Most people use five Claude Code commands. I use twenty-five.
Open Claude Code, type a slash, and a menu drops down with more commands than you will ever bother to read. Most people scroll past it once, learn /clear, maybe /model, and never look again. That is the entire vocabulary for a tool they now spend hours a day inside.
The slash commands are not decoration. They are the difference between a chatbot that happens to live in your terminal and something that remembers your project, undoes its own mistakes, and asks before it does anything dangerous. I have been using this thing every working day for months, and the commands I reach for without thinking are most of the reason it stopped feeling like a novelty and started feeling like a coworker.
That is the embarrassing part. I write software for clients who pay by milestone, which means I spend most of my week in codebases I did not write, on problems nobody at the company can fully explain anymore. The margin between a good week and a bad one is how fast I can orient and how little I have to redo. These commands are that margin. Here are the twenty-five I actually use, grouped the way I think about them rather than the way the docs order them.
The gap between someone who is fast in Claude Code and someone who is quietly fighting it is almost never the prompting. It is the twenty commands they never learned were there.
The ones that keep the context clean
Claude Code has a memory the size of a room, and every message you send fills more of it. Once it is full, the model starts losing the beginning of the conversation and quality drops without announcing itself. A handful of commands keep that room tidy.
/context shows you the room. It draws a grid of what is currently eating your token budget, the system prompt, the files it has read, the whole conversation. The first time I ran it I found a giant log file I had pasted in an hour earlier, still sitting there, still costing me on every single turn. Now I check it the moment the model starts feeling slow or vague.
/cost is the bill. It tells you what the session has spent so far. If you are on a plan where that number matters, this is the command that kills the surprise at the end of the month.
/compact shrinks the history without binning it. It summarises everything so far into something smaller and keeps going. Reach for it when a session has run long but you still need what came before. /clear throws it all out. Empty room, clean start. This is the one everyone already knows, and they still under-use it. When a conversation has gone stale and the model is looping on a bad assumption, do not argue with it. Clear and begin again.
Two more live here because they change what the room costs. /model swaps the engine, so you can drop to something fast and cheap for grunt work and move up to the strongest model for the bug that has beaten you twice. /effort turns the thinking up or down. More depth for the gnarly problem, less for the mechanical edit. I leave it low and reach for high the second something stops being obvious.
The ones that let you break things safely
People are timid with AI in a real codebase for one honest reason. Fear that it will wreck something they cannot get back. These five are the seatbelt.
/rewind is the one I would not work without. It rolls back both the conversation and your files to an earlier point, so a wrong edit costs you a keystroke to undo instead of an afternoon. Before I found it I was committing constantly out of paranoia. Now I let Claude attempt the risky thing, and if it goes sideways I rewind and it never happened.
/branch lets you try an alternative without touching your main line. Fork the attempt, see whether the other approach is actually better, keep the winner. It is for the moments when two designs both look plausible and you would rather see them than argue about them.
/sandbox gives Claude an isolated environment to work in. When you want it to run something you do not fully trust, or install things you do not want anywhere near your real machine, this is the wall around it. /diff shows you exactly what changed before you accept anything, and I run it reflexively before every commit. Reading the diff yourself is the one step you should never hand off.
/security-review scans the code for vulnerabilities before it leaves your machine. On client work I run it on anything that touches auth, file uploads, or a database query built from user input. It has flagged things a tired human review would have waved straight through.
The ones that make Claude yours
Out of the box Claude Code knows nothing about your project or your habits. These four teach it once, so you stop repeating yourself every session.
/init scaffolds the CLAUDE.md file, the single document the model reads every time before it starts working. It writes a first draft by reading your repo, and then you trim it down to the house rules that actually matter. /memory edits that persistent context directly, so the things you keep having to say become things it already knows. Where the build script lives, which tables are sacred, how you like your commits written.
/config opens the settings and lets you shape the CLI itself, the theme, the model defaults, the behaviour you want on by default. /hooks is the power move most people never touch. It attaches your own commands to points in Claude’s lifecycle. Run the linter after every edit, block a dangerous command before it fires, ping you when a long task finishes. This is where Claude Code stops being a chat window and becomes automation you configured yourself.
The ones that widen what it can reach
By default Claude works in one folder with one brain. These four widen the blast radius.
/agents lets you build a team of specialised subagents, each with its own instructions and its own slice of the job. A reviewer, a researcher, a test-writer, spun up to work in parallel and report back. Overkill for small tasks, a genuine multiplier on the big ones.
/mcp connects external tools through the Model Context Protocol, your issue tracker, your database, your docs, so Claude can read from and act on systems that live outside the repo. /plugin manages extensions to the CLI itself, the growing ecosystem of add-ons that bolt new capability on.
/add-dir gives the session a second directory to see. When the thing you are building spans two repos, this is how you let Claude read both at once instead of working half-blind.
The ones nobody tells you about
The odds and ends. Small, and the kind of thing you only find by accident or by reading a list like this one.
/btw lets you ask a side question mid-task without derailing whatever Claude is in the middle of. A quick, off-topic thing, answered, then back to work. /copy drops the last response onto your clipboard, so you can paste an explanation or a snippet somewhere else without fighting your terminal’s text selection.
/pr-comments pulls the feedback on a pull request into your session, so you can work through review comments locally instead of living in a browser tab. /doctor diagnoses the install when something is broken and tells you how to fix it. It is the first thing to run when Claude Code is misbehaving and you have no idea why.
/remote-control hands the session to your phone, so you can kick off a long task at your desk and keep half an eye on it from the train. /teleport syncs a web session down to your terminal, so work you started in the browser carries on where your real tools live.
None of these matter much in isolation. What matters is the reflex of reaching for them in sequence. A normal hour for me looks like this:
/init # teach it the repo once
/context # what is eating my tokens right now
/effort high # this bug is not obvious, think harder
# ...let it work, read the plan before it edits...
/rewind # wrong approach, roll it back cleanly
/diff # read exactly what changed
/security-review # before this goes anywhere near a PRLearn three you have never run
Do not try to memorise all twenty-five. Nobody works that way. Open the slash menu on your next session and pick three you have never touched, the unglamorous ones, /rewind, /effort, /hooks. Use them until they are muscle memory. Then pick three more. The tool was always this capable. You were just talking to it with a fifth of its vocabulary.
The full list, with every flag and option, lives in Anthropic’s slash command reference. Worth a scroll even if you think you know them all, because the list keeps quietly getting longer.