AI Agents 與 Skills 及提示詞：有什麼不同？

No. A prompt runs once and lives in the conversation, where it competes with everything you send afterwards. A skill is a reusable file your AI loads automatically when a task matches, so the same expertise applies across every conversation without you pasting anything, and it is as loud on message forty as on message one.

Usually not the file format - both are a package with a SKILL.md document inside. What differs is how the document is written. A skill describes a role and a menu of things it can do, so you stay in charge of the conversation. An agent describes a short fixed intake, a handful of questions it asks before producing anything, and one named output packet that comes out the same shape every time.

If your task is a single kind of output and you want to stay in control of each request, a skill is usually enough. If it is a multi-step process where the deliverable should never vary in shape - and where a colleague on their first week should get the same result you would - an agent is the better fit, and it can use skills internally.

Not the ones sold as skill files, including ours. An agent file runs one session with you answering its intake questions. It does not browse the web on its own, does not connect to your CRM, bank, repository or ad account, and does not run while you sleep, unless the underlying assistant already has that access independent of the file.

Two questions. Is there a named deliverable - a Launch Packet, a Client Acquisition Packet - or a described role? A named packet means agent; a described role means skill. And does the description mention an intake list, specific inputs the file will ask for before producing anything? That gating behaviour is an agent trait; skills do not usually work that way.

A curated collection of ready-to-use prompts for a topic or role. You copy and run them manually, task by task, which suits work that varies each time. Packs are $9, from a single-purpose generator to a full role library, and they need no installation at all.

Yes, and most people end up doing exactly that: prompts for quick asks, skills for repeatable tasks, and agents for end-to-end workflows that lean on those skills. They are layers rather than rivals - a capable agent uses skills, and a skill replaces a prompt you would otherwise paste every time.

It did not stop working, it got buried. Prompts live in the conversation history, and as you add follow-ups, pasted documents and new requests, the original brief competes with material that is more recent and more specific. A skill avoids this because it is loaded from a file each time rather than read out of the scrolling history.

Skills need code execution enabled, under Settings then Capabilities. Anthropic's help centre currently lists Skills on Free, Pro, Max, Team and Enterprise, while its Academy tutorial lists Pro, Max, Team and Enterprise, so if you are on the free plan check Settings then Capabilities for your own account rather than trusting either page. On Team and Enterprise an owner controls that switch for the whole organisation, which is why the menu is sometimes greyed out on a work account. Where the Skills menu is not available, or in ChatGPT, Gemini or Copilot, paste the SKILL.md contents into custom instructions instead.