Claude for sales: Projects, knowledge, and data controls

Setting up Claude for sales work comes down to four moves: build a Project that carries your company context into every conversation, load a small set of documents as project knowledge, set the data controls that keep deal information where it belongs, and understand what sharing and memory actually do on your plan. Budget about an hour. After it, every session starts already briefed on your product, your ICP, and your rules — instead of starting from zero.

Before you start

Audience: individual seller — no administrator required for anything in the core setup. Team and Enterprise sharing notes are marked where they apply.
Plan requirements: Claude Projects work on the free plan; Free accounts currently have a documented maximum of five projects.
Prerequisites: a claude.ai account, your core sales documents in any common format, and ten honest minutes on what you actually sell and to whom.

Step 1: create your sales project and write its instructions

Go to claude.ai, open Projects, and create a new project — name it for the job, not the tool: "B2B sales — [your company]." Project instructions are standing orders that apply to every conversation inside the project. Write them once, specifically. A working starting block, customized:

Project instructions — starting block

"You are a sales assistant for [COMPANY]. We sell [PRODUCT in one to two sentences]. ICP: [COMPANY SIZE, INDUSTRY, BUYER TITLES]. Competitors: [NAMES]. Differentiators: [ONE, TWO]. Average deal [AMOUNT], typical cycle [LENGTH], CRM is [NAME]. Writing rules: direct, peer-to-peer; emails 100–150 words ending in a specific proposed next step; never 'synergy,' 'circle back,' 'touch base,' or 'I hope this finds you well.' Proposals follow situation, problem, solution, proof, investment. Never invent statistics, customer names, or capabilities. When my request is ambiguous, ask one clarifying question instead of guessing."

The last two sentences do more work than everything above them. An assistant that is forbidden to invent and allowed to ask is the difference between drafts you edit and drafts you distrust.

What does not belong in project instructions. One-time task directions belong in the chat prompt, not the standing orders — an instruction written for today's email will quietly distort every conversation after it. Confidential customer data does not belong here either; the knowledge base and the data controls in Step 3 govern that. And resist being exhaustive: keep the block under roughly 1,500 words. Claude re-reads it on every single message, so a focused instruction outperforms a complete one.

Step 2: load project knowledge — small and current beats large and stale

Add documents through the knowledge panel inside the project. Start with five: a one-to-two-page product overview, your ICP definition, two or three case studies, a pricing or packaging overview, and one strong past proposal. Add competitive battlecards if you have them. Two behaviors worth knowing. By default, chats inside a project do not share context with one another — a fact established in one conversation does not travel to the next unless it lives in the knowledge base. If you enable Claude's memory features, project memory can carry context between a project's chats; check your memory settings and decide that deliberately, because the knowledge base remains the mechanism you fully control. Second, when knowledge grows large, Claude switches to a retrieval mode that pulls relevant sections per request — on paid plans Anthropic documents expansion up to ten times the working context. Anthropic's pages currently differ on whether that expansion applies on the free plan, so treat it as a paid-plan guarantee and a free-plan possibility. Either way, retrieval works best when each document is clearly titled and single-purpose.

Create separate projects per workflow once the first one earns its keep — outreach, discovery and qualification, proposals, account research. Separate projects keep each instruction set sharp instead of blended, and on the free plan's five-project cap, those four plus a sandbox is exactly the budget.

Step 3: set data controls and understand permissions

Before deal data goes in, spend five minutes in settings. Review the privacy controls for whether your conversations may be used for model training and set them to your comfort and your company's policy — on Team and Enterprise plans, organizational controls apply and your admin owns the defaults. Know the sharing model before you assume it: full project sharing is a Team and Enterprise capability; on individual plans, a project is yours alone, and "sharing a project with a colleague" means they build their own from your instruction text. Each project also keeps its own separate memory space, so context from your pipeline-review project does not bleed into your outreach project.

What not to paste is its own discipline — customer identifiers, unreleased pricing, and anything under NDA deserve the same caution here as in any tool. The standing rules live on the best-practices page; read them once and they generalize to every AI tool you touch.

Step 4: connections and where automation lives

Claude connects to CRMs and other tools through connectors — HubSpot, for example, ships an official one-click connector. Connected, Claude can read live pipeline data instead of pasted exports. Two routing rules keep this simple: persistent connector configuration belongs here on the AI tool setup page, and a dedicated CRM-connection setup article is planned for it; scheduled, triggered, or integrated recipes built on those connectors live on the automated workflows page. Set up the foundations in this article first; connections compound them later.