Skip to content
tl;dr on Claude Projects
- A Claude Project has its own chat history, instructions, and a file panel labeled Context..
- Projects are available on all plans. Free accounts have a five-project cap.
- Files are capped at 30MB each, with no limit on how many.
- I filled a fresh project with text files until it stopped accepting them. It reached 845,271 words, registering as 98% filled, and refused any further files.
- Project capacity and the context window are separate. The project holds several times more.
- Chats within a project are not visible to each other. Only Instructions and Context carry across.
- Best use: a fixed reference library you can query repeatedly. Anything with a date on it will go stale.
Claude Projects are workspaces that hold sets of files and instructions, so every chat you start inside one begins with the same context. A normal chat starts empty apart from retained memory. A Project chat starts pre-loaded.
So how much fits, and what happens when it fills? I added files to a blank project until it wouldn’t take any more. The total was 845,271 words, and it refused the next file.
During my research, I wasn’t able to find a specific number that projects are capped at, or what the Project limit is on a free account. That’s less an omission than a sign of how fast this is moving. Everything below carries a date, and everything got checked against a live account rather than lifted from a page.
What a Claude Project Is, And What It Isn’t
Three terms in Claude get mixed up regularly.
Projects vs Chats and Memory
A regular chat starts empty. A project chat starts with everything you put in it. Claude’s memory does build a picture of you over time and brings that forward on its own, but a Project is fed specific information for recall.
Take a company’s brand guidelines. Upload that into a chat and Claude retains elements of it, but not for recall. Upload it into a Project and it becomes a ground-truth corpus, influencing every chat in that project.
Project vs Skills
Anthropic puts it this way: a Project is what Claude needs to know, a Skill is how it should be done. Skills run in any chat, in a project or outside one.
A Note On Naming
Claude is developed constantly, and names change with it. Anthropic’s documentation says “project knowledge”, but the panel is called Context in the product. I’ll use Context throughout.
Plans, Limits, And Who Actually Gets What
Projects are available on every plan, including free, and the caps differ by tier.
| Plan | Price | Projects | Sharing | Retrieval | Context window |
|---|---|---|---|---|---|
| Free | $0 | Up to 5 | No | Listed as available Retrieval page says all plans. Projects page says paid only. | 200K |
| Pro | $17/mo annual, $20 monthly | Unlimited | No | Yes | 200K |
| Max | From $100/mo | Unlimited | No | Yes | 200K |
| Team | $20/seat annual, $25 monthlyPremium seats $100 and $125 | Unlimited | Yes, within the org | Yes | 200K |
| Enterprise | Custom | Unlimited | Yes, within the org | Yes | 500K |
Prices were correct at the time of writing, July 31, 2026. Seat pricing and the Team and Enterprise breakdown are covered here.
Plenty of articles say Projects need a paid plan. On a live free account, you get five.
Retrieval is worth knowing about. When a project gets too big to load whole, Claude searches your files rather than reading all of them. Anthropic’s two help pages describe who gets it differently, and testing turned up something odder than either version, which I’ll come to when I fill a project up.
Setting Up A Claude Project
Head to claude.ai/projects, or click Projects in the sidebar, then New project.
There are two fields, one for what you are working on and one for what you want to achieve. Both are for your own reference and are not used by Claude.
On Team or Enterprise you can set visibility: private, or open to the whole workspace.
Next, add your files. The panel sits on the right, named Context on paid plans or Files on Free. Anything added here is used by every chat in the Project.
Then set your project instructions. They apply to every chat you start.
Team and Enterprise get a Share option: add people by name, email, or bulk invite list.
The Context Panel: What Goes In, And Where It Stops
On a paid account, the Context panel takes up to 30MB per file, with no cap on the count.
Ten document formats: PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, and XLSX, though XLSX needs code execution and file creation switched on. Text extraction only, so a scanned PDF with no text layer uploads fine and reads as nothing.
I built three markdown files of around 94,000 words each, about 4,000 lines apiece, and planted a canary sentence in one: a name and a case reference that appear nowhere else. Then I uploaded one file at a time, opening a fresh chat after each and asking for the canary in a single prompt, deleting that chat before the next test. Paid account, Opus 5, 31 July 2026. Reaching the ceiling meant uploading the three files repeatedly, so the last run held three copies of the canary.
Two Different Measurements
One file registered at around 11% of project capacity. That same file runs at about 125,000 tokens of a 200,000 context window, the measure that dictates how much Claude reads in one go. A single file is almost all of it.
The panel tracks project capacity, a far larger pool. The context window is a different measurement, and it is never shown to you.
The files were large, so search mode kicked in immediately.
Upload anything substantial and you get retrieval from the first file. Write prompts as though Claude is searching, and get specific. I found my canary every time because the prompts asked for concrete details.
Hitting The Wall
Each file was about 11% of project capacity. After nine uploads, 845,271 words, I reached 98% capacity. I tried one more and it was refused.
Doing the math, the real ceiling sits somewhere around 862,000 words, though that’s extrapolated from a percentage meter, so treat it as approximate.
With an almost full project, could Claude still find what was needed? Yes. Nothing was dropped, nothing was truncated, and the answers held up. The upload gets refused and your existing base sits exactly as it was.
Retrieval Held
In every test I ran, at 11%, 22%, and up to 98%, I surfaced my canary line every time, in a fresh chat with the previous ones deleted.
One account, one day, one model, one canary, and three near-duplicate files. Directional, not a benchmark.
What Should Go In Context
Anything that stays true. Style guides, completed research, contracts, finished reports. Anything evergreen.
Anything with a date on it will go stale in place, and Claude has no way of telling which of your files is the current one.
Claude Project Instructions That Hold Up
Claude Project instructions load into every chat you start, which makes length a standing tax. Write the format spec, cut the explanation.
One of the best bits of guidance on this isn’t for Claude either. Taiki Minami runs customer success at Speee, a Japanese company where the sales team records every call with tl;dv. He built more than twenty prompts for pulling handoff notes out of transcripts, testing each one repeatedly before releasing it to the team. He ended up with four rules that work for Ask tl;dv, and they apply just as well to Claude Projects.
Lock The Output Format
The best instructions come with specificity. Uniform output makes everything downstream possible, and vague instructions lead to vague results that may still look fine at a glance.
Give Proper Nouns
Again with specificity: name your accounts, products, and people in the instructions. I did this with my canary test and got good results. Had I only asked “What’s the case number?” it may not have worked out which one I meant.
Three-level Breakdown Per Item
Do not write big long paragraphs. Instructions need to be short, dense with information, and shaped the same way each time. Prose may feel more natural, but it can end up with lopsided results.
Ask For Source Reference
In Minami’s case, this is timestamps. For a contract, it could be specific clauses. Specificity with the source matters as much as the output.
Here’s a block you can paste and edit.
OUTPUT FORMAT
Respond in this structure every time. No preamble, no summary paragraph.
## [Account name]
**Status:** [one line, max 15 words]
**Last contact:** [date, meeting type]
### What they want
- Stated goal: [one line]
- Underlying reason: [one line]
- What success looks like to them: [one line]
### Open concerns
- Concern: [one line]
- Raised by: [name, role]
- Source: [file name, timestamp or page]
### Commitments made
- What: [one line]
- By whom: [name]
- By when: [date, or "not stated"]
- Source: [file name, timestamp or page]
Use plain text. No emojis. No bold except the labels above.
If a field has no answer in the files, write "not stated."
Never infer a date, a number, or a name that isn't written down.
KNOWN NAMES
Accounts: [list your accounts here, exactly as spelled]
Products: [list your product names here]
People: [list names and roles here]
Transcripts mangle proper nouns. If something in the files
resembles a name on these lists, use the spelling from the list.
If it resembles nothing on these lists, quote it exactly as it
appears and flag it.
SOURCING
Every factual claim needs a source: file name plus timestamp,
page, or section. A claim with no source is not usable.
If you cannot source it, leave the field as "not stated"
rather than answering from general knowledge.Claude Project Examples By Use Case
Using Claude Projects can be really helpful for everybody, whatever their function. The core things are fixed reference material in Context, everything clearly laid out in Instructions; depending on your role, what you include shifts.
Sales. Closed-won transcripts, competitor pages, current pricing. Instruct it to answer only from the files and cite the account and date. Good for asking which objections keep recurring.
Research. Source PDFs, your notes, the methodology doc. Instruct it to quote page numbers and flag anything it can’t source. Keeps a long project consistent across weeks of separate chats.
Support. Product docs, past escalations, the escalation policy. Instruct it to give the documented answer first, then any workaround. Stops answers drifting between agents.
Engineering. Architecture docs, style guide, key modules. The GitHub connector needs manual syncing, so a Project holds a snapshot of the repo rather than its current state.
Where A Claude Project Stops Working
Feed call transcripts into a Project manually and they have a shelf life. More calls happen, and they aren’t in there. You end up querying a photograph of your pipeline.
Connecting Google Drive directly is one way around it. Claude reads the current version rather than what you uploaded months ago, though the limits are narrow: Google Docs only, added by URL or from recent documents. GitHub connects too, but syncing is a button somebody has to remember to press. And Scheduled runs recurring tasks inside the project, which is Anthropic’s own move at the freshness problem.
That last one is the idea Speee is testing, feeding accumulated tl;dv meeting data into Claude on a recurring basis to generate coaching feedback. Recurring is the operative word.
All three come with the same catch. They are private-project only. Share the project with your team and the connector option disappears.
Which is the problem MCP was built to solve. Custom connectors are the other route, worth the setup if your data changes daily.
When Claude Projects Doesn’t Work
Three problems come up more than any others, and all three have simple fixes.Claude is not using your files
Check the line count on the file card in your Context panel. If a large PDF is showing a handful of lines, it’s a scan with no text layer. It uploaded fine, and Claude can read almost none of it. Run it through OCR and upload it again. If the line count looks right, the question is the problem. Once a project is big enough to use search mode, Claude retrieves against your wording rather than reading everything. Name the file, name the client, name the document type.Instructions are being ignored
Are they actually in your project instructions, or did they go into the description field when you set the project up? Claude reads the instructions. It does not read the description.There is no Projects button
Go to claude.ai/projects directly. And if you’re on the free plan and can’t create a sixth, that’s the cap rather than a fault.Claude Projects in 2026: What To Keep, What To Connect
Building a Claude Project is the easy part. The key is knowing what to put into it and what to leave outside. For Context, use style guides and finalized reports. Anything that has stopped changing.The rest should stay where it’s generated and get queried live, because a copy of something that shifts is out of date by the time you need it. If that material is meetings, transcripts, and live call data, an MCP server keeps it current rather than leaving you a static export from last quarter. tl;dv records calls as they happen and its MCP sits on Pro and above, so it’s a paid tier. The free plan has no meeting cap, so you can build up a few weeks of calls first.FAQs About Claude Projects
Are Claude Projects free?
How much can a Claude Project hold?
I filled one to 845,271 words across nine files, registering as 98% of capacity. The tenth upload was refused rather than truncated. Project capacity is a separate measurement from the 200,000 token context window and holds several times more, so the panel percentage is not telling you what Claude can read at once.
Do chats in a Claude Project share context?
No. Each chat inside a project starts fresh and cannot see what happened in the others. Only your Context files and project instructions carry across. If you want something available in every conversation, it has to go into one of those two places rather than being mentioned in a chat.<
When does a Claude Project switch to search mode?
Does Anthropic train on files in a Claude Project?
By default, Anthropic uses conversations from Free, Pro, and Max accounts to train its models unless you switch that setting off. Team and Enterprise sit outside it. Project files are not addressed separately in the policy, so treat this as applying and check the current terms yourself.
