What Claude Projects Does and How to Use It

Claude Projects is not a search engine. It does not index your documents and retrieve the relevant one when you ask a question.

It is a persistent context environment: a space where you upload documents and instructions that Claude has access to throughout every conversation in that Project, without you re-entering them each time.

The difference matters. A knowledge base retrieves. A Projects context informs. The company that understands this distinction builds a more useful system than the one that expects Projects to function like a searchable document library.

This article explains exactly what Claude Projects does: the mechanics of context persistence, document upload, custom instructions, and team sharing.

Not a theoretical company-wide knowledge base. A practical, maintained context environment that produces operationally specific outputs for every team member who uses it.

Pre-publication note: this article describes Claude Projects as of mid-2026. Verify the current feature set, context limits, and collaboration capabilities at docs.claude.com before publication.

The mechanics — what actually happens when you upload a document

What context persistence means

When you upload a document to a Claude Project, the document’s content becomes part of the context that Claude has access to in every conversation within that Project.

(Verify supported file types at docs.claude.com.)

The practical meaning: when you open a new conversation in the Project and type “draft a customer notification for a delayed shipment,” Claude is already working with the uploaded communication standards document. It does not need to be re-pasted, re-described, or re-referenced. It is present.

What “present” means — and what it does not

“Present” does not mean “retrieved from a search index.”

Claude does not search the document for the most relevant passage the way a knowledge management system does. The document is part of the context it reasons from.

The practical guidance: Projects works best with focused, well-organised context documents. Not with bulk document dumps.

What custom instructions do

Custom instructions are persistent instructions that apply to every conversation in the Project, without appearing in the conversation itself.

They shape how Claude behaves in that Project: what format it uses, what workflow it follows when specific inputs are provided, what quality standards it applies.

The operational use:

The Billing Project’s custom instructions specify that when claim denial data is pasted, Claude should categorise by denial code, prioritise by claim value and deadline, and draft appeal letter stubs for the top claims.

Every billing coordinator who opens the Project and pastes their denial data gets the denial triage output. They do not write the instruction themselves. The instruction runs silently every session.

What conversation history does

Within a Claude Project, past conversations are accessible to team members with Project access. This enables:

• The AI system owner to review team sessions, identify which produced the best outputs, and understand what inputs and approaches work best
• Team members working on multi-week projects to reference prior sessions within the Project context

What it does not do: the conversation history from Session A does not automatically inform Session B. Each new conversation starts fresh with the Project’s documents and custom instructions, not with the accumulated context from every prior conversation.

To reference a prior session, the team member explicitly includes relevant information from it.

The company knowledge base use case — what works and what does not

What works: the focused operational context library

The Projects configuration that produces the strongest operational results is a focused library of purpose-built context documents: each document written specifically for its operational function, at 200 to 500 words, updated regularly.

Documents that work well in Projects:

These documents are short enough to be fully utilised in every conversation, specific enough to produce operationally distinctive outputs, and focused enough to be maintained accurately as the company evolves.

If you want a framework for building these documents before loading them into Projects, see what an AI context pack is and how to give AI full business context.

What does not work: the bulk document dump

The Projects configuration that produces frustration: uploading the full employee handbook, all past client proposals, the company’s contract library, all historical emails, and the previous year’s grant submissions.

The expectation is that Claude will function as a comprehensive searchable company memory. It will not.

Three reasons this configuration underperforms:

1. Context window dilution

The more undifferentiated content is present, the less attention any specific document receives in any given conversation.

2. Relevance confusion

When a customer notification session has access to the employee handbook and the contract library simultaneously, the outputs may be influenced by irrelevant content.

3. Maintenance impossibility

A Projects context containing 200 uploaded documents cannot be maintained accurately. The AI system owner cannot track what is current, what is outdated, and what is contradicting what.

The knowledge base use case that works in Projects is not “all company knowledge available to Claude.” It is “the specific operational knowledge Claude needs to do this function’s work well, organised and maintained specifically for that purpose.”

How to build the Projects context as a genuine operational system

The document architecture for a $15M company with three function Projects

Shared Company Project (all-team access)

Uploaded documents:

• Company overview (200 words): what the company does, who its customers are, what makes it distinctive
• Brand voice guide (250 words): how the company sounds, vocabulary to use and avoid, the relationship tone
• Core terminology guide (150 words): the specific terms the company uses for its products, services, and processes

Total context: 600 words. Well within the context window. Present and fully utilised in every conversation.

Operations and Customer Service Project

Uploaded documents:

• Customer communication standards (250 words): tier-based communication conventions, common situation framing, specific vocabulary
• Back-order and exception vocabulary (200 words): the specific language for delay situations, supplier issues, recovery communications
• Operations briefing format (150 words): the structure and sections of the weekly management briefing

Custom instructions: specify the back-order notification workflow, the exception communication workflow, and the weekly briefing workflow.

Total context: 600 words plus custom instructions. Fully utilised in every conversation.

Billing Project

Uploaded documents:

• Payer communication vocabulary (300 words): appeal structure for common denial codes, regulatory citation format, escalation language
• Compliance reporting standards (200 words): report format, required elements, narrative conventions

Custom instructions: specify the denial triage workflow, the appeal letter workflow, and the compliance report workflow.

The maintenance cadence

Monthly review

The AI system owner reads through each Project’s context documents for accuracy. Updates any document where:

• The company’s practice has changed
• The vocabulary has evolved
• Quality feedback has identified a gap

Quarterly refresh

Review the custom instructions for each function Project against the quality data from the conversation history. Update any instruction that is consistently producing outputs requiring significant editing.

When a significant event occurs

A rebrand, major client relationship change, regulatory update, or product/service change triggers an immediate update of the relevant context documents. Do not wait for the monthly review cycle.

Common questions on Claude Projects

”How many documents can you upload to a Claude Project?”

Verify the current upload limits at docs.claude.com. Projects supports a 200K context window (approximately 500 pages) and scales further through RAG mode.

The operational guidance: keep each Project’s context focused rather than maxing out the upload capacity.

”Can team members in a Project see each other’s conversations?”

Shared Project conversations and the Project’s knowledge base are accessible to team members with Project access. Individual conversations are private by default unless shared. Verify the current visibility and sharing mechanics at docs.claude.com, as these features have evolved.

”What happens when the context window fills during a long conversation?”

Claude Projects uses RAG mode when the Project knowledge approaches context limits, which expands capacity while maintaining response quality.

For individual conversations that become very long, the earlier parts of the conversation receive proportionally less attention. For long, multi-part tasks: break the work into separate sessions within the Project rather than continuing a single conversation indefinitely.

”How does Projects compare to ChatGPT’s Custom GPTs?”

Both serve similar operational purposes: persistent context for team deployments.

The key differences: Claude Projects makes all uploaded documents present in every conversation simultaneously. ChatGPT’s Custom GPTs retrieve relevant content from uploaded files.

For focused operational context documents (200 to 500 words each), both architectures work. For multi-document operational deployments where simultaneous context from multiple documents matters: Claude Projects tends to produce more consistent multi-document utilisation on long-form operational writing tasks.

Want the Projects architecture designed and the context documents built — so your shared Claude workspace is running and maintained from week two?

Claude Projects functions as a genuine operational context system: not a searchable knowledge base, but a persistent context environment that makes Claude work from the company’s specific operational knowledge in every session.

The configuration that works: focused, well-organised context documents at 200 to 500 words each, custom instructions that specify the workflows for each function, and an AI system owner who maintains both as the company evolves.

Path one: build your first focused context document today. Take your customer communication standards. Write a 250-word version that covers tone, vocabulary, and structure for the most common communication types. Upload it to a new Project. Run three notification drafts against it. Compare those outputs to what Claude produces from a blank session. The difference is your business case for the full deployment.

Path two: bring in a partner. Phos AI Labs builds the Projects architecture and the context documents for operations teams across every sector. Thirty minutes, no deck. Start here.