Custom Skills & Slash Commands

Skills and slash commands are how you give Claude Code reusable, on-demand instructions. A skill is a markdown file (SKILL.md) with YAML frontmatter that Claude invokes when a task matches its description; a slash command is a prompt template you trigger explicitly by typing /name. Both turn a repeated workflow ("review this diff", "write a migration", "audit for security") into one invocation instead of a re-typed paragraph.

Where they live (scope & precedence)

Location determines who gets the skill or command and whether it is shared:

When a project skill and a personal skill share a name, the project skill wins. So to keep a personal variant without overriding the team's, give it a distinct name (e.g. /review-fast rather than redefining /review).

SKILL.md frontmatter reference

A skill is a folder under .claude/skills/<name>/ containing SKILL.md. The YAML frontmatter controls when and how it runs:

A minimal skill:

---
description: Review the current git diff for bugs and security issues. Use when the user asks for a code review or before a commit.
allowed-tools: Read, Grep, Bash
context: fork
---

Run `git diff --staged`, then review each change for correctness bugs,
missing error handling, and security issues. Report findings grouped by
severity. Do not modify files.

When to fork the context

Set `context: fork for skills that do exploratory or verbose work — brainstorming, large-codebase analysis, research. Forking keeps all that intermediate output out of the main conversation so abandoned approaches don't bias later implementation, and only the skill's conclusion returns. Run inline (current`) when the skill needs the surrounding conversation or its step-by-step output is what you want to see.

Skills vs commands vs CLAUDE.md vs hooks

These four configuration surfaces are constantly confused; the exam tests telling them apart by *when each loads*:

The key distinction: CLAUDE.md is always-on context; skills/commands are pulled in only when relevant (progressive disclosure, keeping the base context small); hooks are executed by the harness, not decided by Claude.

Common setup mistakes

• Vague description — if the description doesn't name the task and its cues, Claude won't trigger the skill when you expect. Be specific about *when* it applies.
• Forgetting allowed-tools — omitting it can leave a skill with more access than it needs; list the minimum.
• Editing the file but not seeing changes — confirm the skill is in the right scope (.claude/skills/ for project, ~/.claude/skills/ for personal) and that a same-named project skill isn't shadowing your personal one.
• Using a skill where a hook belongs — if you want something to run automatically on every tool use, that's a hook, not a skill.