Claude Skills vs Agents
These two things are often confused because they both show up in the same conversations about "extending Claude." They're not alternatives to each other — one runs inside the other.
The one-line version
Agent
The AI system that takes actions
Claude Code, Claude.ai, Codex CLI, Gemini CLI. Reads your code, runs commands, generates output. The runtime.
Skill
The instructions that make it better
A plain text SKILL.md file installed into an agent. Defines a structured process for a specific type of work. No server, no permissions.
Skills don't replace agents. Skills run inside agents. You install skills to give your agent better processes for specific tasks.
Side-by-side comparison
| Agent | Skill | |
|---|---|---|
| What it is | An AI system that reasons and acts | A plain text instruction file |
| Examples | Claude Code, Claude.ai, Codex CLI, Gemini CLI | RAG Architect, Code Reviewer, Legal Review |
| Format | Running software / web interface | SKILL.md file on disk |
| Setup | Install the agent application | One npx command or paste into instructions |
| Runs code | Yes | No — read-only text |
| Has memory | Session-based or project-based | Persistent — always loaded |
| Scope | General purpose by default | Task-specific by design |
| Modifiable | Configured via settings | Edit the .md file directly |
| Shareable | Platform-specific | Yes — installable from GitHub |
How they work together
The relationship is straightforward: an agent is a vehicle; a skill is a specialised driver's manual for a particular type of road. You install skills into agents to make them better at specific tasks — without changing the agent itself.
A developer using Claude Code might install five skills: one for systematic debugging, one for PR review, one for API design, one for writing database migrations, one for incident response. Claude Code is the same agent for all of them — each skill gives it a structured process for a different type of work.
How it fits together
Agent
Claude Code
Installed skills
Code Reviewer · RAG Architect · CI/CD Builder
Result
Consistent, structured output for each task
Skills are also portable across agents. Install the same Code Reviewer skill into Claude Code, Codex CLI, and Gemini CLI — they all read the same SKILL.md format.
What about subagents?
Subagents are specialised agents that run as part of a larger multi-agent workflow. In an orchestrated system, one agent manages the overall task and delegates subtasks to subagents — each focused on a specific capability.
Skills work exactly the same way with subagents. A subagent responsible for code review gets the Code Reviewer skill installed. A subagent handling legal document work gets the Legal Review skill. Each subagent has skills matched to its role.
The pattern
Orchestrator agent coordinates → Subagents execute subtasks → Skills define each subagent's process → Consistent output at every step.
When to focus on each
If…
Your agent gives inconsistent output for a repeated task
Focus on
Install a skill
Because
The agent has the capability — it just needs a defined process. A skill structures how it approaches the task every time.
If…
Your agent can't access data or tools it needs
Focus on
Look at MCP Servers
Because
This is a capability problem, not a process problem. A skill won't help if the agent genuinely lacks access to external data or systems.
If…
You want to automate a complex multi-step workflow
Focus on
Consider a multi-agent setup
Because
One agent orchestrating several specialised subagents, each with their own skills, handles complex workflows better than one agent with many skills.
If…
You do the same task repeatedly and want better results
Focus on
Start with a skill
Because
It's the lowest-friction improvement. Install in seconds, no configuration, zero risk. Most "my agent isn't good enough at X" problems are process problems, not capability problems.
Which agents support Skills?
The SKILL.md format is an open standard. These agents all support it:
Frequently asked questions
Can I build my own agent with skills?
Skills extend existing agents — they don't build new ones. If you want to create a custom agent, look at the Claude API or the Agent SDK. Skills are for making existing agents (Claude Code, Codex CLI, etc.) better at specific tasks.
How many skills can one agent have installed?
Claude Code supports multiple skills simultaneously. Most professional setups run 5–10 skills. The practical limit is context window size — all installed skills are loaded into context at the start of each session.
Are skills specific to Claude or do they work with other AI agents?
The SKILL.md format is an open standard adopted by multiple platforms. Skills on findskills.co are tagged by agent compatibility — many work with Claude Code, Codex CLI, Gemini CLI, and OpenClaw. Currently 405 skills listed.
What's the difference between a skill and a custom agent?
A custom agent is built with code — it has its own logic, tools, and deployment. A skill is a plain text instruction file that extends a general-purpose agent. Building a custom agent takes days; installing a skill takes seconds. Start with a skill unless you genuinely need capabilities that don't exist in any agent yet.
Related comparisons
Browse 405 Agent Skills
Source-verified skills for every agent — organised by profession, installable in one command.