Skip to main content
What this unit solvesA slash command binds a frequently-used prompt or workflow to a single /name invocation. The key difference from a Skill: a command is triggered by you deliberately; a Skill is invoked by the model based on its own judgment. This unit covers command format, authoring, $ARGUMENTS parameter injection, and the decision criterion for choosing a command over a Skill.

​
Learning objectives

  • Explain the difference and division of responsibility between a command (user-triggered) and a Skill (model-invoked).
  • Create a custom command file under .claude/commands/ or ~/.claude/commands/.
  • Use $ARGUMENTS, $0, $1, and named-variable syntax to give a command dynamic capability.
  • Use disable-model-invocation: true to prevent Claude from invoking a specific command on its own.
  • Decide whether a given requirement belongs as a command or a Skill, and state the criterion.

​
1. What a command is

A slash command is an encapsulated prompt you trigger by typing /name at the start of a Claude Code session. It does not depend on the model recognizing your intent — what you type is what runs (as of 2026-06, per the official commands documentation [1]). Three key properties:
  1. Deliberately triggered: you don’t type it, Claude won’t suggest it. The decision is entirely yours.
  2. Reusable: write it once; every future invocation is identical.
  3. Version-controlled: files in .claude/commands/ can go into Git and be shared across the team.
Commands vs. pasting a promptYou can paste the same prompt repeatedly, but every time you have to trim parameters and adjust phrasing yourself. A command packages that work: you type /<name> <args> and the logic stays consistent. That consistency pays dividends wherever the output format is fixed (commit messages, PR summaries, changelogs) or for inspection workflows you run repeatedly.

​
2. Commands vs. Skills: deliberate vs. automatic

.claude/commands/<name>.md and a Skill (SKILL.md) now share the same underlying mechanism in Claude Code [2]. Their trigger timing differs: Simplified decision rule:
  • High certainty, fixed workflow → command.
  • Has side effects, you don’t want the model deciding when to run it (deploy, send message, commit) → command + disable-model-invocation: true.
  • Should depend on reading user intent (code review, doc search, refactor suggestions) → Skill.
Workflows with side effects default to commandDeployments, pushes, Slack sends, writes to external systems — anything that is hard to undo — default to command + disable-model-invocation: true. You do not need Claude to trigger a deployment on its own because it “looks like the right time.”

​
3. Authoring: file location and structure

File locations (as of 2026-06): The filename is the word after /: .claude/commands/deploy.md responds to /deploy; .claude/commands/fix-issue.md responds to /fix-issue.

​
3.1 Minimal form

When you trigger /commit, this text is injected into Claude’s current conversation and Claude uses it to generate a commit message.

​
3.2 Using frontmatter to configure behavior

Commonly used frontmatter fields (as of 2026-06, per the official Skills and commands documentation [1, 2]):
Why disable-model-invocation: true mattersSuppose you write a /deploy command. Without the flag, the model may see “this code looks ready to deploy” in conversation and invoke /deploy on its own — while you are still reviewing. With disable-model-invocation: true, the command only runs when you explicitly type /deploy; the model cannot see it at all. For any workflow with side effects — deploy, send, push — add this flag by default.

​
4. Parameter injection: $ARGUMENTS and indexed syntax

When you trigger a command, everything you type after /name is injected into the command body.
1

Simplest approach: $ARGUMENTS

Use $ARGUMENTS to receive everything typed after /name:
Typing /fix-issue 123 replaces $ARGUMENTS with 123; Claude receives “Fix GitHub issue 123 following our coding standards. …” [2].If the command body contains no $ARGUMENTS, Claude Code appends the arguments as ARGUMENTS: <your input> at the end of the body — Claude still sees them [2].
2

Positional index: $0, $1, $2

To split multiple arguments, use $0, $1, $2, or $ARGUMENTS[0]:
Typing /migrate-component SearchBar React Vue gives $0 = SearchBar, $1 = React, $2 = Vue.Wrap multi-word arguments in quotes: /migrate-component "Search Bar" React "Vue 3".
3

Named parameters

Declare arguments in frontmatter and reference them by name in the body:
Typing /commit feature-x 456 gives $branch = feature-x and $issue = 456 [2].

​
4.1 Other available variables

​
5. Dynamic context injection: !`command`

Command body content can use !`shell command` to execute a shell command before the body is sent to Claude and embed its output inline [2]:
When the command is invoked, Claude Code executes each !`cmd` in sequence, substitutes the output, and only then passes the entire body to Claude as a prompt. From Claude’s perspective it receives plain text that already contains the git diff output. Multi-line commands use the fenced code block form:
Two traps with dynamic injection
  1. The execution environment is your shell. When you type /deploy, !`cmd` inside the command runs in your local shell (macOS/Linux defaults to bash; Windows defaults to PowerShell, configurable [2]). Do not write sensitive operations into a command file.
  2. Org-level disable setting: organizations can set "disableSkillShellExecution": true in settings.json, replacing all dynamic injections with [shell command execution disabled by policy]. Do not rely on this mechanism for critical steps in commands destined for public repos.

​
6. Tool comparison

Naming clarificationCommands and Skills share the same underlying mechanism in Claude Code: as of 2026-06 both are unified, and both .claude/commands/<name>.md and .claude/skills/<name>/SKILL.md produce a /name command [2]. The difference is that the former is a single file while the latter is a directory (including references/ and scripts/; see 04-4 Skills). This unit focuses on the “user-triggered” face shared by both. OpenAI Codex, GitHub Copilot, and Cursor each have a “prompt file” or “prompt template” concept, but the paths and mechanisms differ significantly; the common denominator is “prompt files support $ARGUMENTS-style variables” — consult each vendor’s current documentation for details.

​
7. Hands-on exercises

30-minute practice
  1. Write a /commit command: include disable-model-invocation: true and allowed-tools: Bash(git *) in the frontmatter; use !`git diff --cached` and !`git log --oneline -10` to pull in the staged diff and recent commit style, then ask Claude to output a Conventional Commits message.
  2. Create a global command: write ~/.claude/commands/weekly-review.md that uses $ARGUMENTS to receive your focus topic; start a new project session and confirm /weekly-review works across projects.
  3. Test the misfire guard: write a /deploy-prod command deliberately without disable-model-invocation; mention “looks like it’s ready to deploy” in conversation and observe whether Claude invokes it on its own. Then add disable-model-invocation: true, repeat the same sentence, and confirm Claude no longer self-triggers.

​
8. Common pitfalls

Anti-pattern list
  • Turning every workflow into a command: too many user-triggered invocations increase workflow friction. Use a Skill for cases where the model should judge the right time (code review, doc search).
  • Vague description: you will forget what the command does in three days. Describe the trigger condition and expected output precisely; “help me do X” is not a description.
  • Writing deploy/push/send commands without disable-model-invocation: the model may run them before you have confirmed, producing irreversible side effects.
  • Committing command files with sensitive data to a public repo: command bodies go into version control, so API keys, absolute paths, and internal URLs all leak.
  • Dynamic injection that depends on unverified state: writing !`./scripts/deploy.sh` when that script’s permissions or content are not fixed destroys reproducibility. Deterministic deployment workflows belong in a Hook (see 04-6 Hooks) or a standalone CI step.

​
Self-check

The bar for passing this unit
  1. Can you state the fundamental difference between a command and a Skill? (The trigger subject differs.)
  2. Can you write a command that accepts $ARGUMENTS in under five minutes?
  3. Can you identify which commands should carry disable-model-invocation: true?
  4. Do you have a prompt you type three or more times a week? Should it be a command or a Skill? State your criterion.

​
Sources and further reading

Factual claims are grounded in official documentation; fast-changing items are annotated as of 2026-05.
  • [1] Anthropic, “Commands,” code.claude.com, 2026. [Online]. Available: https://code.claude.com/docs/en/commands (as of 2026-06; covers built-in and custom commands)
  • [2] Anthropic, “Extend Claude with skills,” code.claude.com, 2026. [Online]. Available: https://code.claude.com/docs/en/skills (as of 2026-06; covers the unified command/skill mechanism, $ARGUMENTS, and complete frontmatter reference)
  • [3] Anthropic, “How Claude remembers your project,” code.claude.com, 2026. [Online]. Available: https://code.claude.com/docs/en/memory (as of 2026-06; CLAUDE.md and modular rules)