TL;DR: Claude Projects prompting splits into more pieces than most guides admit: project knowledge (files and text you upload), project instructions (a standing behavior block), and, since Anthropic shipped a broader memory feature, project memory: an auto-generated summary you never touch directly. This page is about the first two and how they interact with the third, sourced from Anthropic's live docs on 3 September 2026.
If you have ever wondered whether the PDF you uploaded to a Claude project three weeks ago still gets read on every new chat, or whether your project instructions and your uploaded files are competing for the same slot, you have found the real ambiguity in Claude Projects prompting. Anthropic's own documentation gets some of it right and contradicts itself on the rest. This page separates project knowledge from project instructions from the plain prompt, then covers the newer complication: a memory feature that now runs quietly underneath every project whether you set it up or not.
What is a Claude project, and who can actually use one?
A project is a self-contained workspace: its own chat history, its own knowledge base, and (on Team or Enterprise) its own sharing permissions. Anthropic's help center describes it directly: "Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases." Within one you upload documents, add context, and keep chats scoped to that one topic.
Access is broader than a lot of "Claude Projects" content assumes. Anthropic states plainly, in more than one article, that "Projects are available to all users, including those with free Claude accounts. Free users can create a maximum of five projects." That is not a beta or a paid-only feature. It ships on the free tier.
There is a catch worth flagging rather than smoothing over. The same personalization article that confirms free access later summarizes the three personalization tools as: profile instructions for account-wide settings, skills for tone and format, and "Use project instructions when you need specific guidance or context for a particular project (paid plans only)." Free accounts get projects, per the paragraph above it, but the article's own closing summary marks project instructions paid-only. Both lines are live on the same page as of this writing. Anthropic has not resolved it, so treat "paid plans only" in that one line as an inconsistency to verify against your own account rather than a rule to plan around.
On Team and Enterprise, projects add real collaboration: an owner can share a project to specific members or the whole org, with "Can view" (see contents, chat, cannot edit) or "Can edit" (modify instructions and knowledge, add members) permission levels. None of that changes what goes into project knowledge; it changes who can add to it.
What's the difference between project knowledge and project instructions?
These are two separate fields in the same project, and they do different jobs.
Project knowledge is the files tab. Anthropic's help center: "You can upload relevant documents, text, code, or other files to a project's knowledge base, which Claude will use to better understand the context and background for your individual chats within that project." This is reference material (documents, transcripts, code, prior outputs) that more than one chat in the project needs to see.
Project instructions are a separate text block, set once per project, that shapes behavior rather than supplying facts. Anthropic's own example: "instructing Claude to use a more formal tone or answer questions from the perspective of a specific role or industry." Setting it up is a two-step process documented at Claude Project Instructions: 18 Copy-Paste Templates in more depth than this page covers: click "Set project instructions." Write the block, save it, and it applies to every chat in that project from then on.
The practical split: if the same rule should shape every response in the project regardless of what's being asked, write it into instructions once. If it's a fact, document or piece of reference content a chat might need to look up, upload it as knowledge instead. Repeating a standing rule inside every prompt, or uploading a one-line behavioral note as a "file," both work against the design.
Does Claude load your entire project knowledge base, or search it?
It depends on size, and Anthropic's own pages disagree about who gets the larger mode.
Below the context window threshold, Anthropic's comparison-with-Skills article describes projects in flat terms: they "provide static background knowledge that's always loaded when you start chats within them." That's accurate for a knowledge base that fits comfortably in context.
Once your uploads approach that limit, the mechanism changes. The dedicated RAG article explains it plainly: "Claude seamlessly enables RAG mode to expand capacity by up to 10x while maintaining response quality." Instead of loading every file into the request, Claude uses what Anthropic calls a project knowledge search tool to retrieve only the passages relevant to your current question. So "always loaded" and "searched" are not actually in conflict: they describe the same feature at two different sizes, automatically, with no setting to force either mode.
The context window your project knowledge has to fit inside depends on the model doing the chatting, per Anthropic's own context-window article:
| Model family | Context window on paid plans |
|---|---|
| Claude Fable 5.1, Opus 5, Sonnet 5 | 1M tokens |
| Claude Opus 4.8, 4.7, 4.6; Sonnet 4.6 | 500K tokens |
| Every other current Claude model | 200K tokens |
A knowledge base that comfortably fits a Sonnet 5 chat's 1M-token window can quietly tip into RAG mode the moment you switch models mid-project to something with a 200K window. That's a real, if underdocumented, reason a project can feel different after a model switch that had nothing to do with your prompt.
How do you prompt so retrieval actually finds the right file?
Once RAG mode is active, the phrasing of your prompt affects what gets retrieved, and Anthropic publishes concrete advice on this rather than leaving it to guesswork. Four points from its own best-practices section, verbatim in substance:
- Upload comprehensively upfront. "Add all relevant documents and files to your project upfront. The more context Claude has access to, the better it can assist you."
- Name files descriptively. "Well-named files help Claude understand and retrieve the right information more effectively."
- Group related material. "Group related documents in the same project to enable Claude to draw connections between different sources."
- Reference documents by name in the prompt itself. "When asking questions, you can reference specific documents by name to help Claude focus its search."
That last point is the one people skip. A retrieval-backed project responds differently to a vague question than to one that names its source:
Weak: "What's our refund policy?"
Better: "Check refund-policy-2026-08.md and summarize the exceptions
section for a Team-plan customer who bought via AppSumo."
The second version gives the search tool an actual filename to match against, on top of the topic. If your project is genuinely knowledge-heavy (dozens of files, several document types), naming the source is the single highest-leverage habit Anthropic's own docs recommend and the one most people never adopt.
What belongs in project knowledge, project instructions, or just the prompt?
With both fields covered, here's the decision in one place:
| Goes in... | Use it for | Scope |
|---|---|---|
| The prompt (this message) | The specific task, this turn's constraints, anything transient | One message |
| Project instructions | Standing tone, role, format, or process rules | Every chat in the project |
| Project knowledge | Reference documents, code, prior decisions multiple chats need | Every chat in the project, loaded or searched |
| Project memory | Context Claude infers from your conversations, not something you write | Every chat in the project, generated automatically |
If you find yourself pasting the same paragraph of context into every new chat in a project, that paragraph almost certainly belongs in project knowledge or project instructions instead. And if you're maintaining the same instructions block by hand across five different projects, note that Anthropic gives you no built-in way to template or reuse it: no gallery, no version history. A prompt library that any client can reach, covered in Best Prompt Managers for Claude Users, is the usual workaround for that specific gap.
That fourth row, project memory, is the part most existing guidance on this topic has never had to account for. It shipped recently enough that it changes the answer to whether context carries across chats automatically, so it gets its own section.
What is project memory, and how does it differ from project knowledge?
Project memory is not something you upload. It's a running summary Claude builds from your conversations inside a project, as part of a broader memory feature Anthropic has rolled out to claude.ai chat. Per Anthropic's own memory documentation: "Memory is on by default for Free, Pro, and Max plans on the web, Claude Desktop, and Claude Mobile (update to the latest version of the app). On Team and Enterprise plans, memory is off by default and can be turned on by an owner."
Crucially, it is scoped per project, separately from your knowledge base: "Each project has its own separate memory space and dedicated project summary, so the context within each of your projects is focused, relevant, and separate from other projects or non-project chats." Mention a deadline moved in one chat, and a later chat in the same project already knows, without you uploading anything or writing it into instructions.
This creates a real, currently unresolved tension with older documentation. Anthropic's guide to creating and managing projects still states, in a note that predates the memory rollout: "Context is not shared across chats within a project unless the information is added into the project knowledge base." Taken literally, that note is now wrong for any account with memory turned on: memory carries context across chats in a project automatically, without anything being added to the knowledge base. Both pages are live as of 3 September 2026. Treat the older note as stale rather than authoritative.
A few mechanics worth knowing before you rely on it:
- Turning memory off gives two distinct choices: "Pause memory" keeps what's already saved but stops new saves, while resetting it is destructive. Anthropic's settings page states plainly that reset "Permanently deletes all memories including project memories." That action cannot be undone.
- Chat search, a related capability, is boundary-aware: you can ask Claude to search individual project conversations, and Anthropic notes that those "searches are limited to within each specific project"; it will not surface a conversation from a different project by accident.
- Anthropic's personalization overview article, the one page that lists "Instructions for Claude", project instructions and skills side by side, does not mention memory as an option anywhere on the page, as of 3 September 2026, a real gap in a document whose entire purpose is listing the ways to personalize Claude.
Is project memory the same as Claude Code's Auto Memory?
No, and conflating the two is the easiest mistake to make in 2026, because Anthropic named both features "memory" within weeks of each other on different products.
The memory covered above lives on claude.ai: the chat product, Projects included, plus Cowork when it runs in the cloud. Claude Code's memory system is a different product entirely, documented on a different domain, for coding sessions in a terminal or IDE rather than chats on claude.ai. Its own documentation is titled plainly "How Claude remembers your project" and describes two components: CLAUDE.md files, which it defines as "instructions you write to give Claude persistent context", and auto memory, defined as "notes Claude writes itself based on your corrections and preferences". Read that title again: Claude Code's own memory page uses the word "project" to mean a code repository, not a claude.ai Project. Same word, two unrelated systems, on two different domains, both current.
Auto memory writes to a scope Anthropic's own comparison table calls "Per repository, shared across worktrees". It loads fresh into the first 200 lines or 25KB of every session. That is nothing like project memory's dedicated per-project space described above. If you're managing standing context for an AI coding agent rather than a chat project, that surface, and how CLAUDE.md fits alongside it, is covered separately in CLAUDE.md Best Practices. If your actual goal is wiring tools into that coding surface rather than a chat project, that's a different configuration step again, covered in Using MCP Inside Claude Code.
Can you use project knowledge through the Claude API?
No, not as a way to prompt with it, at least. Anthropic's Messages API reference, the endpoint every third-party integration actually calls to send a prompt, does not mention projects at all: the word does not appear once across the full reference, fetched 3 September 2026. Project knowledge, project instructions and project memory are all claude.ai product surfaces, not parameters you can set on a messages.create call.
The one place "project" does appear in Anthropic's API docs is the Compliance API, and it is not for prompting. It is restricted to Claude Enterprise organizations, for eDiscovery and account-deletion workflows: Anthropic's own description says "Projects bundle related chats together with custom instructions, knowledge base content, and attached files or text documents. The Compliance API exposes project metadata, project details, and the list of attachments belonging to a project." That's retrieval and deletion for legal and compliance review, gated behind a Compliance Access Key an Enterprise org has to provision. It is not something an individual developer building on the API can reach, and not a way to prompt against a project's knowledge programmatically.
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 4.8★ on the Chrome Web Store.
Create An AccountDoes Prompt Architects work inside Claude projects?
Yes, in the narrow sense that matters. Claude is a listed integration on our own /integrations page (checked 3 September 2026), and the enhance button appears in the compose box wherever it renders on claude.ai, a project chat included, since it's the same input field as any other Claude conversation.
What it cannot do is touch project knowledge or project memory. The extension enhances the prompt you're about to send; it has no visibility into your uploaded files, your project instructions, or Anthropic's auto-generated project summary, and no way to read or write either. That's an architectural boundary, not a missing feature: a browser extension sees what's in the compose box, nothing Anthropic stores server-side. Our own free plan is 5 prompt enhancements per day, forever, per the FAQ page, checked the same day.
The short version
Claude Projects prompting is three fields, not one: the prompt for this message, project instructions for standing behavior, and project knowledge for reference material multiple chats need, loaded directly below the context threshold and retrieved via RAG above it. Naming a specific document in your prompt is the one concrete lever Anthropic documents for improving retrieval. Layered underneath all of that, automatically, is project memory: a per-project summary Claude builds from your conversations, unrelated to Claude Code's Auto Memory despite the shared name, and unreachable from the Messages API either way. Anthropic's own docs disagree with themselves on two separate points here: RAG's plan availability, and whether context carries across chats without a knowledge-base upload. As of 3 September 2026, neither has been resolved. Check your own account before planning around either page alone.