The short answer
A prompt pack is text: a document of instructions you copy and paste into a conversation yourself, every time you want the behavior. A Claude Skill is a file on disk with metadata Claude checks automatically, so it can trigger itself when your request matches, without you pasting anything. An agent is the whole system doing the work, the model, the tools it can call, the memory it carries between steps, of which an installed skill is one ingredient, not a competing thing.
The confusion is understandable. Search for "AI prompt packs" and you will find products that are one PDF of copy-paste text, sitting next to products calling themselves "skill packs" that are the exact same PDF with a new label. The difference this guide draws is not about marketing names, it is about what the file actually does once you have it, which is a mechanical question with a mechanical answer.
What is a prompt pack
A prompt pack is a static piece of writing, almost always a document or a list, containing prompts someone else wrote for a category of task: cold email openers, blog outlines, resume bullet rewrites. Nothing about it is executable on its own. You open the document, find the prompt that matches what you are doing right now, copy it, and paste it into a chat window. If you want the same behavior next week, you do the exact same copy-paste again, because the pack has no way to notice you are doing that task and no way to run itself.
That is not a design flaw, it is simply what the format is: portable text, nothing more. A prompt pack works in any chat interface that takes text, which is a real advantage, and it costs nothing to maintain since there is no format to break. What it cannot do is persist, trigger itself, or enforce that you actually followed the steps in order. Most products that sell AI prompts and skills in the same storefront are, underneath the label, selling this format for both.
What is a Claude Skill
A Claude Skill is a folder with a required file, SKILL.md, that opens with two pieces of YAML metadata (a name and a description) followed by plain-markdown instructions. The mechanical difference from a prompt pack is entirely in that metadata: Claude reads every installed skill's name and description at the start of a session, and reads the full instructions only when your request matches a description's stated trigger. Nobody has to copy-paste anything; the skill notices its own moment and loads itself.
That single mechanism, matching a request against a description before reading the body, is what the sibling guide in this set, What are Claude Skills?, covers in full, including the frontmatter rules and the three-stage loading model. The short version for this comparison: a skill is a prompt pack that learned to trigger itself.
What is an agent
"Agent" gets used loosely, but the useful definition is specific: an agent is a model paired with tools it can call (file access, code execution, a browser, an API) and a loop that lets it take more than one step toward a goal without a human approving each individual step. Claude Code is an agent. Claude running inside the API with tool access configured is an agent. A single skill is not an agent by itself, it is one capability an agent draws on when the moment calls for it, the same way one procedure in a reference manual is not the whole employee who follows it.
This is the distinction that trips people up most: a skill and an agent are not on the same axis, so they are not really comparable. You do not choose between "a skill" and "an agent" the way you choose between a prompt pack and a skill. You choose an agent (Claude Code, claude.ai, or a custom build on the API), and then you decide which skills to install into it.
The differences, laid out
- Persistence. A prompt pack persists nowhere; it lives in a document you manage yourself. A Claude Skill persists on disk, in
~/.claude/skills/or a project's.claude/skills/, until you remove it. An agent persists as the running system itself, the thing the skill gets installed into. - Trigger mechanism. A prompt pack has none, a human decides when to paste it. A skill triggers itself by matching your request against its description. An agent is not triggered at all, it is the thing doing the triggering and the acting.
- Portability. A prompt pack is the most portable of the three, plain text works anywhere. A skill's content is equally portable as markdown, but its auto-trigger mechanism is specific to how Claude reads SKILL.md, so a skill needs to be adapted, not just copied, to work as, say, a custom instruction in a different tool.
- What it can execute. A prompt pack executes nothing; it only shapes what the model says next. A skill can bundle real scripts the agent runs, not just instructions to reason through. An agent is the actor with the actual capability, filesystem, network, code execution, that both prompts and skills are ultimately trying to direct well.
- Composability. Prompt packs do not compose; each one is used in isolation. Skills compose deliberately, one skill's instructions can hand off to a differently named skill for the next stage of a pipeline. An agent is what holds several composed skills together as one working system.
The same job, three ways
Take one concrete job: sorting a messy inbox into what needs a reply today. As a prompt, that is a paragraph you paste into a chat: "Here are my unread emails, sort them into act, waiting, read, or archive." It works, once, for the emails you pasted. Tomorrow's inbox needs the same paragraph pasted again, and if you forget one of your own sorting rules between sessions, the model has no way to know that.
As a Claude Skill, that same job is the inbox-triage skill this site ships as a free, full sample: a description that states the exact trigger ("the inbox has more than about 15 unread"), nine ranked decision rules applied top-down, five named edge cases (a long thread, two asks in one email, an angry tone), and a fixed output schema. You do not paste a paragraph. You open your inbox, and if it is genuinely messy, Claude reaches for the skill on its own because the description matches.
The agent in both cases is the same thing: Claude Code or claude.ai, doing the actual reading of your inbox and writing of the sorted output. The prompt and the skill are two different ways of directing that same agent toward the same job; only one of them remembers itself.
When to use which
- Reach for a plain prompt when the task is genuinely one-off, or when you are still figuring out what the instructions should even say. Packaging something you have not stabilized yet is wasted effort.
- Reach for a Claude Skill once you are giving Claude the same instructions for the third time this month, and you are on a surface, Claude Code, claude.ai, or the API, that can actually read installed skills. Writing it down once and letting it trigger itself is strictly less work than re-explaining it forever.
- Think in terms of the agent when the question is bigger than one procedure, when you are deciding which tool to run your work through at all, or which set of skills to install into it as a working kit rather than one-off capabilities.
Do they compete, or combine?
They combine, not compete, and treating them as rivals is the whole source of the confusion. A prompt is the rough draft of an instruction; a skill is that instruction packaged so an agent can find and run it on its own; the agent is what actually carries it out. Nothing here is a replacement for anything else, each one is a different layer of the same stack.
If you already own a pack of plain prompts and are wondering whether it was money well spent, the honest test is whether you would ever want that instruction to trigger itself instead of being pasted again. If yes, the guide on how to write a Claude Skill shows exactly how to wrap those same words into something that does.