If you have used Claude Code for more than a week, three words keep showing up: skills, subagents and hooks. They sound interchangeable. They are not. Each answers a different question, and picking the wrong one is why a setup that works in a demo falls apart on real work.
For background, see How Agent Skills Work Under the Hood and the AI Tools Pricing Index.
Think of it as hiring: a skill is the training manual, a subagent is a colleague sent off to do one job alone, and a hook is a rule on the wall that applies whether anyone remembers it or not.
Skills: the training manual
A skill is a folder anchored by a SKILL.md file: short frontmatter with a name and a description of when it applies, followed by the instructions. Project skills live in .claude/skills/ and personal skills in ~/.claude/skills/, so a team skill can sit in the repository next to the code it describes.
The clever part is when the skill loads. Only its name and description sit in context at the start of a session. The full instructions load only when the task matches that description, and supporting files load only when the instructions call for them. That is progressive disclosure, and it is why a team can keep dozens of skills available without every session paying for all of them.
For a small business, a skill is the right home for anything you keep explaining the same way: a “publish a blog post” skill holding house style and the SEO checklist, or a “monthly client report” skill holding which numbers to pull. If you paste the same checklist into chat for the third time, that checklist wants to be a skill.
The limit matters too. A skill only loads when the model judges it relevant, and its instructions are guidance, not enforcement. If a rule must hold even on a rushed day, a skill is the wrong tool for it. That job belongs to a hook.
Subagents: the colleague sent off alone
A subagent is a separate AI assistant for one specific type of task, defined as a Markdown file with frontmatter (name, description, optionally a tool list and model) followed by a system prompt, stored in .claude/agents/ for a project or ~/.claude/agents/ for personal use. Built-ins include Explore for read-only research, Plan for plan-mode research, and General-purpose for complex multi-step work.
The point of a subagent is isolation. It runs in its own context window and returns only a summary to the main conversation, which is exactly what you want when a side task would flood the main session with search results, logs or file contents you will never look at again. The main conversation stays focused on the decision; the noisy work happens somewhere else.
That isolation has costs. A subagent does not see your main conversation, so pass it the exact files and goal it needs. Its requests count toward the same usage limits as the main conversation, and Anthropic warns at startup when the combined descriptions of custom subagents pass 15,000 tokens, so keep descriptions short and move detail into the system prompt.
A good first subagent is a reviewer: read-only tools, a clear checklist, one job. It cannot edit anything by accident.
Hooks: the rule on the wall
A hook is the most boring of the three, and the most reliable: a command, HTTP endpoint, tool call, prompt or subagent that runs automatically at a specific point in Claude Code’s lifecycle. Register hooks in settings.json, at user level, project level, or as a local uncommitted file, or define them in a skill or subagent’s frontmatter.
The events tell you what hooks are for. SessionStart fires when a session begins. PreToolUse fires before a tool call executes and can block it. PostToolUse fires after a tool call succeeds, a natural place to run a formatter on an edited file. Stop fires when Claude finishes responding, a place to run tests before the session is allowed to end.
Because a hook fires on an event whether the model likes it or not, it is deterministic in a way a skill can never be. A PreToolUse hook can refuse a dangerous command or access to a secrets file, and a Stop hook can refuse to finish until the tests pass. Configuration lives outside the main context window, so hooks cost almost nothing in context terms, but they run with your user’s permissions, so read every script before enabling it.
The three side by side
| Building block | What it is | When it fires | Best for | Watch out |
|---|---|---|---|---|
| Skill | A SKILL.md folder of instructions | When a task matches its description | Repeatable procedures | Guidance only, not enforcement |
| Subagent | A separate assistant, own context | When Claude delegates a side task | Noisy exploration, review | Needs exact instructions; counts toward usage limits |
| Hook | A command tied to a lifecycle event | When the event fires, every time | Rules with no exceptions | A bad hook can block real work; test it first |
How they work together on one job
Take a monthly client report. The main session loads a reporting skill, which sets the structure, the numbers and the tone. A research subagent gathers the figures and returns a short summary instead of a wall of exports. Meanwhile a PreToolUse hook refuses any attempt to open the secrets file, and a PostToolUse hook formats the finished report the same way every month.
No single block does that job alone: the skill cannot stop the secrets file, the subagent cannot guarantee the format, and the hook cannot teach the structure.
Start with one of each
- Write one skill: The checklist you paste most often, with a precise description of when it applies.
- Create one subagent: A reviewer with read-only tools and a single job, given exact files and a clear goal each time.
- Add one hook: One rule with no exceptions, such as formatting after every edit, tested on a safe copy first.
- Review after a week: Delete what nobody triggered. A small set that is actually used beats a large set that is mostly decoration.
Frequently Asked Questions
Do I need all three to start?
No. A repeated procedure means a skill, cluttered sessions mean a subagent, a forgotten rule means a hook.
Is a skill just a long prompt?
Not quite. A prompt serves one conversation, while a skill is stored, described and loaded on demand across conversations, with supporting files that load only when needed.
Can a subagent see my main conversation?
No. It starts with its own system prompt plus basic details such as the working directory, not your chat history, so pass it the goal, the files, and what a good result looks like.
Do hooks replace permissions?
No, they complement them. Use a hook for the rule and permissions as the backstop, and review both when either changes.
Teach with skills, isolate with subagents, and enforce with hooks. Get that division right and an AI coworker becomes a colleague you can hand real work to.



