# What are Claude Projects, and how do they keep context?

Date: 2024-07-04 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/claude-projects-they-keep-context/

**Answer:** Claude Projects are workspaces in Claude.ai where you group related chats with a shared set of documents, called project knowledge, and a set of custom instructions. Every new chat in the project starts with that knowledge and those instructions already in view, so you stop re-explaining your situation and re-uploading the same files each time.

Anyone who uses a chatbot for ongoing work knows the routine: open a new chat, paste the style guide again, explain the project again, upload the same report again. Anthropic calls this the "cold start" problem, and on June 25 it introduced Projects in Claude.ai for Pro and Team subscribers to solve it.

A project is a folder with memory attached. Inside it you keep the chats about one piece of work, and alongside them two things that Claude reads at the start of every chat in that project: the **project knowledge**, documents and text you upload, and the **custom instructions**, a standing brief on how Claude should respond.

## How Projects keep context

A language model has no lasting memory between conversations. What it knows in a given chat is what fits in its **context window**, the amount of text it can consider at once, measured in tokens (small chunks of words). Each project offers a 200,000-token context window, which Anthropic compares to a 500-page book.

Projects use that window deliberately. When you open a new chat inside a project, your knowledge files and instructions are placed into the context before your first message, so Claude answers as if it had read your material and your brief. The chat still works like any other conversation from there.

What carries over from chat to chat is what you put in the knowledge and instructions. A conclusion reached in last Tuesday's chat is not automatically known in today's. If a decision matters for future work, add it to the project knowledge, for example as a short "decisions so far" note, and every later chat will have it.

Projects can use Claude 3.5 Sonnet, Anthropic's newest model. Anthropic also states that data and chats shared within Projects are not used to train its models without a user's explicit consent.

## A worked setup: a volunteer newsletter

Imagine you edit the monthly newsletter of a neighborhood association. Here is how a project for it could look.

**Project name:** Association newsletter

**Custom instructions:**

```prompt
You help edit the monthly newsletter of the Riverside Neighbours Association.
Readers: about 400 households, many retired, some new families.
Tone: warm, plain, practical. Short paragraphs. No jargon.
Always use the spelling and naming rules in "Style sheet."
Never invent dates, names, prices or places: if a detail is missing
from my notes or the knowledge files, ask me for it.
Each article ends with one line on how to get involved or learn more.
```

**Project knowledge:**

| File | Why it is there |
|---|---|
| Style sheet (1 page) | Street names, how to write dates and times, words to avoid |
| Three past newsletters | Shows the voice and length that readers know |
| Association bylaws summary | So meeting and voting details are described correctly |
| Contacts and venues list | Correct names, rooms and addresses |
| Decisions so far (running note) | Choices made in earlier chats, such as the new column name |

**A chat inside the project:**

```prompt
Draft the lead article for October from these notes:
[paste meeting notes on the fall cleanup day]
About 250 words.
```

Because the style sheet, past issues and venues list are already in context, the draft should use the right street names, the familiar voice and the correct meeting room, without you supplying any of it in this message. Results vary, so check every name and date against your notes before printing.

> **Try it:** Pick one recurring task, such as a weekly report or a course you teach. Create a project, write five lines of custom instructions, and add two or three reference files. Then ask the same question in a regular chat and in the project, and compare what changed.

## Sharing with a team

On the Team plan, colleagues can work in a shared project, and members can post snapshots of useful conversations to the project's activity feed. That turns a good chat into an example others can learn from, such as the prompt that produced a clean meeting summary or the conversation that fixed a tricky spreadsheet formula. Anthropic has said it plans to let projects draw knowledge from popular apps and tools through integrations in the coming months.

## What should go in a project's knowledge?

Put in the material you would hand to a capable new colleague on their first day: the reference documents they would need to consult again and again. Anthropic's own examples are style guides, codebases, interview transcripts and past work. Good candidates share three traits: they are stable, they are relevant to most tasks in the project, and they are accurate.

Favor short, focused documents over large archives. A one-page style sheet will shape every answer; a 300-page folder of old minutes mostly adds noise and uses up context that could hold something more useful. If you have a long source, consider writing a summary of the parts that matter and uploading that instead.

Keep the knowledge current. Replace outdated versions rather than adding new ones beside them, so Claude is never weighing two contradictory price lists. A running "decisions so far" note is often the most valuable file of all, because it is how one chat's conclusions reach the next.

Leave out what does not belong. Sensitive personal data, passwords and confidential client records should stay out unless your plan and your organization's rules clearly allow it. And remember that instructions and knowledge do different jobs: rules about tone and format go in the custom instructions, facts and reference material go in the knowledge.

## Sources

- Collaborate with Claude on Projects: https://www.anthropic.com/news/projects
- Claude 3.5 Sonnet: https://www.anthropic.com/news/claude-3-5-sonnet