Skip to content
EN DE

Claude Projects: Your First Persistent Workspace

L3 Lesson 2 of 5 — Context as Infrastructure
1
2
3
4
5

In Lesson 1, you understood why persistent context matters. Now let’s build one. Claude Projects Persistent workspaces in Claude.ai that bundle a knowledge base (uploaded documents), custom instructions, and chat history in one place. Every chat within a project automatically has access to the full project context. are the most concrete implementation of the “explain your job once” concept — a workspace that brings together knowledge, rules, and conversations.

A Project consists of three elements:

ElementWhat It DoesAnalogy
Knowledge BaseDocuments that Claude uses as referenceThe shelf with the handbooks
Custom InstructionsRules for behavior, tone, and formatThe onboarding document
ChatsIndividual conversations within the projectMeetings in the same office

Every chat within a project automatically has access to the knowledge base and instructions. No copying, no explaining — the context is just there.

Important: Individual chats do not share their history with each other. Only the knowledge base and instructions are shared across the project. If you discuss something in Chat A, Chat B doesn’t know about it — unless you add it to the knowledge base.

Go to claude.ai/projects (or via the sidebar → “Projects”) and click ”+ New Project”. Enter a name and brief description.

Note: Claude cannot see the project name or description — they’re for your organization only. Everything you want Claude to know belongs in the Instructions or Knowledge Base.

Click ”+” in the right sidebar and upload files:

File TypeFormatsLimit
DocumentsPDF, DOCX, TXT, CSV, HTML, JSON, EPUB30 MB per file
SpreadsheetsXLSX (requires Code Execution enabled)30 MB per file
ImagesJPEG, PNG, GIF, WebP8000 × 8000 px

What to upload? Anything Claude should reference when answering: style guides, project documentation, competitive analyses, data sheets, templates.

What not to upload? Sensitive data you don’t want to share with the AI provider. Check your plan’s privacy policy first.

Click “Set project instructions” and write your rules. Here’s an example for a content team project:

You are a content strategist for [Company Name].
Context:
- Audience: CTOs and VP Engineering at companies with 100-500 employees
- Tone: authoritative but approachable (see Brand Guide)
- Channels: Blog, LinkedIn, Newsletter
Rules:
- Every blog post draft includes headline and meta description
- Use the uploaded SEO keyword list
- Flag claims that need fact-checking
- Paragraphs max 4 sentences

Begin a new chat within the project. Claude immediately has access to all uploaded files and follows your instructions — without you having to repeat anything.

Upload the documents you’d otherwise copy-paste into every chat: style guides, process descriptions, templates.

Summarize background knowledge in a short document that Claude uses as a foundation: “Who are our customers?”, “What’s our product strategy?”, “How do we use certain terms?”

Update the knowledge base regularly. Remove outdated documents, add new ones. The knowledge base is the source of truth for the project — if it’s outdated, the answers will be too.

You are [role] for [context].
Rules:
- [Rule 1]
- [Rule 2]
- [Rule 3]

Sufficient for most use cases. Clear, concise, maintainable.

For every analysis, follow this process:
1. Summary (2-3 sentences)
2. Key findings (3 points)
3. Recommended action
4. Open questions

Good for recurring tasks with the same workflow.

My context:
- [Industry, team, product]
Output rules:
- When uncertain: say so, don't guess
- No recommendations without evidence
- Plain English, avoid jargon

For projects where quality standards matter more than a fixed workflow.

Claude offers three tiers that work independently or in combination:

TierScopePurpose
Profile PreferencesAll conversations, account-wideGlobal context about you
Project InstructionsSingle projectProject-specific rules and context
StylesPer conversation or globalTone and formatting (no content context)

Profile Preferences provide the “who I am.” Project Instructions provide the “what this project needs.” Styles provide the “how it should sound.” Combining all three creates the strongest context you can build on Claude.

PlanProjectsKnowledge BaseRAGSharing
FreeUp to 5Basic uploadsNoNo
ProUnlimitedFullYes (automatic)No
TeamUnlimitedFullYesYes (team)
EnterpriseUnlimitedFullYesYes (organization)

On paid plans, the knowledge base automatically scales beyond the Context Window The span of text an AI model can 'see' at once — your entire conversation history plus the current input. Everything within this window influences the response. — Claude activates RAG Retrieval Augmented Generation — a technique where the AI model retrieves relevant information from a knowledge base before responding. This extends capacity beyond the context window. automatically and can process up to 10× more content.

Create a Claude Project for a recurring task from your workday. Write custom instructions using the “Role + Rules” pattern. Upload a reference document. Test with 3 typical requests.

Use the project for a week. Note where Claude deviates from the desired behavior. Adjust instructions with targeted changes — one rule per problem. Observe whether results improve.

Exercise 3: Knowledge Base vs. Instructions

Section titled “Exercise 3: Knowledge Base vs. Instructions”

Experiment: What belongs in instructions (rules, behavior) vs. the knowledge base (facts, references)? Move content between the two. Observe the difference.

Claude Projects are one of several ways to build persistent context. In the next lesson, you’ll learn about Custom GPTs — OpenAI’s approach to the same problem, with its own strengths and trade-offs. The comparison helps you understand which tool fits which situation.

Part of AI Learning — free courses from prompt to production. Jan on LinkedIn