Commands vs Skills vs Plugins

Both Claude Code and Codex CLI use slash commands, but the concepts underneath are different. This doc breaks down what each term means and how the two tools compare.

Core Concepts

Commands (Built-in)

Fixed-logic operations that ship with the tool. You type /command and it executes immediately — no AI reasoning involved.

Claude Code examples: /help, /clear, /compact, /cost, /config, /model, /resume, /permissions, /plan, /diff

Codex CLI examples: /review, /fork, /permissions, /skills

Commands cannot be customized or extended by users.

Skills (Prompt-based Workflows)

Skills are instructions that teach the AI how to do something. They're defined in markdown files and the AI decides when and how to follow them.

Key difference from commands: skills are model-driven. The AI reads the instructions and executes them using its available tools. Commands are code-driven — they run fixed logic regardless of the AI.

Claude Code: Skills live in SKILL.md files with YAML frontmatter.

.claude/skills/my-skill/SKILL.md

---
name: my-skill
description: When to auto-trigger this skill
allowed-tools: Read, Grep, Edit
---

Instructions for Claude go here...

Codex CLI: Skills also use SKILL.md files with a similar structure. Currently feature-gated (codex --enable skills).

Plugins (Distributable Packages)

Plugins bundle skills, agents, hooks, MCP servers, and config into a shareable package. Think of a plugin as a "skill pack" with extras.

Claude Code: Mature plugin ecosystem with an official marketplace.

# Install from marketplace
/plugin marketplace add EveryInc/compound-engineering-plugin

# Or install from a git URL
/plugin install https://github.com/user/my-plugin

Codex CLI: Plugin/marketplace system is still emerging.

Side-by-Side Comparison

Concept	Claude Code	Codex CLI
Built-in commands	~50+ (/help, /clear, etc.)	Smaller set (/review, /fork)
Skills	Mature, auto-discovered from SKILL.md	Gated feature, growing catalog
Plugins	Marketplace + manual install	Emerging
Auto-trigger	Yes — Claude reads skill descriptions and decides when to use them	Yes — implicit skill matching
Skill location	~/.claude/skills/ (personal) or .claude/skills/ (project)	Similar directory structure
Namespacing	Plugin skills: /plugin-name:skill-name	N/A currently

Autonomy / Permission Models

Both tools have tiered permission systems controlling what the AI can do without asking.

Claude Code

Permissions are configured in settings.json with allow/deny lists for specific tools:

{
  "permissions": {
    "allow": ["Read", "Grep", "Glob"],
    "deny": []
  }
}

Can also run in "bypass permissions" mode where all tools are auto-approved. Individual skills can restrict their own tool access via allowed-tools in frontmatter.

Codex CLI

Three named modes:

Mode	Reads	Writes	Commands	Network
Suggest (default)	Yes	No	No	No
Auto Edit	Yes	Yes	No	No
Full Auto	Yes	Yes	Yes	Yes

Planning vs Execution

Claude Code

No built-in plan-then-build separation in vanilla Claude Code. The /plan command enters "plan mode" where Claude explores and writes a plan before implementing.

Compound Engineering plugin adds this formally:

• /ce:plan — Research + write a plan.md
• /ce:work — Execute from a plan.md
• /ce:brainstorm — Explore ideas before planning

Codex CLI

Built-in plan-then-execute flow:

1. Agent reads the codebase
2. Develops a plan and streams it
3. Proposes edits (approval in most modes)
4. Executes and tests iteratively

Performance and Strengths

Area	Claude Code	Codex CLI
Complex refactors	Stronger autonomy and consistency	Improving
Terminal/DevOps tasks	Good	Leads (77.3% Terminal-Bench 2.0)
Context window	Large (varies by model, up to 1M)	Smaller (4K for 5.4 mini)
Multi-agent	Skills can spawn parallel agents	Single agent
IDE integration	VS Code, Zed, JetBrains	Terminal only

Key Takeaways

1. Commands = fixed operations, same in both tools
2. Skills = AI-readable instructions. Claude Code's are more mature; Codex's are catching up
3. Plugins = packaged bundles of skills + config. Claude Code has a marketplace; Codex doesn't yet
4. Planning = Codex has it built-in; Claude Code gets it via the Compound Engineering plugin
5. Permissions = Both have tiered autonomy. Claude Code is more granular (per-tool allow/deny). Codex has named modes.