Claude Code commands: the ones I use and how to make your own

Most people who start with Claude Code end up typing the same long prompts again and again. Check the open tasks, look at today's calendar, flag what is late. Word for word, pasted from a note.

That is exactly what a command is for. Once you know which commands already exist and how to make your own, Claude Code stops feeling like a chat window and starts feeling like a tool you set up once.

This guide covers the built-in commands worth learning, the terminal flags that start a session, and how to write custom commands with arguments.

What are Claude Code commands?

Claude Code commands are shortcuts you type to control a session or run a saved prompt, and they come in three kinds: built-in slash commands, command-line flags, and custom slash commands you write yourself.

Type / at the start of the prompt and Claude Code shows every command available right now. That menu is the source of truth. Anthropic ships new versions often, and commands get added, renamed or folded into others. If a command in this guide is missing in your version, the menu will tell you what replaced it.

Which built-in slash commands should you learn first?

The built-in commands to learn first are the ones that manage context, because a full or messy context is the most common reason Claude starts making odd choices. These are the ones I use daily:

The habit that matters most is /clear between tasks. A conversation that started on an invoice script and drifted into a blog post carries both into every answer. Clearing costs nothing. What should survive belongs in CLAUDE.md or memory, which is covered in my guide to Claude Code memory.

/compact is the softer version. It keeps you in the same task but shrinks the history. You can steer it: /compact keep the final pricing decisions and the file list tells Claude what the summary must hold on to.

Two shortcuts sit next to the slash commands. Start a line with ! to run a shell command directly, without Claude interpreting it. Type @ followed by a path to point Claude at a specific file.

Which terminal commands start Claude Code?

The terminal commands that start Claude Code are claude plus a few flags, and the most useful ones are -p for one-off runs and --continue or --resume for picking up old work.

1. claude opens an interactive session in the current folder.
2. claude "fix the failing test" opens a session with a first prompt already sent.
3. claude -p "summarize today's errors" runs once, prints the answer and exits. This is the one for scripts and cron jobs.
4. claude --continue reopens the most recent conversation in this folder.
5. claude --resume lets you pick an older conversation from a list.

-p is where Claude Code turns from an assistant into infrastructure. You can pipe data into it, like cat errors.log | claude -p "group these by cause", and use the output in the next step of a script. Combined with a custom command, a scheduled job can run the same routine every morning without you opening a terminal. Add --output-format json when another program needs to read the result.

How do you create a custom slash command?

You create a custom slash command by saving a Markdown file in .claude/commands/, where the file name becomes the command and the file body becomes the prompt. That is the whole setup.

1. Find the prompt you keep retyping. If you have pasted it three times, it qualifies.
2. Create the folder. .claude/commands/ in your project, or ~/.claude/commands/ for a personal command that works everywhere.
3. Save the prompt as name.md. The name is what you will type, so keep it short.
4. Add a description in the frontmatter, so the / menu explains what it does.
5. Type /name in Claude Code. If it does not show up, open a new session.

Here is a full example. Say you run a small agency and every Monday you want a client status overview:

---
description: Weekly status update for one client
argument-hint: [client name]
allowed-tools: Read, Grep, Glob
---

Write a weekly status update for $ARGUMENTS.

1. Read clients/$ARGUMENTS/notes.md and the files changed this week.
2. List what was delivered, what is in progress, and what is blocked.
3. Flag anything waiting on the client.
4. Keep it under 200 words, plain language, no jargon.

Do not send anything. Put the draft in drafts/$ARGUMENTS-status.md.

Typing /client-status acme replaces $ARGUMENTS with acme everywhere. For more than one input you can use $1, $2 and so on, like /compare $1 $2.

The frontmatter fields worth knowing:

Two more tricks. A line like !`git status` runs that shell command first and puts the output into the prompt, so Claude starts with fresh facts. It needs the matching Bash permission in allowed-tools. And @path/to/file inside a command pulls that file into context. Subfolders work too: .claude/commands/sales/follow-up.md still runs as /follow-up, and the folder shows up in the description so you can tell similar commands apart.

Should you write a command or a skill?

Write a plain command when you decide the timing, and write a skill when Claude should also recognize the moment by itself. Recent versions of Claude Code merged the two: a skill in .claude/skills/client-status/SKILL.md also answers to /client-status.

For anything that sends, publishes or deploys, I keep the timing with a human. Either as a plain command or as a skill with disable-model-invocation: true. How to write a skill description that Claude actually picks up is in the skills guide.

Do custom commands actually save time?

Custom commands save time mostly by removing the rewriting and the forgetting, not by making Claude faster. The prompt is the same every run, so the output is consistent, and nobody on the team needs to remember the exact wording.

That consistency is the real point. METR ran a randomized trial in 2025 and found that seasoned open-source maintainers were 19% slower with AI help than without it, while they predicted the opposite. Speed does not come from the model alone. It comes from the setup around it. And in the Stack Overflow 2025 Developer Survey, 51% of professional developers said they use AI tools every day. Daily use is where retyping the same prompt adds up.

Commands are also how a Claude Code setup becomes something you can hand over. A colleague who clones the project gets your commands with it. I run large parts of my business through Claude Code, with me reviewing what goes out, and I walked through that daily routine in this article. If you want ready files to start from, my starter template files include a settings file and a starter skill, and there are more on the free resources page.

Where should you start?

Start with one command for the prompt you typed most this week, and use /clear between tasks from today. That alone changes how Claude Code feels.

After a week, look at what you still type by hand. Each repeated prompt is a command waiting to be written. Each command that Claude should trigger itself is a skill. Each rule that must never be broken is a hook. Small files, one at a time, until the setup does the routine work and you do the judging.