A Claude Project can hold far more material than Claude can consider in a single response. Once project knowledge approaches the active model's context limit, Claude can switch from loading that material directly to retrieving only the parts it considers relevant. That distinction should shape the files you add and the instructions you give each conversation.
This guide explains how to use Claude Projects around those limits. You will set up a Project and see what changes when the knowledge base or conversation becomes large.
The short version
A Project is a workspace with reusable instructions, a knowledge base, and related conversations. Anthropic says project instructions and knowledge apply across chats in that project, but it also states that context from one chat is not shared with another unless you add it to project knowledge. Claude's separate memory and chat-search features can bridge some of that gap when enabled.
Use a Project for work that needs the same source material or rules more than once. Use an ordinary chat when the task is self-contained. If you are deciding between the two major workspace products, the companion guide to ChatGPT Projects covers the OpenAI side without turning this article into a feature-by-feature comparison. If your decision extends beyond workspace features, the ChatGPT vs Claude comparison for AI agents looks at shipped model configurations.
What Claude receives in a Project chat
It helps to separate the available context into layers. They persist differently and should hold different kinds of information.
| Layer | What belongs there | What carries forward |
|---|---|---|
| Project instructions | Role, response rules, source hierarchy, output format | Anthropic says these instructions apply to every chat in the project. |
| Project knowledge | Canonical documents, reference material, approved examples | Files in the knowledge base are available across the project's chats. |
| Current chat | The immediate task, working draft, corrections, decisions made during that thread | The full transcript remains part of that conversation. It is not automatically added to project knowledge for the next chat. |
| Project memory and chat search | Selected remembered topics and earlier project conversations | Availability depends on settings and plan. Memory is project-scoped; chat search must be invoked and searches only within that project. |
Those boundaries come from Anthropic's project setup documentation. Its memory documentation describes memory as saved topics rather than a copy of every conversation, with a separate memory space for each project.
That leads to a simple rule: put durable truth in project knowledge or instructions. Do not leave a decision only in a long chat and assume every future conversation will receive it verbatim.
How to create a Claude Project
Open Projects in Claude's left navigation, select New Project, and give the workspace a name. Anthropic notes that Claude does not receive the project's name or description as context, so repeat any information that should affect its responses in the instructions or knowledge base. The same setup page says Projects are available to free accounts and that free users can create up to five projects.
Write instructions for behavior, not the current assignment
Select Set project instructions and add the rules that should survive every new chat. Keep the current deliverable in the chat itself. Anthropic's guidance recommends concise project instructions for general context, core guidelines, and Claude's role because instructions consume context alongside the conversation and tools (usage and length limits).
A useful starting template is:
Purpose: Help with [stable project purpose].
Audience: Write for [reader or stakeholder].
Source rules:
- Treat the files in Project knowledge as the source of truth.
- Name the source file for factual claims.
- If the sources conflict or do not answer the question, say so.
Output rules:
- Use [format].
- Keep [stable terminology or style rule].
- Ask before [high-impact or irreversible action].
Avoid adding this week's topic, deadline, or draft-specific exception here. Those details will become stale and affect unrelated chats.
Add canonical material to Project knowledge
Use the + control in the knowledge panel to upload the documents that Claude should be able to reference in every project chat. Anthropic says content in this area is processed as context across the project, while a file attached only to an individual chat belongs to that chat, according to the same create and manage Projects page.
Prefer one current policy over several near-duplicates. Give files descriptive names, and state which source wins when documents disagree. This is especially important after retrieval activates, because Claude may search by the words in your request and the available document metadata.
Start with a source check
Before asking for a deliverable, ask Claude to identify the files it can use and flag obvious conflicts or gaps. For the real task, name the relevant document when you know it. Anthropic specifically recommends referencing files by name to focus project-knowledge retrieval in its RAG for Projects documentation.
For example:
Use onboarding-policy-current.pdf as the authority for eligibility.
Use support-examples.csv only for tone examples.
Before drafting, list any question the policy does not answer.
This prompt does not guarantee a correct answer. It does make the source boundary inspectable.
How much knowledge can a Claude Project hold?
Anthropic does not publish one absolute token ceiling for every Project. Its RAG documentation gives a relative limit: when project knowledge approaches or exceeds the current context window, Claude can automatically enable retrieval and expand project capacity by up to ten times. Anthropic does not document the exact trigger threshold or a plan-by-plan absolute Project capacity on that page.
There is also a live documentation conflict. The dedicated RAG article says the feature is available on all Claude plans, while the general Projects overview says enhanced knowledge with RAG is paid-only. The project setup page also describes automatic RAG specifically for paid plans. Until Anthropic reconciles those pages, check your own project for the RAG indicator rather than treating either plan statement as settled.
Retrieval changes what Claude sees
Before retrieval is needed, Claude can use in-context processing. When RAG activates, it searches project knowledge and loads only the material it judges relevant to the question instead of placing the entire knowledge base into active context, according to Anthropic's explanation of RAG. Activation and any return to in-context processing happen automatically; Anthropic does not provide a manual mode switch.
The operational consequence is easy to miss. Stored in the Project does not always mean visible to the model during this answer. A broad request such as “review our policy” gives retrieval less direction than a request naming the policy, topic, and required evidence. For work that depends on exact wording, ask Claude to quote the relevant passage and identify the file before it interprets it.
For an exhaustive comparison across a large library, split the work into scoped queries and reconcile the results in a final pass. Retrieval is designed to select relevant material, so one broad prompt is a poor audit trail for proving that every document was considered.
File types and upload limits
The file limit for Project knowledge is 30 MB per file. Anthropic lists the number of Project files as unlimited, subject to the total knowledge capacity, and warns that additional token limits can apply after text extraction. Every file still has to fit, and the page does not promise that every byte enters every answer.
Claude's supported document formats include PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, and XLSX. XLSX uploads require code execution and file creation to be enabled. Supported image formats are JPEG, PNG, GIF, and WebP. The current list belongs to Anthropic's file-upload documentation; convert anything outside that list before relying on it as project knowledge.
For non-PDF documents, Anthropic says Claude extracts text and does not interpret embedded images. PDF handling depends on length: it analyzes text and visual elements in files of up to one hundred pages, then uses text-only processing for PDFs from page one hundred one through page one thousand. PDFs beyond that published page ceiling are rejected. The Project's per-file size limit can stop an upload sooner.
If a chart, scan, or layout carries the meaning, do not hide it inside a non-PDF office file. Convert it to a supported format that preserves what Claude needs to inspect, then verify the upload with a question about that specific element.
How model context limits affect Projects
Storage and active context are separate. The context window is the amount of information the selected model can work with in one chat. Project instructions, the current conversation, retrieved knowledge, and enabled tools all draw on that working space. Anthropic advises keeping instructions concise and notes that tools and connectors are token-intensive in the same length-limit guidance.
On paid Claude chat plans, the current published context limits vary by model:
| Models in Claude chat | Published context window |
|---|---|
| Claude Fable 5.1, Opus 5, and Sonnet 5 | One million tokens |
| Claude Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6 | Five hundred thousand tokens |
| Other models on paid plans | Two hundred thousand tokens |
Project storage can exceed those model context limits through RAG, so the table is not a storage guarantee. The active limit can change when you select a different model, and a portion of the context is reserved for Claude's response. Anthropic also says that paid users with code execution enabled get automatic context management: as a long conversation approaches its limit, Claude summarizes earlier messages while preserving the full transcript for later reference in the same paid-plan context documentation.
RAG and conversation summarization solve different problems. RAG selects from a large Project knowledge base. Automatic context management compresses an individual conversation as it grows. Neither turns the context window into unlimited working memory.
What carries between conversations?
Project instructions and knowledge are the dependable shared layer because you put them there explicitly. Anthropic's setup documentation still states that context is not shared across Project chats unless it is added to Project knowledge.
Claude now has additional continuity mechanisms:
- Chat search: Paid users can ask Claude to search earlier conversations. Inside a Project, the search is limited to conversations in that specific Project and appears as a tool call, according to Anthropic's chat-search documentation.
- Project memory: Claude keeps a separate memory space and project summary when memory is enabled. Anthropic describes memory as a set of saved topics, not a transcript mirror. It is on by default for Free, Pro, and Max accounts; Team and Enterprise owners control whether it is available in their organizations, according to the same memory documentation.
Do not use either feature as the only record of an approval, requirement, or final decision. Move durable information into a decision log in Project knowledge, then update or replace it when the decision changes. That gives future chats an explicit source rather than asking memory to reconstruct one.
Sharing a Project without sharing every chat
Project sharing is documented for Team and Enterprise organizations. Members with Can view permission can access the Project's contents, knowledge, and instructions and can start chats there. Members with Can edit can also change instructions and knowledge and manage access (project sharing documentation).
A shared Project does not make its members' conversations visible by default. Anthropic says chats remain private unless the user manually shares a snapshot. Messages added after that snapshot also stay private until the user updates it, according to the same sharing and visibility documentation.
This separation matters for governance. Reviewers may share the same policy library and instructions while keeping their working conversations private. Put material the whole group must see in Project knowledge, not in one member's chat.
Troubleshoot the failure mode, not the wording
Claude ignores a file you uploaded
First, confirm that the file is in Project knowledge rather than attached to another chat. If RAG is active, name the file and the section you need. Ask Claude to identify the source passage before answering. If the file format or size is the issue, compare it with Anthropic's supported formats and Project upload limit.
A new chat misses a decision from an older chat
Ask Claude to search for the earlier Project conversation if chat search is available, then move the final decision into Project knowledge. Check the Memory settings if you expected project memory to apply. A related chat sitting in the same Project is not the same as its full transcript being part of the next prompt.
Claude uses an outdated policy
Remove the obsolete file or label it clearly as superseded. Add a source-priority rule to the Project instructions and cite the current file in the task prompt. A knowledge base with conflicting versions creates an evidence problem before it creates a prompting problem.
The answer becomes generic after the Project grows
Look for the RAG indicator or project-knowledge search tool. Narrow the request to named documents and ask for supporting excerpts. If the task needs complete coverage, query bounded subsets and keep a coverage table outside the conversation so omissions are visible.
Teammates cannot see a Project chat
Sharing the Project gives access to the common workspace, not to each private chat. Share the conversation separately or move the agreed output into Project knowledge. The same visibility guide documents both permissions independently.
A setup checklist that respects the limits
- Put stable behavior in concise Project instructions.
- Put canonical facts and approved examples in Project knowledge.
- Keep task-specific instructions in the current chat.
- Name the source files that matter when asking a question.
- Check whether RAG is active before assuming the whole library is in context.
- Treat memory as selected continuity and keep a separate verbatim record when one matters.
- Move durable decisions out of private chats and into a maintained source.
- Verify file format, extracted content, and visual handling before relying on an upload.
Claude Projects are useful because they preserve a controlled working environment across related chats. They work best when you know which parts are persistent, which parts are retrieved, and which parts still belong to one conversation.