Projects: Persistent Workspaces
Beginner
A Project is a dedicated workspace in Claude.ai that bundles its own files, knowledge, and instructions. Instead of re-uploading the same documents and re-explaining context every chat, you set it up once — and every conversation in the Project starts already informed.
- What a Project bundles, and why that beats re-pasting context every chat
- The four-step setup, and the one step people skip
- How to curate project knowledge so retrieval stays sharp
- How this Project differs from the coordinator-and-threads Projects in Claude Code
Why use a Project
- Grounded answers. Add your documents (a handbook, specs, notes) and Claude answers from them — a built-in, no-code flavor of RAG.
- Persistent context. Project instructions act like a scoped system prompt for everything inside it.
- Organized. All the chats about one topic/client/initiative live together.
Set one up
Guided walkthrough1 of 4
- One Project per topic, client, or initiative. A Project that covers three unrelated things retrieves badly for all three.
- The files and text it should always know — the handbook, the spec, the style guide. This is what Claude answers from.
- The step people skip. Role, conventions, what to do and what to avoid. Without it, you have a search index rather than a colleague.
- Every conversation inside the Project inherits the knowledge and the instructions automatically.
Great use cases
- A client/account workspace (their docs + your notes).
- A codebase or product knowledge base for Q&A.
- A writing project with your style guide and past pieces (so drafts match your voice).
- Study for a course, with the syllabus and materials loaded.
A starting point for project instructions
Project instructions skeleton
You are helping with [project/client name]. Context: [what this project is, who it serves, current stage] Always: - Answer from the project knowledge first; say so when you are going beyond it. - Use [our terminology: list the 3-5 terms that matter]. - Match the tone in [named file in the knowledge base]. Never: - Invent figures, dates, or quotes. If it is not in the knowledge, say it is not in the knowledge. - [Project-specific constraint: e.g. reference the deprecated v1 API.]
Tips
- Curate the knowledge — relevant, current files beat dumping everything (noise hurts retrieval).
- Keep instructions tight and true (same rule as custom instructions).
- Prune when things change. A superseded spec left in the knowledge base is worse than no spec, because Claude has no way to know it lost an argument with the new one.
- Don't add sensitive data you're not comfortable storing — see Privacy.
Two things called "Projects"
They are genuinely different products, and the names overlap during the rollout:
| This Project (Claude.ai) | Claude Code Projects | |
|---|---|---|
| What it is | A workspace with knowledge + instructions | A coordinator that delegates to parallel threads |
| What it does | Grounds your chats in your files | Scopes, splits, and executes work for you |
| Runs code? | No | Yes — each thread is a cloud session on its own branch |
| Who has it | Broadly available | Beta, select Pro/Max Claude Code users at launch |
If you are reading about coordinators, threads, and pull requests, that is the Claude Code one.
Check yourself
0/2- A Project bundles knowledge plus instructions so every chat inside it starts informed.
- Instructions are the step people skip — without them you have search, not a colleague.
- Curate and prune: superseded files actively mislead, they do not just add noise.
- The coordinator-and-threads Projects in Claude Code is a different product that shares the name.