Claude Code Commands: Built-ins, Skills and the Ones We Run

What are Claude Code commands, and which ones are built in?

Three kinds of thing answer to a slash. Built-in commands are “coded into the CLI”, as the commands reference puts it, and it listed 63 of them on 26 September 2026, from /help and /clear to /goal and /deep-research. Bundled skills are prompts Anthropic ships that behave like commands. Custom commands are skills you write, and the same page ends with the only sentence that matters for them: “to add your own commands, see skills”. The built-ins a builder reaches for first, with the docs' own one-line descriptions:

/help prints the full list inside the CLI, which is more reliable than any third-party list, this one included, because the set changes with releases. The rest of this page is about the third kind.

How do you create a custom command?

Write a skill. The skills documentation is explicit that the older mechanism was folded in: “custom commands have been merged into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way.” Skills add a directory for supporting files, frontmatter that decides whether you or Claude may invoke them, and automatic loading when a request matches the description. Locations, per the same page: personal skills in ~/.claude/skills, project skills in .claude/skills checked into the repo, plugin skills in the plugin, and nested .claude/skills folders for monorepos.

The smallest skill that does a job, shaped like one of ours (details simplified, the structure is exact):

---
name: friday-update
description: Write the weekly client update from this week's Slack thread and meeting notes. Use when Marco asks for the Friday recap or a Monday kickoff.
argument-hint: [client]
disable-model-invocation: true
---

# Friday update

1. Read the client's CLAUDE.md and this week's Slack channel.
2. Pull the Granola notes for every call with the client this week.
3. Draft the update in the house voice: what shipped, what moved, what is next.
4. Post it as a Slack draft. Never send it.

Two fields carry the weight. description is what Claude reads to decide whether to invoke the skill on its own, so it should say when, not only what. disable-model-invocation: true makes a skill user-only, which is right for anything that posts, sends or spends. The frontmatter reference on the docs page lists the rest; the ones we use in production:

Which custom commands do we actually run?

24 project skills, counted from the workspace on 26 September 2026, most of them symlinked in from a shared repository so the agency team runs the same versions. They fall into seven families, and the largest three are the ones that produce output for other people: articles, ad campaigns and social posts. This guide was itself produced by the first of them, the article skill, which runs a demand check, a top-three audit, a draft, a blocking quality gate and the commit, in that order.

Three of them, described by what they do rather than what they are called:

• The article pipeline. One skill file of about 116 KB that carries a registry of four projects, a house benchmark, and every gate an article has to clear. It is invoked by name or by phrases like “new blog post on X”, and it produced the rewrite of every guide on this site on 26 September 2026.
• The weekly client update. Takes a client as its argument, reads that client's context file, the week's Slack thread and the meeting notes, and posts a draft, never a message. User-invoked only, because it touches a client channel.
• The morning brief. Reads the cash-flow priority model and surfaces the one or two actions that matter today. The only personal skill in the set, and the one that runs most often.

What broke while running 24 custom commands, and what fixed it?

Four failures, each of which is now a rule in the skill files themselves.

The pattern across all four: a skill fails quietly when it depends on something outside the file, a shell dialect, a symlink, a schedule, an ID. The fix each time was to move the check into something that cannot pass by accident.

When should a command be a subagent instead?

When the work would swamp the conversation. The subagents documentation defines them as “specialized AI assistants that handle specific types of tasks. Use one when a side task would flood your main conversation with search results, logs, or file contents you won't reference again: the subagent does that work in its own context and returns only the summary.” Each “runs in its own context window with a custom system prompt, specific tool access, and independent permissions”, and it “sends its own requests, which count toward the same usage limits”.

The two compose: a skill's frontmatter can set context: fork to run the whole skill in a subagent, which is how our long pipelines keep their tool output out of the session they were started from. In practice a skill is the default and a subagent is what you reach for when the skill's own output is the problem.

How does this compare with Lovable's slash commands?

Closely enough that a skill written for one is a short edit from the other. Lovable's Skills are workspace-level markdown with “a name, a description that tells Lovable when to use it, and markdown instructions that Lovable follows when the skill applies”, applied automatically on a matching request or by typing a slash and picking one, shared with every project in the workspace, on all plans. Its /goal command starts a run that keeps working “until the goal is achieved” for up to 10 hours, the same idea as Claude Code's /goal. The difference is what sits behind the command: a codebase and a shell on one side, a hosted app with its own backend on the other.

That is why the prompts on this site are written as structured briefs rather than one-liners: a brief with a Context line, numbered features and safe-guards is a skill without the frontmatter, and it runs in either tool. The advanced prompting guide covers Lovable's side of that, and the library holds the briefs.

Sources

All checked on 26 September 2026.

• Claude Code docs: Commands
• Claude Code docs: Extend Claude with skills
• Claude Code docs: Create custom subagents
• Lovable docs: Define reusable instructions with skills
• Lovable docs: Set a goal for Lovable