AI Agents 與提示詞：真正重要的差異

A prompt is one request: you paste it, you get an answer, and it is gone when you close the tab. A skill is a method the assistant loads whenever a matching task comes up, so it survives across sessions without you pasting anything. An agent takes multiple steps toward a goal, decides what to do next based on what it just found, and uses tools to get there. All three run on the same underlying model - the difference is how much of the process they own.

Give it a goal where the correct second step depends on what the first step returns. If it runs the same fixed sequence regardless of what it found, it is a prompt chain with an agent label on it. If it reads the result and changes what it does next, it is an agent. If it cannot decide its own next step, it is not an agent, whatever the marketing says - and a large share of what is currently sold as an agent fails this test.

Buy a prompt when you can describe the entire requirement in one sentence and the output format is predictable - a single cold email, one employee survey, one budget breakdown. Buy an agent when a competent specialist would need to ask you questions before starting, or when the right next step depends on what the previous step turned up. If you find yourself pasting the same prompt for the third time this week, you wanted a skill.

1,900+ skills at $19, 60+ prompt packs at $9, from a single-purpose generator to a full role library, and 90+ agents at $39. The gap reflects depth rather than file size: a prompt is one well-built request, a skill is a working method that persists across sessions, and an agent runs an intake and adapts to your answers. Everything carries a 30-day money-back guarantee.

Three places. They compound their own errors, because a wrong step three becomes the foundation for steps four through eleven and the mistake ends up stated confidently. They are expensive per task, since planning, searching, retrying and verifying burns far more tokens than a single answer and vendors meter agentic runs separately. And they are hardest to trust exactly where the stakes are highest - anything irreversible like sending, paying, deleting or publishing. Let the agent prepare the action and keep the final click yourself.

A skill is a ZIP file with SKILL.md at the root. In the Claude desktop app open Customize, then Skills, upload the ZIP and toggle it on; it then loads by itself when a matching task appears. 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 can disable it organisation-wide, which is why the menu is sometimes greyed out on a work account. Where the Skills menu is not available, or in ChatGPT or Gemini, paste the contents of SKILL.md into custom instructions instead. Prompt packs need no installation at all.