Bytes, AI, Claude Code, Skills

What?

Skills are custom slash commands for Claude Code. They follow the Agent Skills open standard (agentskills.io) and give Claude detailed playbooks for specific tasks — code review, deployment, commit workflows, etc.

Why?

Skills let you encode repeatable workflows and domain knowledge so Claude behaves consistently. Instead of re-explaining your conventions every session, you write it once as a skill and invoke it with /my-skill.

How?

Skill Structure

my-skill/
├── SKILL.md           # Instructions (required)
├── template.md        # Optional template for output
├── examples/          # Optional example outputs
└── scripts/           # Optional helper scripts

Skills live at one of three levels:

Level	Path	Scope
Personal	~/.claude/skills/<name>/SKILL.md	All your projects
Project	.claude/skills/<name>/SKILL.md	Current project
Plugin	<plugin>/skills/<name>/SKILL.md	Where enabled

Creating a Skill

mkdir -p ~/.claude/skills/code-review

Write SKILL.md with frontmatter and instructions:

---
name: code-review
description: Reviews code for best practices, security, and test coverage.
allowed-tools: Read Grep
---
 
When reviewing code, analyze:
1. **Structure**: Is the code organized logically?
2. **Error handling**: Are errors caught and handled?
3. **Security**: Any SQL injection, XSS, or auth issues?
4. **Tests**: Is coverage adequate?
5. **Performance**: Any obvious inefficiencies?
 
For each issue found, explain the problem and suggest a fix.

Then invoke:

/code-review src/auth.ts

SKILL.md Frontmatter Reference

Field	Description
name	Display name (lowercase, hyphens, max 64 chars)
description	What it does — Claude uses this to auto-invoke
argument-hint	Autocomplete hint, e.g. [issue-number]
disable-model-invocation	true to prevent Claude auto-invoking (manual / only)
user-invocable	false to hide from / menu (background knowledge)
allowed-tools	Space-separated tools Claude can use without asking
model	Specific model override
effort	low, medium, high, max
context	fork to run in isolated subagent
paths	Glob patterns limiting when skill activates

Dynamic Variables

Use these in your SKILL.md:

• $ARGUMENTS — all arguments passed
• $0, $1, $N — specific argument by index
• ${CLAUDE_SESSION_ID} — current session ID
• ${CLAUDE_SKILL_DIR} — skill’s directory path

Dynamic Context with Shell Commands

Prefix a shell command with ! to inject live data before Claude processes:

---
name: pr-summary
context: fork
allowed-tools: Bash(gh *)
---
 
## PR Context
- Diff: !`gh pr diff`
- Comments: !`gh pr view --comments`
 
Summarize this PR.

Invocation Control

Config	You invoke	Claude invokes	Use case
Default	Yes	Yes	General utilities
disable-model-invocation: true	Yes	No	/deploy, /commit (side effects)
user-invocable: false	No	Yes	Background knowledge

Top Skills

Popular skills worth trying:

Skill	Purpose
Superpowers	TDD methodology, atomic tasks, test-driven approach
Frontend Design (official)	UI/UX design quality and best practices
Remotion	Animations, timing, audio, captions, 3D content

Installing Skills from Marketplace

/plugin install <skill-name>

Bundled Skills

Claude Code ships with these built-in:

• /simplify — review and improve code quality
• /batch — plan and execute large-scale changes
• /loop — run commands on recurring intervals
• /claude-api — build and debug Claude API apps

Skills vs Plugins

• Skills — single-purpose command, simple SKILL.md, no namespace, quick to create
• Plugins — complete packages bundling skills, hooks, MCP/LSP servers, agents; namespaced (/plugin:skill), shareable via marketplace