Claude Projects vs Artifacts: What to Use When (2026)

These five now live one click apart in the Claude sidebar, and almost nobody can say what separates them. Projects and Artifacts are the pair people ask about most, because both feel like "the place my work lives." Then Skills arrives with instructions of its own, Cowork with tasks and folders, and plain chats keep working well enough that you wonder what configuring nothing costs you.

Existing comparison pages fail one of two ways: they describe the older chat-only Projects while you're looking at the redesigned Claude Code version, or they list features without saying which to pick. So this is the decision version: what each thing is as of September 2026, one table on the axes that matter, a section per pairing, a decision tree, and the five misuses that cost the most time.

TL;DR

What each of the five actually is

Project. Anthropic ships two. The redesigned Claude Code Projects, announced September 17, 2026 in "Projects redesigned: from folder to conversation," is one ongoing conversation where Claude "scopes the request, delegates the work, coordinates parallel threads, reviews the outputs, and assembles the finished result." Each thread is a Claude Code cloud session on its own branch; public beta, Pro and Max only. The earlier Projects in claude.ai chat and Cowork are "self-contained workspaces with their own chat histories and knowledge bases" — you upload reference files, write instructions, start chats inside. Every plan has them, free accounts capped at five.

Artifact. Anthropic's definition is the cleanest line here: "anything Claude makes for you that you'd put in front of someone: a design, a deck, a document, a dashboard, or a small interactive tool." It opens beside the conversation, you can edit it, revisit it, share it by link. Available on Free, Pro, Max, Team and Enterprise. Claude Design, Slides and Docs are artifact types — all three in beta on paid plans.

Skill. "Folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks" — loaded only when a task makes them relevant. On every plan including Free, in beta for Claude Code and the API, and an open standard at agentskills.io.

Cowork. The agentic surface: it "uses the same agentic architecture that powers Claude Code, with no terminal required," running tasks in the cloud on paid plans. Two things reframe it. Cowork tasks and chats are merging — the notice reads "Claude Cowork is now just Claude," rolling out to Pro and Max — and Cowork has its own projects, workspaces "with their own files, context, instructions, and memory."

Chat. One conversation — useful, disposable, the baseline every other feature extends.

The table that decides it

How each pairing breaks down

Projects vs Artifacts

The pair most people mean, and they are not alternatives. A project is where work lives — repositories, instructions, memory, files in its Library. An artifact is a thing a piece of work produced: one page at a private URL that "updates in place as the session continues." Anthropic's suggested uses: a pull-request walkthrough with annotated diffs, a dashboard built from session data, an investigation timeline filling in as Claude works.

Both fail in opposite directions when neglected: an unused project leaves five unconnected chats each needing the same background retyped; an unused artifact leaves a transcript 40 messages deep when a teammate needed one page.

So: the project for the recurring body of work, the artifact for the single thing you'd send someone. The Library is where both meet — it collects the files you add and what Claude produces.

Projects vs Skills

Anthropic answers this one directly: "Projects provide static background knowledge that's always loaded when you start chats within them. Skills provide specialized procedures that activate dynamically when needed and work everywhere across Claude."

Read "always loaded" as the cost. Project knowledge is your house style guide, API contract, brand rules — material relevant to nearly every message in that workspace. A skill is how to do a thing: run the release checklist, apply the migration pattern, produce the report in your format. Skills use progressive disclosure, so Claude loads only what a task needs — which is why the same procedures parked in project knowledge would quietly eat your context window.

The practical split: knowledge and constraints in the project, procedures in the skill, the skill invoked from inside project threads. Our Claude Skills tutorial walks one from nothing.

Projects vs Cowork

Not comparable, because they aren't the same kind of thing: Cowork runs tasks; a project organises them. Cowork even contains its own projects — workspaces with their own instructions, context, memory and scheduled tasks, memory "scoped to the project, so what Claude learns in one project doesn't carry over to others."

The line is blurring on purpose: Anthropic is folding the Chat/Cowork switch away ("Claude Cowork is now just Claude"), so within months the honest framing is three things — conversations, projects, artifacts — with Cowork as the engine underneath rather than a tab you pick.

One boundary still matters: Cowork reaches your machine, Claude Code project threads do not. A thread runs in an Anthropic-hosted sandbox against GitHub repositories and the files, folders and Google Drive folders you upload — not a local database or a service behind your VPN. If the work only exists locally, that's a Cowork or local-session job, and for the model-cost side see our Claude Cowork + DeepSeek guide.

Projects vs chats

The most useful sentence in the projects documentation is a warning: "Context is not shared across chats within a project unless the information is added into the project knowledge base." Two chats in the same project do not remember each other — which trips almost everyone, because the sidebar makes a project look like a folder of conversations.

Chats are for work that ends; projects for work that keeps producing tasks, each with its own memory kept separate from your non-project chats. The trigger is simple: the moment you've re-explained three facts in a fourth chat, move them into project knowledge or instructions — and in Claude Code, open a project instead of another session.

Decision tree

Is the result a page you'd hand to someone?
  yes -> Artifact (built from wherever the work lives)
  no  -> Is it background Claude needs in nearly every message
         of a recurring area of work?
           yes -> Project knowledge + project instructions
           no  -> Is it a procedure for HOW to do one kind of task?
                    yes -> Skill
                    no  -> Does the work outlast one session and keep
                           producing tasks?
                             yes -> Project (Claude Code beta: Pro/Max only)
                                    needs local files or your VPN?
                                      -> Cowork or a local session
                                    only repeats on a schedule?
                                      -> a routine on its own
                             no  -> Chat. Configure nothing.

Five misuses

1. A project as a filing cabinet. Twenty files, no instructions. Knowledge is retrieved by relevance, so a pile mostly sits unused. Write the brief first; add files when a thread demonstrably needed one.
2. Procedures in project instructions. Anthropic's size guidance for the sibling mechanism, CLAUDE.md, is "target under 200 lines" because longer files reduce adherence. Project instructions accept 16,000 characters; spending them on multi-step how-tos is exactly how you get rules Claude follows inconsistently.
3. An artifact as your internal tool. "An artifact is a capture of work: one self-contained page with no backend, so it can't serve multiple routes." It can't authenticate viewers and relative links don't resolve. For a hosted tool with a backend, deploy it yourself.
4. Treating Claude Code Projects as the team feature. "A project belongs to one user" during the beta — no sharing, no org controls. For shared team context, the older chat Projects support Can view / Can edit on Team and Enterprise, or let artifact links carry the output.
5. Leaving everything at defaults. A new project runs every thread on Opus at high effort, and as of September 2026 the documented enforced limit is 200 new threads per day across all your projects — no fixed concurrency cap. Set the thread model deliberately.

FAQ

What is the difference between Projects and Artifacts in one sentence?

A project is a persistent workspace holding context and directing work; an artifact is a single shareable output a conversation produced. One is an input container, the other an output object — which is why "which one" is usually the wrong question.

Are Artifacts being replaced by Projects?

No, and they aren't converging. Artifacts are on every plan including Free and are the delivery format for Claude Design, Slides and Docs; Projects are where work is organised. The docs cross-reference rather than absorb: the redesigned beta's Library "collects the files you add and the artifacts produced by Claude."

Can an artifact read from a project?

Not directly — an artifact is a page, not a worker. What you can do is have a project thread produce or refresh the page, then wire that page to external systems through your account's connectors: on Pro, Max, Team and Enterprise, artifacts read and write to tools like Asana, Google Calendar and Slack over MCP. Artifact storage is capped at 20 MB, text-only, and is deleted if you unpublish.

Is Cowork the same as Claude Code?

Same engine, different surface. Anthropic describes Cowork as using "the same agentic architecture that powers Claude Code, with no terminal required," aimed at knowledge work rather than repositories. Claude Code has the terminal, the IDE and the thread-coordinating Projects beta; Cowork reaches local folders and browsers through the desktop app. See the Cowork setup guide for what it does with a folder.

Which of these cost usage limits?

All of them, but differently. Projects draw on the same plan limits as your other sessions and "running several threads at once uses it faster." Artifacts count toward your plan's usage limits too. A skill costs nothing until a task loads it, then costs context. The cheapest habit: a chat for anything that ends in one sitting.

Do Skills work in Claude Code and Projects?

Yes, with a wrinkle. Skills are in beta for Claude Code, and a project's threads load skills committed under .claude/skills/ in its repositories plus the skills you enable for your claude.ai account. What they don't inherit is anything installed only on your machine — threads are cloud sessions, so local plugins, MCP servers and tools must be routed in deliberately.

I just want to run an artifact outside Claude. Which is that?

Neither — that's an export. The tutorial on running Claude Artifacts locally covers copying the source to a plain HTML file, and the Claude Artifact Downloader does it without opening the code panel. For React components, the React Artifact Viewer runs the JSX directly; What is a React artifact? explains what you're holding.

Related

• Claude Code project examples — what to build in a project
• Claude Projects templates — paste-ready scaffolds
• Claude Artifacts tutorial — building and sharing artifacts
• Claude Skills tutorial — turning a procedure into a Skill
• Run Claude Artifacts locally — getting an artifact out
• Claude Code subagents tutorial — under a thread's work