TL;DR
A Claude Project bundles three things that persist across every chat inside it: uploaded Knowledge files, Custom Instructions (a per-project system prompt), and a list of chats. It does not share chat history. As of June 2026 Projects are available on every plan including Free, you can upload unlimited files at up to 30MB each, and Claude automatically switches to retrieval (RAG) once your Knowledge approaches the ~200K-token context window, expanding effective capacity by up to 10x. The trick is to treat Knowledge as a curated citation list, write tight Instructions, and write decisions back into a file — chats forget, files do not.
What a Project actually shares
Most people use a Project as a glorified file folder, then wonder why answers drift when they open a new chat. Here is the real model.
A Project carries exactly three things into every new chat opened inside it:
- Knowledge — your uploaded files (PDF, DOCX, CSV, TXT, HTML, code, and more). Available to every chat in the Project without re-pasting.
- Custom Instructions — a per-project system prompt: role, audience, voice, and constraints. Applied to every chat automatically.
- The chat list — a folder of conversations grouped under the Project, so you can find them later.
What it does not share is conversation memory. Two chats in the same Project cannot see each other’s messages. If a decision needs to carry over, you write it into a Knowledge file and re-upload. That single fact explains most “why did Claude forget” complaints.
Who this is for
Anyone who keeps re-pasting the same context into Claude. If you run a book draft, a long research thread, a recurring client engagement, or a side codebase where you keep saying “here is the spec again,” Projects pay off in the first week. One-off questioners can skip the setup overhead.
Three concrete cases where Projects beat plain chats:
- A novel with a 12-character cast. Load
characters.md,worldbuilding.md, and the latest outline once; every chapter draft inherits the same voice rules. - A consulting engagement. Brand guidelines, last month’s deliverables, and a running
decisions.mdgive every email and deck the same tone. - A side codebase. README, architecture notes, and
coding-conventions.mdkeep code-review chats consistent across weeks.
For a book or article series specifically, pair Projects with a real Claude writing workflow so each chat starts from a fixed voice and outline rather than a blank slate.
Knowledge vs Instructions: what goes where
This is the decision that makes or breaks a Project. The split is not cosmetic.
| Knowledge (files) | Custom Instructions (per-project prompt) | |
|---|---|---|
| Holds | Source material: specs, transcripts, brand guides, code | How to behave: role, voice, constraints, success markers |
| Size | Unlimited files, up to 30MB each (June 2026) | A text box; keep it under ~1,000 words |
| Changes | Static snapshots — edit locally, then re-upload | Edit in place; applies to the next message instantly |
| Best for | ”What is true” — facts Claude should ground answers in | ”How to act” — the standing rules for this Project |
| Failure mode | Stale files silently desync from your local copies | Vague prose (“be helpful”) that adds nothing |
Rule of thumb: if it is a fact Claude should cite, it is Knowledge. If it is a behavior Claude should follow, it is an Instruction.
How Knowledge scales (and where it gets sloppy)
You can upload unlimited files, but capacity is not free. As of June 2026, Anthropic’s Claude models (Opus 4.7 and Sonnet 4.6) run a 200K-token context window on consumer plans. Two regimes follow:
- Below the limit: every file loads into context on every prompt. Precise, but token-heavy, and answers slow as the pile grows.
- At or above the limit: Claude switches to retrieval-augmented generation (RAG). Instead of loading everything, it searches your files and pulls only the relevant chunks. A visual indicator appears showing the Project is RAG-enabled. This expands effective capacity by up to 10x. (Anthropic’s RAG-for-projects docs.)
RAG is available on every plan — Free, Pro, Max, Team, and Enterprise — and it kicks in automatically; you do not toggle it. The practical takeaway: keep Knowledge lean anyway. Once retrieval is doing the work, irrelevant files become noise that competes with the chunks you actually want. A focused 5-file Project retrieves better than a 50-file dumping ground.
Step by step
- Create a Project named for the outcome, not the topic. “Q3 launch plan” forces you to define done; “marketing” invites scope creep.
- Add 3 to 7 Knowledge files. Treat each like a citation: if you would not link to it in the final deliverable, it does not belong. Supported formats include PDF, DOCX, CSV, TXT, HTML, and code, up to 30MB per file.
- Write Custom Instructions. Cover role (“you are my book editor”), audience, voice (“plain English, no marketing copy”), a hard constraint (“never invent statistics; ask for sources”), and one or two success markers. Keep it under ~1,000 words.
- Open a chat and lead with the sub-goal. First message: “Today: outline chapter 4 using the existing voice from
chapter-3.md.” When your recurring output is a single iterable document or small app, pair this with a deep Claude Artifacts workflow. - Write decisions back to a file. When a chat reaches a thesis, a chosen direction, or a kill-list, summarize it in a paragraph, append to
decisions.md, and re-upload. This is how state survives across chats that cannot see each other. - Prune every two weeks. A healthy long-running Project ends up with roughly 30% less Knowledge than it started with. Outdated drafts and dead ends quietly degrade retrieval.
Sharing a Project with your team
Project sharing is a Team or Enterprise feature; personal Free, Pro, and Max accounts keep Projects private to you. (Claude Pro is $20/month, $17 on annual billing, as of June 2026; Max is $100 or $200/month.) On Team and Enterprise, when you create a Project you choose its visibility:
- Public — everyone in your organization can view and use the Project. Individual chats inside it stay private unless you share a specific chat.
- Private — only invited members can view or use it. You assign each member can view or can edit; editors can change Instructions and Knowledge.
In both cases, sharing a Project shares its Knowledge and Instructions, not your chats. Your conversations remain private unless you explicitly share an individual chat. (Anthropic’s project sharing docs.)
Common mistakes
- Dumping 30 files because storage is cheap. Whether you are under the context limit or in RAG mode, more files mean noisier retrieval, not smarter answers.
- Naming the Project after a topic. “Client X” invites scope creep; “Client X Q3 deliverables” names a finish line.
- Skipping Instructions entirely. Without them, every chat starts from Anthropic’s defaults instead of your context.
- Treating Knowledge files as live documents. They are snapshots. Edit a file locally without re-uploading and your chats silently keep reading the old version.
- Mixing personal experiments with billable client work in one Project. Voice drifts, and so does your trust in the output.
- Never pruning. Retrieval quality decays as dead files accumulate.
FAQ
- Do Projects share chat history?: No. Projects share Knowledge files and Instructions, not conversations. Two chats in the same Project cannot read each other. To carry a decision over, write it into a file and re-upload.
- What is the practical Knowledge limit?: There is no hard file cap — unlimited files, up to 30MB each. But Claude’s context window is ~200K tokens (June 2026). Below that, all files load every prompt; at that point RAG turns on automatically and retrieves only relevant chunks, raising effective capacity up to 10x. Either way, a lean, focused file set retrieves better than a large one.
- Can I share a Project with my team?: Only on Team and Enterprise plans, where you set a Project public or private and grant members view or edit access. Free, Pro, and Max keep Projects private to your account.
- Is RAG something I have to turn on?: No. It activates automatically when your Knowledge approaches the context limit, and a visual indicator shows the Project is RAG-enabled. It also reverts to full-context mode if your Knowledge shrinks back below the threshold.
- One mega-Project or many?: Many, scoped by outcome. One Project per deliverable is far easier to prune than one Project holding “all my work,” and it keeps retrieval focused.