- 01
- September
"What are Claude Projects?" — the short answer is a Project is a separate workspace in Claude with its own document library and its own standing instructions. Every chat opened inside that room sees the same files and the same rules, with no re-uploading. This article covers how much a project can hold, how sharing and permissions work, and five project blueprints an organisation can stand up on day one.
In one line: A Claude Project = project knowledge + custom instructions + its own chat history. Projects exist on every plan (free accounts are capped at five), but sharing a project with colleagues requires the Team or Enterprise plan.
How a project differs from an ordinary chat
Most people start with Claude the same way: open a new chat, attach a file, ask a question, close it. Tomorrow the same files get attached again, and the same organisational context gets re-typed — "we are a contractor, this is our chart of accounts, budget is split across three funding sources" — every single time.
A project fixes that by moving context out of the conversation and into the room. Files and instructions belong to the room, not to the chat. Anyone who opens a new chat inside the project inherits the same context immediately.
| Aspect | Ordinary chat | Project |
|---|---|---|
| Reference documents | Re-attached in every chat | Uploaded once; every chat in the room sees them |
| Standing instructions | Retyped each time, or forgotten | Set once in project instructions, applied everywhere |
| Chat history | Every topic mixed in one list | Scoped to the room, easy to find later |
| Sharing | One conversation at a time | The whole room including its knowledge (Team / Enterprise) |
| Answer consistency | Depends on how much context each person typed | Everyone gets the identical context |
That last row is the one that matters most in an organisation. The recurring complaint is rarely "the AI was wrong" — it is "two colleagues asked the same question and got different answers", because one attached the current regulation and the other attached last year's copy.
The three parts of a project
| Part | What it is | Real-world example |
|---|---|---|
| Project knowledge | The file library Claude draws on in every chat in the room | Procurement regulations, chart of accounts, SOP manuals, contract templates |
| Custom instructions | Standing rules that set the role, tone and output format | "Always cite the clause number. If the answer is not in the uploaded documents, say so." |
| Chats | Conversation history that lives in this room only | "Review the September TOR" / "Summarise the audit findings" |
The part people underrate: custom instructions carry more weight than most teams expect. A single line — "if the answer is not in the project knowledge, say you could not find it; never guess" — cuts vague, unsupported answers more effectively than uploading ten more files. See also how to verify AI answers before trusting them.
How much can project knowledge hold, and when does RAG kick in?
The first question every IT team asks is "how many files?". Per Anthropic's own documentation (checked 1 September 2026), the number of files is not the constraint — per-file size and total context capacity are.
| Item | Limit | Notes |
|---|---|---|
| File size in project knowledge | 30 MB per file | Text extraction only, except multimodal PDFs |
| Number of files in a project | No fixed cap | Bounded by total context capacity instead |
| Attachments in a single chat | 500 MB per file · up to 20 files per chat | A separate allowance from project knowledge |
| PDFs | Up to 1,000 pages | Charts and images are read up to 100 pages; beyond that, text only |
| RAG mode | Expands capacity by up to 10x | Enabled automatically as knowledge approaches the context limit; paid plans only (Pro, Max, Team, Enterprise) |
RAG (retrieval-augmented generation) means that instead of stuffing the entire library into context, the system retrieves only the passages relevant to the question. The library can therefore grow without answer quality collapsing. Nothing needs configuring — Claude switches modes on its own. The trade-off worth knowing: in RAG mode answers are grounded in the retrieved passages, not the whole corpus, so very broad prompts such as "summarise everything in here" perform worse than specific ones.
The companion number is model context size. The Team plan uses a 200K token context, while Claude Opus 5 and Sonnet 5 support a 1M token context on all paid plans when chatting directly with Claude. A full plan-by-plan breakdown is in every Claude plan limit compared.
Sharing: view-only versus edit
Projects can be created on any plan, including free accounts (capped at five projects). But sharing a project with colleagues is a Team and Enterprise feature — which is the main reason organisations move off individual subscriptions to Claude Team.
| Permission | What it allows | Who should have it |
|---|---|---|
| Can view | Chat in the room and use its knowledge, but cannot change files or instructions | Everyday users in the department — the safest default |
| Can edit | Add or remove knowledge files and change the custom instructions | One or two content owners per room, accountable for keeping documents current |
Projects can be shared with named individuals, with a bulk list of email addresses, or with the whole organisation; recipients find them under the "Shared with me" tab. Because permissions ride on user accounts, this ties directly into user management in Claude Team and Enterprise.
Before you share org-wide: project knowledge has no per-file access control. Anyone who can open the room can read every file in it. If a document contains personal data, salaries or cost prices, put it in a separate project from the start — do not rely on instructions to hide it. See configuring Claude Team for Thailand's PDPA and data governance for Claude in the enterprise.
Five projects an organisation can stand up on day one
If you have just switched on Claude Team and cannot decide where to begin, these five rooms pay back fastest, because the documents already exist — nothing new has to be written.
| Project | What goes in the knowledge | Instruction worth setting |
|---|---|---|
| Procurement rules | Procurement regulations, practice manuals, previously approved TOR examples | Always cite the clause; if not found, say not found |
| HR policy | Leave rules, benefits, employee handbook (with individual data removed) | Answer in plain language and point to the page number |
| Internal ERP manual | User guides, chart of accounts, period-close steps, recurring help-desk questions | Answer as numbered steps using the exact menu names in the system |
| Quotes and contracts | Quotation templates, standard terms, legal-approved clauses | Never invent new terms; select only from approved clause text |
| Management reporting | Monthly report formats, KPI definitions, approved sample reports | Use the KPI definitions in the knowledge; do not invent formulas |
Start small, see value fast: the first room to build is whichever topic generates the most repeat questions — usually procurement rules or the internal system manual. Both get asked at the help desk every week, and the answers already live in existing documents.
Three mistakes that make projects fail
1. One room for everything. A project called "Company documents" holding procurement rules, HR policy and marketing decks answers worse than three small rooms, because a single question pulls in unrelated material.
2. Nobody owns document currency. When a regulation changes, someone has to remove the old file. Keeping both the old and new versions in the same room is more dangerous than having neither, because Claude will answer from both without knowing which one is in force — the same failure mode organisations already know from passing Excel files around.
3. Treating a project like a database. Projects hold documents, not balances. Upload today's stock report and tomorrow the number in the library is already wrong. Anything that needs current figures has to come from the system of record.
Projects and ERP: where the line sits
The workable division is simple: the ERP holds the numbers that are true; the project holds the rules and the know-how. Stock balances, committed budget, approved purchase orders — those belong in an ERP system with an audit trail and named approvers. Questions like "which clause allows a direct-award purchase?" or "what is our period-close sequence?" are exactly what a project is for.
To be direct about our own product: Saeree ERP does not yet connect Claude to read live data inside the system, and the in-product AI assistant is still in development. What does work today, and works well, is using a project as the knowledge base for how to use the system — put the user manual, the chart of accounts and the questions the help desk answers repeatedly into a single room, and let users ask there before raising a ticket. That reduces help-desk queue time without changing a line of the system.
Organisations that want to go further — letting Claude read live data — should look at connectors and MCP, not at uploading reports into project knowledge.
Where projects fit, and where they do not
| Good fit | Poor fit |
|---|---|
| Repeat questions about regulations, manuals and policies that sit still | Numbers that change daily, such as stock or remaining budget |
| Drafting documents from templates the organisation already approved | Documents that require digital signatures and an audit trail |
| Getting a team to answer customers with one consistent set of wording | Personal data or individual compensation in a widely shared room |
| Onboarding new staff so they can self-serve the manual first | Work where a wrong answer causes immediate harm with no human review |
Conclusion
Claude Projects are the cheapest way to get a whole team answering from the same source, because context moves out of individual heads and into a room everyone shares. The numbers to remember are few: 30 MB per knowledge file, no cap on file count, RAG automatically extending capacity up to tenfold on paid plans, and sharing requiring Team or above.
What decides success is not how many files you upload. It is whether someone is accountable for the library being current. One room with a clear owner beats ten rooms with none.
A knowledge base nobody keeps current is more dangerous than no knowledge base at all — a wrong answer sounds more authoritative than no answer.
- The Saeree ERP team
References
- Anthropic — What are projects? (checked 1 Sep 2026)
- Anthropic — Retrieval augmented generation (RAG) for projects
- Anthropic — Upload files to Claude (file limits)
- Anthropic — Context window on paid Claude plans
Rolling out Claude Team in your organisation?
We help design your first set of projects, configure sharing permissions, and handle Thai invoicing and documentation end to end.
Get advice / request a quoteTel 02-347-7730 | sale@grandlinux.com


