Put them in a Claude Project. A project holds files and instructions that Claude reads as background for every chat inside it, and on the Team and Enterprise plans you can share one project with everyone in the organisation. That gets the facts in front of Claude. It doesn’t tell you whether Claude used them, whether they’re still true, or who changed them, and those are the three things that go wrong.
Setting up a project
Projects are on every Claude plan, including Free, which allows five. Create one, give it a name, and add two things.
The first is project knowledge: the files Claude reads. Each file can be up to 30MB, and you can add as many as fit in Claude’s context window. On paid plans, when the knowledge gets near that limit Claude switches to retrieval, which Anthropic says expands a project’s capacity by up to ten times by loading only the parts relevant to each chat.
The second is project instructions. These apply to every chat in the project and, Anthropic notes, only to chats within that project. This is where the voice description goes, and where you tell Claude how to treat the facts: use only these figures, name only these clients, say when something isn’t in the files.
One thing Anthropic’s pages make clear: a chat’s context doesn’t carry to other chats in the project. If Claude learns something useful in one conversation, it’s gone in the next unless someone adds it to the knowledge.
What to put in the facts file
Keep it to what you’d defend in front of a client. What the company does and for whom. The figures you can stand behind, each with where it came from and the date. The clients you’re allowed to name, and the ones you’re not. Credentials, prices, the words you use for your own products and the words you don’t.
Write the source and the date against every figure, even though Claude won’t ask for them. In six months you’ll need to know whether “more than 100 projects” was counted in January or last week, and nothing in the project will tell you.
Sharing it with the team
On Claude’s Team and Enterprise plans a project can be public to everyone in the organisation or private to invited members. Members get one of two roles. “Can view” lets them use the project’s knowledge and instructions and chat inside it. “Can edit” lets them change the instructions and the knowledge and manage who’s in.
Chats stay private even in a shared project, and an owner can switch project sharing off for the whole organisation. Anthropic’s pages don’t describe any approval step or version history for changes to a project’s instructions or files, so an edit by anyone with edit access is live for everyone at once.
On Free, Pro and Max, each person’s projects are their own.
Where it stops working
Claude reads the facts, and nothing reads the draft back against them. If a figure in a proposal doesn’t match the figure in the project, nobody finds out unless they check by hand. Anthropic’s pages describe how project knowledge is loaded; they don’t describe any check of what Claude writes against what the files say.
Facts go stale without telling you. The file says what it said on the day someone uploaded it, and Claude will use it with the same confidence in a year.
And the project stays in Claude. A colleague who writes in ChatGPT or Copilot has none of it, and when they switch tools they set it up again from memory.
What people get wrong
They upload the whole website and the whole brand book and assume Claude will find the right bit. Retrieval helps, but a short, dated list of the facts you want used beats a hundred pages of the facts that happen to exist. They also forget that Claude’s memory belongs to the account, so a correction one person makes in conversation never reaches anyone else.
How BeigeCheck does it
BeigeCheck holds your company’s facts in six areas, each fact with its source and version, and Claude reads them through a read-only connection whenever you name BeigeCheck in what you ask. The same approved facts reach ChatGPT and, on Business and above, Microsoft 365 Copilot.
Nothing changes without a proposal and an approver, and every approved change is a version with a name and a date. When a draft comes back, the Beige-ometer checks every claim in it against the facts you’ve approved and flags the ones they don’t support, before it goes anywhere.
It also proposes new facts from your own documents and from the questions your people answer, so the record grows without anyone maintaining a file by hand.