Claude skills vs. projects vs. instructions: which should you use?

Claude gives you four ways to stop repeating yourself: account-wide instructions, Projects, memory and skills. They overlap enough to be confusing, and putting something in the wrong place means Claude either ignores it when you need it or drags it into chats where it doesn't belong.

The short version: instructions are for what's always true about you, Projects are for one piece of ongoing work and its documents, memory is what Claude picks up on its own, and skills are for tasks you repeat that need to be done a specific way.

The short answer

Instructions: what's always true about you

Your instructions apply to every conversation. To set them, open Settings, choose Your Claude, and fill in Instructions. (In the older Settings layout, it's Account, then Instructions for Claude.)

Good instructions are short and true everywhere: "I'm a second-year nursing student. Explain things with clinical examples, and keep answers brief unless I ask for detail."

What doesn't belong there: step-by-step procedures for one task, like how to format your meeting minutes. Instructions are part of every chat, so a long procedure for one job would come along to all the others. That's what skills are for.

Projects: one piece of work and its documents

A Project is a workspace with its own chats, its own knowledge (documents, text and code you upload) and its own instructions, which you add with Set project instructions. Claude uses those only in chats inside the project. Projects are available on every plan; free accounts can have up to five.

Projects fit work that comes with a set of documents: a course with its syllabus and lecture slides, a client with their brand guidelines, a thesis with its sources and your advisor's feedback.

Two things to know:

• Chats in a project don't see each other. Only what's in the project's knowledge is shared between them.
• Each project has its own memory. To keep it completely separate from the rest of your account, turn off Use account memory on the project's page.

Projects are less suited to a task you do across many kinds of work. If you'd have to paste the same instructions into five projects, that's a sign it should be a skill. (Anthropic is also rolling out a new version of Projects in stages, so the screens may look different for you.)

Memory: what Claude picks up on its own

With memory on, Claude saves useful context as you chat, as separate topics: your role, the projects you're working on, people you mention, how you like things done. You can also say "remember this." By default it leaves out sensitive topics like health, politics and religion.

Memory is on by default on the Free, Pro, Max and Ultra plans; on Team and Enterprise plans, owners decide. To see or edit what Claude remembers, go to Settings → Your Claude and choose Manage next to Your memory. To have a chat that isn't saved to memory or history at all, start an incognito chat with the ghost icon in the upper right.

Memory is great for context, but it isn't a way to define how a task should be done. You don't decide exactly what gets saved, or when it's used. If you need the same steps or format every time, write them down as a skill.

Skills: a task done the same way every time

A skill is a folder with a SKILL.md file of instructions, and sometimes templates, reference files or scripts. Claude starts out seeing only each skill's name and a short description. It reads the rest only when a request matches. That means you can install many skills without filling every conversation with instructions.

Skills fit jobs you repeat that need a method:

• meeting minutes in a set format, with no invented owners or due dates (Meeting Minutes);
• citations in APA 7, with missing details flagged instead of guessed (Citation Formatter);
• code reviews that show the input that breaks the code (Code Review).

Skills you turn on are available across your chats, not tied to one project, and they're portable: the same .zip works on claude.ai and in Claude Code, and you can share it with anyone. On claude.ai they need code execution turned on. See how to install a skill and how to write your own.

How they work together

They aren't either-or. Here's how a graduate student might set things up for a thesis:

• Instructions: "I'm a public health grad student. Be direct and skip the pep talk."
• A Project called Thesis, with the proposal, the advisor's latest feedback and the reading list as knowledge, and project instructions with the research question.
• Skills for repeated jobs in and outside the project: Paper Summary for each new paper and Citation Formatter for the reference list.
• Memory picks up the rest along the way, like the defense date or which chapter is next.

Which one should you use?

Ask these in order:

1. Should Claude know this in every chat? Put it in your instructions.
2. Does it belong to one piece of work and its documents? Make a Project.
3. Is it a task you repeat, with steps or a format? Make it a skill, or find one that already exists.
4. Is it background you'd rather not keep up to date yourself? Let memory handle it.

Two signs you need a skill: you keep pasting the same instructions into different chats or projects, or your account instructions have grown into a multi-step procedure for one particular job.

In Claude Code

Claude Code has its own version of instructions: CLAUDE.md files. Claude reads them at the start of every session, from ~/.claude/CLAUDE.md for all your projects and from a CLAUDE.md at the root of each repository. Anthropic suggests keeping one under roughly 200 lines, since every line is loaded every time.

Skills work the same way as on claude.ai. A rule of thumb that follows from how each is loaded: put the commands, conventions and constraints that matter in every session in CLAUDE.md, and put procedures you only need for some tasks, like reviewing a pull request or cutting a release, in a skill.