Configuring a Project

Creating one, and what Claude can actually see

Anthropic describes Projects as self-contained workspaces with their own chat histories and knowledge bases. They are available on every plan, including Free, where you can create up to five. You create one from the Projects area with “+ New Project”, give it a name and a description, and on Team and Enterprise plans choose whether it is private or visible to the organisation.

Instructions are set with “Set project instructions”, and they apply to every chat started inside that Project. Knowledge is added with the “+” button on the project’s main page: documents, text files, and snippets you paste in. Both are configured once and then quietly do their work each time anyone opens a new chat in the room.

Instructions: the standing brief

Project instructions answer a single question: what should be true of every answer produced in this Project, regardless of the specific task? That is a narrower question than it sounds, and the narrowness is what keeps instructions useful. Anything that applies only to one task belongs in that task’s prompt. Anything that applies to all your work everywhere belongs in your account-level instructions instead — the layering is covered in 5.3.

This project is for our marketing team.
Be helpful and write in our brand voice.

This Project produces campaign copy for
a B2B software firm selling to finance
teams at mid-sized companies.

Always: product names and capitals
from the brand guide here; plain
language; UK spelling and dates.

Never: a customer result that is not
in the approved claims document; a
name not on the reference list.

If the knowledge here does not cover
something, say "not covered in project
knowledge" rather than filling the gap.

Four habits make instructions work. Write what to do rather than what to avoid, because positive instructions give Claude something to aim at. Say why a rule exists when the reason is not obvious, since Claude generalises from explanations. Keep them short enough that a new colleague would read them. And include one instruction about gaps — something that turns “the knowledge does not say” into a visible line rather than an invented answer.

Knowledge: what to put on the shelf

Project knowledge is the material every chat in this Project might need. The limits are worth knowing precisely: files added to a project are capped at 30MB each — smaller than the 500MB allowed for a file uploaded into a single chat — and the number of files is not capped, but the total content has to fit within Claude’s context window. Supported document formats include PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB and JSON, with XLSX requiring code execution and file creation to be enabled.

How the content is read matters as much as whether it fits. PDFs of up to 100 pages are analysed for both text and visual elements such as charts; longer ones, up to a 1,000-page limit, are processed as text only. For formats such as Word documents Claude extracts text only and cannot interpret embedded images — so a policy whose key table is a screenshot is, in effect, not in your knowledge base at all.

The last branch deserves emphasis because it is the mistake teams make when they are being thorough. Loading last year’s handbook alongside this year’s does not give Claude more context; it gives it two conflicting sources with no way to tell which governs. The same goes for three drafts of the same policy. Curate deliberately: one current version of each thing, dated in the filename where versions matter.

What happens when the shelf gets full

On the paid plans — Pro, Max, Team and Enterprise — Claude handles a large knowledge base differently. When a project’s knowledge approaches or exceeds the context window, retrieval-augmented generation switches on automatically, expanding capacity by up to ten times. Instead of holding every document at once, Claude uses a project knowledge search tool to find the most relevant passages for your question. You will see it searching, and no configuration is required; if the knowledge shrinks below the threshold, the Project can revert to the standard approach.

This changes how you should write and name things. Anthropic’s guidance for RAG-mode projects is to upload comprehensive material up front, use clear descriptive filenames, group related documents together, and name the specific document in your question when you know which one holds the answer. The same instinct appears in Anthropic’s long-context prompting guidance, which recommends asking Claude to quote the relevant passages before answering, so the answer is anchored to text you can check.

Two final points of scope. Sharing a Project with colleagues, with view or edit permissions, is a Team and Enterprise capability, and administrators can disable sharing entirely. And each Project keeps its own memory space and its own project summary, separate from your other work — which is part of why a Project is the right container for a body of work rather than a folder of chats. Keeping all of this current over time is 5.4.

Traps the wrong answers are built from

You should now be able to

• Create a Project and set instructions that apply to every chat within it.
• Decide what belongs in project instructions, in a single prompt, or in account-level instructions.
• Curate project knowledge to current, clearly named documents within the file limits.
• Explain how RAG mode changes retrieval in a large project and how to name files for it.
• Write an instruction that makes uncovered questions visible instead of invented.
• Check a Project’s readiness before sharing it with a team.