
This is the most expensive mistake I see people make. Claude starts heading down the wrong path, and they just sit there and let it finish, watching it burn tokens building the wrong thing, so they can correct it after. That is double the cost: you pay for the wrong version, then pay again for the fix.
The second you see it going sideways, stop it (hit Esc in Claude Code) and redirect: "Stop, that is not what I meant. I want X, not Y." Catching it ten seconds in instead of two minutes in is one of the biggest token savers there is, and it is just a habit you build.
You can bake good token habits right into Claude so you do not have to remember them. Paste this into your CLAUDE.md file (your project's instruction file) and Claude will follow these rules every session.
Token-saving rules for your CLAUDE.md
# Token & Context Discipline Work efficiently. Treat the context window as a limited resource and follow these rules without being reminded: ## Reading files - Only read the specific files or sections you actually need for the task. Do not read whole folders "to be safe." - When you only need part of a file, read that part, not the entire file. - Prefer a quick search to locate the right place over opening many files. ## Subagents (delegating busywork) - When you spin up subagents for grunt work (reading files, searching, scraping, simple lookups), use the cheap fast model (Haiku) for them. - Keep the expensive thinking (planning, decisions, reviewing) on the main agent. ## Staying on track - If you are unsure what I want, ask me ONE question before building. Do not guess and produce a large wrong output I have to pay to redo. - If I tell you you are off track, stop immediately and re-align. Do not finish the wrong approach first. ## Keeping context clean - When this task is basically done and we are clearly moving to a new one, tell me: "This looks like a new task, want to /clear?" - When we have made a lot of progress and the window is getting full, suggest I run /compact, and remind me I can name what to protect. ## Output - Be concise. Give me the answer and the why, not a wall of restated context. - Do not repeat large blocks of code or text back to me unless I ask.
What To Do With This
Open the CLAUDE.md file in your project (or create one in the main folder), and paste this in. New to .md files? They are just plain text notes Claude reads, see my crash course. You can also keep a personal copy and drop it into any project. Adjust any line to fit how you work, it is yours now.
The Real Win
Tokens stop being a mystery. You can see where they go, reset on purpose instead of by accident, put cheap work on the cheap model, and stop paying for wrong turns. Same quality answers, a fraction of the fuel, and you get way more done before you ever hit a limit.
The Only AI Masterclass You Need
If this guide helped, but you’re looking to go deeper, I got you!! My 30-Day Challenge takes you from saving AI tips you never use to actually building with AI, step-by-step.
I show you exactly how I automated two e-commerce brands, my social media, and most of my personal life, then hand you the agents, workflows & systems to do the same. I’m teaching you every single thing I know with one lesson and one build a day.
Join the AI Masterclass →© 2026 Mariah Brunner. All rights reserved.