Domain 5 · Task Statement 5.1

Projects in Cowork

TL;DR

Create scoped workspaces with isolated chat histories, configure Project Instructions as persistent system prompts, manage the 200K-token Knowledge Base, and control access with permission levels.

What You Need to Know

Without Projects, every Cowork conversation exists in a flat, undifferentiated stream. Your marketing brainstorm, your legal contract review, your quarterly planning — all blended into the same context, with Claude potentially applying the wrong tone, referencing irrelevant data, or confusing separate initiatives. Projects fix this by creating self-contained workspaces, each with its own chat history, instructions, and knowledge base.

Think of Projects as dedicated office drawers. Instead of having all your paperwork in one messy pile, you have one drawer for Marketing, another for Legal, another for Quarterly Planning. When you open the Marketing drawer, Claude only sees the branding guide and previous campaign notes. It doesn't peek into the Legal drawer. This isolation isn't just organisational convenience — it's what prevents context bleed between workstreams.

Project Instructions: the persistent system prompt

Project Instructions are background rules that apply automatically to every conversation within a specific project. You write them once, and they shape Claude's behaviour for every chat in that workspace — without you needing to repeat context, role definitions, or formatting rules at the start of each session.

This is the mechanism that transforms Claude from a generic assistant into a specialist. A project with the instruction "You are a senior B2B SaaS marketing strategist. Use British English. Reference the brand guide in the Knowledge Base for all tone decisions" produces dramatically different output from one with "You are a legal compliance reviewer. Flag any clause that creates liability exceeding £500K."

[!]

Exam Trap: Instructions Consume Context Tokens

Project Instructions count towards the context window, not your message quota. A lengthy set of instructions leaves less room for the actual conversation. If your instructions exceed 1,000 words, audit them for redundancy — every line must earn its place.

The Knowledge Base: grounding Claude in your data

Each project can hold up to 200K tokens of uploaded documents — PDFs, text files, reference material — that Claude references when answering questions. This is how you give Claude access to your company's actual data rather than relying on its general training knowledge.

The Knowledge Base isn't a filing cabinet Claude searches on demand. It's loaded into context for every conversation in the project, which is why the 200K token limit exists. Upload a 50-page brand guide and Claude references it automatically. Upload your company's product catalogue and Claude can answer questions about specific products, pricing, and features without you pasting anything into the conversation.

Context isolation: preventing cross-project bleed

Each project maintains a separate memory space. Information discussed in Project A doesn't leak into Project B. This is essential for:

  • Confidentiality — sensitive HR data stays in the HR project, not surfacing in marketing conversations
  • Accuracy — legal terminology stays in the legal project, not contaminating a casual team update
  • Consistency — each project's instructions and knowledge apply only within that project's boundaries

Without isolation, a conversation about salary negotiations in one project could influence Claude's behaviour in a public-facing content project. Projects prevent this by design.

Permission levels for shared projects

When you share a project with colleagues, two access levels are available:

  • Can View — participants can chat within the project but can't modify the instructions, knowledge base, or settings
  • Can Edit — participants have full access to modify instructions, update the knowledge base, and change project configuration
[~]

Default to Can View

When sharing a project, default to Can View unless the person needs to modify the configuration. This prevents accidental changes to carefully tuned instructions and ensures the project behaves consistently for everyone who uses it.

An important nuance: chats within a shared project are private by default. Sharing a project doesn't expose your conversation history to other members. Each person has their own private chat stream within the shared workspace — they share the instructions and knowledge base, not the conversations.

Common Mistakes

Common Mistake

Cramming marketing, sales, HR, legal, and personal tasks into a single 'Everything' project because it feels simpler than creating separate ones.

Instead: Create one project per distinct workstream. Each project gets focused instructions and relevant knowledge documents. When contexts are isolated, Claude applies the right tone, terminology, and rules to each task without confusion.

Common Mistake

Writing vague Project Instructions like 'be professional' or 'write well' — phrases that Claude interprets differently every time.

Instead: Write specific, measurable rules: 'Use British English. Limit paragraphs to 3 sentences. Always include a Risk Rating header at the top of every response. Reference the brand guide for tone decisions.' Concrete rules produce consistent output.

Common Mistake

Uploading documents to the Knowledge Base once and never updating them — leaving Claude confidently citing outdated pricing, deprecated features, or former team members.

Instead: Treat the Knowledge Base as a living document. Set a calendar reminder to review uploaded files quarterly. When business strategies, product details, or team structures change, update the documents so Claude's references stay accurate.

Starting a project-based task

Before

Help me with my project.

After

Based on our project's technical specs in the Knowledge Base, identify three risks in the proposed architecture and rank them by business impact.

Using uploaded documents effectively

Before

Summarise the document I uploaded.

After

Using the brand guide in our Knowledge Base, audit this draft press release for tone violations. Flag any sentences that contradict our messaging framework and suggest replacements.

Hands-On Activity

Hands-On Activity

Build Your First Scoped Project

15 min

Create a project, upload a reference document to the Knowledge Base, configure Project Instructions, and test that Claude's behaviour is shaped by both the instructions and the uploaded data.

What you will learn

  • Create a new Project and understand its isolated workspace model
  • Upload a document to the Knowledge Base and verify Claude can reference it
  • Write specific Project Instructions that shape Claude's behaviour
  • Confirm that project context doesn't leak into other conversations
  1. 01

    Open the Projects sidebar in Claude and select 'New Project'. Name it 'Brand Voice Test'.

    Why: Creating a dedicated project establishes an isolated workspace. Everything you configure here stays contained — it won't affect your other conversations.

    Expected: A new empty project appears in your sidebar with its own chat area and settings panel.

  2. 02

    Upload a document to the Project Knowledge section — use a style guide, a brand document, or any reference file you have handy.

    Why: The Knowledge Base is what grounds Claude in your specific data rather than its general training. This is how you give Claude access to information it couldn't otherwise know.

    Expected: The document appears in the Knowledge section with a token count showing how much of the 200K limit it uses.

  3. 03

    Click the settings icon and paste these Project Instructions: 'You are a brand specialist. Always reference the uploaded brand guide for tone and terminology. Use British English. Keep responses under 200 words unless asked for more detail.'

    Why: These instructions act as a persistent system prompt. Every new chat in this project will automatically follow these rules without you needing to repeat them.

    Expected: The instructions saved. The settings panel shows your custom instructions are active.

  4. 04

    Start a new chat within the project and ask: 'What is our primary brand colour?' followed by 'Write a 50-word product description using our brand voice.'

    Why: This tests both the Knowledge Base (retrieving specific data from your document) and the Project Instructions (following tone and formatting rules).

    Expected: Claude references your uploaded document for the colour answer and follows your brand voice instructions for the product description, all without you needing to repeat any context.

Practice Question

Practice Question

A Project Manager wants the AI to always use the company's 'Agile-Hybrid' methodology and reference the specific '2026 Tech Stack' document for every new chat. What is the most effective approach?

Sources