Claude Projects vs Skills vs Artifacts vs Connectors: When to Use Each

September 27, 2026 · Claude Features Explained: Projects, Skills, Artifacts, Connectors & MCP (part 1)

▶ Watch on YouTube & subscribe to The Stack Underflow

“Which Claude should I use?” hides two separate decisions. The first is the model: models differ, and so does what each plan includes. The second is the layer: where your context, your procedures, your work and your access should live. This page is about the second decision.

The one-line version: pick the model for the thinking. Pick the layer for the job.

Last verified against the Claude Help Center: 27 September 2026. Claude’s features change often. Every claim below links to its source at the bottom of the page, and the change notes list what we’ve corrected.

Four layers, four problems

“Layers” is our teaching frame, not an official Anthropic term. Everything starts in a chat, and four features each solve a different problem:

The one distinction worth memorising: a project instruction is a rule for this body of work; a skill is a procedure across all of them. It’s a decision heuristic, not a formal product definition.

What Claude actually “remembers”: four different things

People say “Claude remembers” and mean four different mechanisms:

When someone says Claude remembers something, ask which one.

Projects: context for one body of work

A Project is a workspace for one body of work. In the current chat version, you add documents and project instructions once, and that context is available across the chats in that project. On paid plans, when project knowledge approaches context limits, Claude switches to retrieval (RAG) instead of loading everything. Free accounts can create up to five projects.

Heads-up: Anthropic is rolling out a new version of Projects in stages, starting with Claude Code, with chat, Cowork, Team and Enterprise to follow. The mechanics here describe the current chat version. The durable idea is scope: project context belongs to one body of work.

The test: if you keep re-explaining the same background for the same piece of work, that background belongs in a Project.

Skills: a procedure you reuse everywhere

At its simplest, a Skill is a prompt you stopped retyping: your report format, your brand rules, your review checklist, written once. Under the hood, a Skill is a folder of instructions that can also include scripts and resources. You don’t paste it in: when a task needs it, Claude reviews the available skills and loads the relevant one, in a plain chat or inside any project. Skills require code execution to be enabled.

On Team and Enterprise, there’s one more place for standing rules: organization instructions, which owners set and which apply to every conversation in the organisation. Use them for org-wide standards; use a Skill for a reusable procedure.

Artifacts: work you keep

An Artifact is a work product (a document, code, a design, a dashboard or a small interactive tool) that opens beside the conversation, where you can keep editing it with Claude and share it. On paid plans, Claude also has beta starting points for Docs, Slides and Design, and an artifact can connect to the apps you’ve connected to Claude and store data.

Artifacts changed on 16 September 2026: chat artifacts made before that date are “legacy”. They keep working and can still be shared, but new ones use the current system.

A chat message stays in the conversation. An artifact is work you keep editing, reusing and sharing.

Connectors and MCP: live access, least privilege

A Connector gives Claude access to apps and services outside it: your documents, your tickets, your calendar. Some connectors are first-party integrations (such as Google Drive or Slack), and Claude also supports MCP, the Model Context Protocol, for compatible services and custom connectors. MCP is an open standard for connecting AI applications to tools and data.

The part that matters for safety:

• A connector works with the permissions of the account you connect, and it can read data or take actions.
• So grant the least access the job needs: read-only if the integration offers it.
• On Team and Enterprise, an owner has to enable connectors before members can use them.
• Never paste passwords or keys into a chat. Connectors authenticate separately.

They combine

Real work uses the layers together:

PROJECT: client context  ─┐
SKILL:   weekly report    ├──►  ARTIFACT: the report
CONNECTOR: live tickets  ─┘      (kept and shared)
          (least access)

The rule:

• Context for one body of work → Project
• A procedure you repeat everywhere → Skill
• A work product you keep → Artifact
• Live access to an external app or service → Connector, with the least access the job needs

What depends on your plan

The video leaves out plan tables on purpose: plan details change more often than videos can. Here is what the features above depend on, as of 27 September 2026. Check Anthropic’s pricing page before you buy anything.

Pause & Prove: the three questions, fully explained

The video explains the right answer and the most tempting wrong one. Here are all four options for each.

1. A consultant pastes the same client brief, meeting notes and last month’s decisions into every new chat. Which layer fixes this?

• A. Claude’s memory. Tempting: memory picks up useful context as you chat, but it isn’t where you keep a client’s documents for Claude to work from.
• B. A Project. ✓ This is background for one body of work. Upload the brief and notes to a project once, and every chat in that project has them as shared context.
• C. An Artifact. An artifact is work Claude produces; the brief is input you give it.
• D. A Connector. A connector gives live access to outside apps and services. A static client brief doesn’t need that.

2. You write weekly reports for different clients, each in its own project. Every report must follow the same steps, structure and brand rules. Which layer should hold that reusable procedure?

• A. A Project. Tempting: project instructions are rules for that body of work, and these rules have to apply across all of them.
• B. An Artifact. The report can be an artifact, but that’s the output, not the procedure for making it.
• C. A Connector. The rules aren’t data in an outside system, so there’s nothing to connect to.
• D. A Skill. ✓ A reusable report procedure: the steps, structure and brand rules travel across projects. A skill applies across chats and projects, loaded when the task needs it.

3. Your support team wants Claude to answer questions from the live tickets in your company’s tracker. Which setup is right and safe?

• A. Weekly export to a Project. It works, but it goes stale. The question asked for live tickets.
• B. Paste tickets into the chat. You become the integration, and you’re always a ticket behind.
• C. A connector with least privilege. ✓ Live data, with only the permissions the job needs, read-only if the integration offers it. On Team and Enterprise, an owner enables connectors first.
• D. A connector on an admin account. Tempting: right layer, wrong grant. The connector acts with that account’s permissions, far more than reading tickets needs.

Sources

Every claim above was checked against these pages on 27 September 2026:

• What are projects?: projects, knowledge, RAG, the new Projects rollout
• What are skills?: folders of instructions, scripts and resources; code execution
• What are artifacts and how do I use them?: artifacts, templates, apps, storage, 16 September 2026 change
• Use Claude’s chat search and memory: chat search and memory availability and controls
• How do usage and length limits work?: automatic context management
• Use connectors to extend Claude’s capabilities: connectors, owner enablement
• Custom connectors using remote MCP: MCP
• Set organization instructions: org-wide instructions on Team/Enterprise
• Claude Fable models on your plan: model availability by plan

Change notes

• 27 Sep 2026: first published. Before release, two outside reviews corrected the draft: “every tier is the same model” was wrong (model availability differs by plan); memory and chat search exist beyond a single conversation; artifacts have no documented version history; connectors include first-party integrations as well as MCP; the Projects section now notes the rollout of the new Projects version.

Not affiliated with or endorsed by Anthropic. Found a mistake? Tell us in the video’s comments and we’ll correct this page.