| name | install-slash-command |
| description | Use when the user wants to install a new user-level slash command at ~/.claude/commands/. Triggers on phrases like "install a slash command", "add a /command", "create a user-level slash command", "save this as a slash command". |
The user wants to install a new slash command at the user level so it's available across all Claude Code sessions on this machine.
User-level commands live in ~/.claude/commands/ as .md files. The filename (without .md) becomes the invocation, e.g. ~/.claude/commands/foo.md → /foo. Subfolders namespace commands, e.g. ~/.claude/commands/git/sync.md → /git:sync.
Step 1: Gather inputs
You need:
- Name — kebab-case, no spaces, no leading slash. Optionally namespaced as
namespace/name.
- Description — one short line. Shown in the
/ picker and used to decide relevance.
- Prompt body — the instructions Claude will follow when the command fires.
- (optional) argument-hint — placeholder shown after the command name, e.g.
[file] or [pr-number].
- (optional) allowed-tools — restrict tool access, e.g.
Read, Grep, Bash.
- (optional) model — override default, e.g.
haiku for cheap/fast commands.
If the user gave you source material (a message, file, URL, existing prompt), convert it into a clear prompt body. If anything essential is missing (especially name or description), ask once before writing.
Step 2: Validate
- Name:
^[a-z0-9-]+(/[a-z0-9-]+)?$ — kebab-case, optional one level of namespacing.
- Check for collision:
ls ~/.claude/commands/<name>.md — if it exists, confirm with the user before overwriting.
- If namespaced, ensure the subfolder exists (create if not).
Step 3: Write the file
Format:
---
description: {{one-line description}}
argument-hint: {{optional — omit field if not used}}
allowed-tools: {{optional — omit field if not used}}
model: {{optional — omit field if not used}}
---
{{prompt body}}
{{reference $ARGUMENTS if the command takes input}}
Only include optional frontmatter fields the user actually wants — don't emit empty ones.
The prompt body should be written in second person addressing Claude ("You are...", "Read the file at..."), same style as a system prompt. If the command takes runtime input, reference it via $ARGUMENTS (or $1, $2, ... for positional args).
Step 4: Verify
ls -la ~/.claude/commands/<name>.md to confirm the file exists.- Report the invocation string to the user (e.g. "Installed — use
/foo or /foo some-arg").
- Mention they may need to restart the Claude Code session or run
/reload for the command to appear in the picker (new sessions pick it up automatically).
Notes
- Do NOT register the command anywhere else — dropping the file in
~/.claude/commands/ is the entire install.
- Do NOT confuse this with skills (which live in
~/.claude/skills/<name>/SKILL.md and have richer metadata). If the user's request sounds like it wants autonomous triggering on certain phrases, suggest a skill instead — slash commands only fire when explicitly invoked with /.
- Do NOT add this to any plugin marketplace or index — it's a local user-level file.
