如何為 AI Agents 設定 Claude Projects（以及何時改用 Skill）

Anthropic's own distinction settles most cases: projects provide static background knowledge that is always loaded within that project, while skills provide specialized procedures that activate dynamically when needed and work everywhere across Claude. So a body of work - a client, a codebase, a matter with documents attached - wants a project. A method you will want in arbitrary conversations - a review discipline, a drafting standard, a format - wants a skill. The common mistake runs one way: people put methods in projects, then need them in an ordinary conversation and find themselves in the wrong room.

Yes, with a limit. Anthropic states that projects are available to all users including free accounts, and that free users can create a maximum of five projects. That matters because the usual advice of one project per agent runs out at five very quickly once real client and codebase projects are competing for the same slots. Enhanced project knowledge with retrieval is paid-only, and sharing or collaborating on projects is Team and Enterprise.

Instructions are always loaded and shape behaviour in every message of every conversation in the project, which means they permanently occupy room in the context window. Knowledge is material Claude draws on - documents, text, code or other files uploaded so it can better understand the context and background for individual chats - and on paid plans it is retrieved rather than held open. So put procedure, rules, tone and scope in instructions, and reference material in knowledge. A long agent file does not automatically belong in instructions.

No. Once project instructions are saved they apply to every conversation you start inside that project, with no expiry and no reloading. What does not happen automatically is updating: there is no update mechanism, so when you replace an agent file you replace it by hand in every project that uses it. A project you forget will keep quietly running last quarter's instructions, which is why it is worth noting which version is installed where.

Not with hello, because a general assistant answers that politely too and proves nothing. Ask something only the configured version could answer: what will you ask me before you start, what is out of scope for you, what does our style guide say about headings. If the reply is generic, either the instructions did not save or you are in a conversation started outside the project, which is the second most common setup error after pasting the skill file into the chat window instead.

Anthropic's own pages currently disagree, so check rather than trust either. The help centre articles on skills and on creating custom skills both state that Skills are available for users on Free, Pro, Max, Team and Enterprise plans, and that the feature requires code execution to be enabled. The Claude Academy tutorial on the same subject says Skills are available on Pro, Max, Team and Enterprise, with no mention of free. If you are on the free plan, open Settings then Capabilities and see whether Skills appear for your account.

You can, but it usually works badly, because project instructions are a single always-loaded block and two personas in one block compete. The better arrangement is one project per body of work, with the method installed as a skill so it travels into whichever project the work is in. If the agents genuinely need separate always-on configuration and separate document sets, they need separate projects - and on a free account that is a five-slot budget to spend carefully.

Three, in order. Pasting the skill file into the chat window instead of project instructions, which works for one conversation and is gone from the next. Starting a conversation outside the project, which produces a confident generic answer that looks fine. And building an empty project with no documents uploaded, which is all of the setup and none of the benefit - if there is no knowledge to attach, the agent probably wanted to be a skill.