Anthropic dropped a cheat sheet for controlling Claude Code, and there are seven ways to do it, not one. Once you know all seven, Claude goes from helpful to scary good.
The Big Idea First
Most people control Claude Code one way: they put everything in their CLAUDE.md. That works, but it's about 15% of what's possible. There are seven different controls, and each one exists for a different job.
Here's the thread that ties them together, and it's worth holding in your head as you read: every one of these controls is really about two things, when an instruction shows up, and how much weight it carries. Some load always. Some load only when they're relevant. Some never enter the chat at all. Master that and you stop overloading Claude and start actually directing it.
Control 1
CLAUDE.md
The always-on table of contents for your whole project. It's the one file Claude reads automatically at the start of every session, and it stays loaded the whole time. Use it for the stuff that's true every single time: your stack, your commands, your conventions, and pointers to deeper docs.
Control 2
Rules
Specific do's and don'ts. The trick is they can be path-scoped, meaning a rule loads only when Claude touches certain files. A rule like "all API handlers must validate input before processing" can sit quiet until Claude opens something in your API folder, then it fires. That keeps the instruction out of the way until it's actually relevant.
Control 3
Skills
Step-by-step playbooks for a repeatable task. A skill stays dormant, just a name and a short description, until you trigger it or the task matches, and only then does the full procedure load. Perfect for "every time I do X, here's exactly how." Think a release checklist, a code-review pass, or your exact content-publishing flow.
Control 4
Subagents
Little helpers that go do a side task in their own window and report back just the answer. The whole investigation, all the messy intermediate steps, happens in a separate context and never clutters your main chat. Great for deep searches, log analysis, or any job you want done off to the side while your main thread stays clean.
Control 5
Hooks
Automations that fire on their own when something happens, like running your linter after every edit, or pinging Slack when a job finishes. Hooks are actual code that runs on an event, so they're deterministic (they always happen) and they cost almost nothing, because they run outside the chat instead of as instructions Claude has to remember. Use them for anything that should just happen, every time, no judgment needed.
Control 6
Output Styles
These change Claude's entire role, like switching it from a coder into a general assistant or a teacher. Because an output style is written into the system prompt itself, it carries the highest weight of anything here, Claude follows it hardest. Powerful, so use it deliberately: it can replace the default coding behavior unless you tell it to keep that too.
Control 7
Appending The System Prompt
Quick add-on instructions for tone or formatting, just for that one session. It adds to Claude's existing role instead of replacing it, and it doesn't stick around afterward. Handy for one-off jobs ("answer in British English, no bullet points") without permanently changing anything. Heads up: piling on too many appended instructions actually lowers how well any of them get followed.
Don't try to control everything from one place. Match the control to the job:
A quick decision guide
True every session? CLAUDE.md.
True only for certain files? A path-scoped rule.
A repeatable procedure? A skill.
A messy side investigation? A subagent.
Should just happen automatically? A hook.
Changing Claude's whole role? An output style.
A one-off tone or format tweak? Append the system prompt.
The reason this matters is simple: stuffing everything into your CLAUDE.md buries the important parts under things Claude only needs once in a while. When you spread instructions across the right controls, each one shows up exactly when it's relevant and carries the right amount of weight. That's the whole game. That's how you stop overloading Claude and start directing it.
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.