Configuring & Maintaining Claude Projects.

A Project is distinct from a plain one-off chat: every chat started inside a Project belongs to that Project and draws on a shared knowledge base and shared instructions, so a recurring piece of work, a client account, a role, a workflow, doesn't require re-explaining context every conversation. What's shared is the knowledge and the instructions, not the conversational memory itself - each individual chat inside a Project is still its own separate thread, and one chat's back-and-forth is not automatically visible to another chat in the same Project. Two configurable layers do the sharing. Project knowledge is whatever gets uploaded, documents, text, or code the project can draw on. Project instructions is a standing brief telling Claude how to behave inside that project specifically, tone, role, or industry perspective, separate from any single chat's back-and-forth.

Configuring a Project is only half the objective. The other half, informing, maintaining, and updating that configuration as needs change, is just as tested: a Project set up once and never revisited drifts stale exactly the way a document left unedited does. Swapping an outdated uploaded file for a current one, or editing instructions when a role or engagement changes, is the ongoing maintenance half of this skill, not a one-time task.

Setting up the two layers is a UI action, not engineering. To add knowledge, click "+" on the project's knowledge panel and upload documents, text, or code. To add instructions, click "Set project instructions," write the brief, and click "Save instructions." Projects are available to every user, including free accounts, though free accounts are capped at a maximum of five projects.

Context does not silently carry across chats. A single chat's back-and-forth inside a Project does not automatically become project-wide context for other chats, only content explicitly placed in the knowledge base or the instructions persists across every chat in that Project. This is the single most commonly tested misunderstanding in this domain.

Knowledge capacity scales without manual setup. RAG (retrieval-augmented generation) for Projects is now available on every plan, free, Pro, Max, Team, and Enterprise, and activates automatically once project knowledge approaches its context-window limit, retrieving only the most relevant snippets per query instead of loading the whole knowledge base every turn. This expands effective capacity by up to 10x with no configuration required, and the system can revert to standard processing if the content volume later shrinks.

Instructions exist at three separate tiers, and only one is Project-specific. Account-wide "Instructions for Claude" apply to every conversation regardless of Project. Project instructions apply only inside that one Project. Skills add reusable behaviors or capabilities Claude can invoke, a related but distinct "styles" control also exists for response formatting specifically, so a business user configuring a Project should default to Project instructions for anything scoped to that Project alone, not the account-wide layer.

Maintaining a Project (CCAOF-D5-O4) means keeping instructions concise and swapping stale knowledge, not writing a longer brief. Anthropic's own guidance is that Claude performs best when project instructions carry general context, key guidelines, and Claude's role, while task-specific detail for one particular ask belongs in the individual chat instead, because shorter instructions leave more context budget for the knowledge that actually needs to be retrieved. When an underlying document changes, a renegotiated contract, an updated policy, the maintenance action is replacing the uploaded file and revising instructions if the role or tone shifted too, not layering a correction into a single chat that other chats in the Project will never see.