Use a project when a set of chats needs the same background knowledge, a skill when Claude needs a repeatable procedure that should load only for a certain task, custom instructions for broad preferences that apply everywhere, and MCP when Claude needs live access to an outside service. Anthropic describes these as complementary.
The short comparison
Anthropic's help article on Skills draws the lines directly.
- Projects provide static background knowledge that is always loaded in the chats of that project.
- Skills activate only when needed and work across Claude.
- Custom instructions apply broadly to all conversations.
- MCP connects Claude to external services and data, while skills give procedural guidance on using them.
What a skill actually is
The help center describes skills as folders of instructions, scripts and resources that Claude loads dynamically. Claude checks which available skills are relevant to a request, loads only what the task needs, and applies the instructions. The page calls this progressive disclosure and says it helps keep the context window from overloading.
Skills come in several kinds: skills built and maintained by Anthropic (for example document handling for Excel, Word, PowerPoint and PDF), custom skills you or your organization write, organization provisioned skills on Team and Enterprise, and partner skills built to work with those partners' connectors. Skills require code execution to be enabled, and they are listed as available on Free, Pro, Max, Team and Enterprise plans.
A simple decision guide
- Is it knowledge Claude should always have in this body of work, like a product brief? Put it in a project.
- Is it a method, like "build the weekly report in this format using these scripts"? Make it a skill.
- Is it a preference, like tone or language? Use custom instructions.
- Does Claude need to read or change something in another system? Add a connector.
A caution on skills
A skill can include executable scripts, so treat a skill like code you are installing. The help page mentions one safeguard: Enterprise organizations can turn on skill and plugin scanning to check uploads for malicious content. It does not list other protections, so review third party skills yourself before enabling them. Our post on Claude Code and AI coding agent security basics covers the habit in more detail.
Open format
The Agent Skills specification is published at agentskills.io, and the help center says the format is not limited to Claude. That makes a well written skill more portable than a pile of pasted prompts.
Safeguard can scan the repositories and dependencies a skill touches, using the free scan at gold.safeguard.sh/scan.
Quick check
Do skills replace projects?
No. Anthropic describes projects as static background knowledge and skills as task specific procedures that load when relevant.
Are skills only for Claude Code?
No. They also work in Claude Code, and are in beta for API users of the code execution tool, according to the help article.
Sources
Last reviewed: 19 October 2026.
