Stop re-explaining yourself to Claude: build a Project that remembers your work
A Claude Project holds your context once so you stop re-pasting it into every chat. What goes in knowledge, what in custom instructions, and how to scope it.
A Claude Project is a workspace where you set your context once so you stop re-explaining it in every new chat. Stable reference material (a brand guide, a spec, a glossary, past decisions) goes into project knowledge; how Claude should behave (role, tone, output format) goes into custom instructions. After that, every new chat inside the project already knows what you're working on.
The familiar pain: you open a fresh chat and spend the first ten minutes re-explaining who you are, what the product is, the tone you want, what's already decided. A project removes those ten minutes. Here's how to build one that actually remembers, instead of blending contexts.
What is a Claude Project, exactly?
A project groups your chats around a shared context. It has two settings you configure once:
- Project knowledge - the files and documents Claude sees in every chat in this project. Everything reference-like goes here: the product description, a style guide, key decisions.
- Custom instructions - how Claude should behave here specifically: the role it plays, the tone it answers in, what it should never do.
Both are picked up automatically. Open a new chat inside the project and the context is already loaded. Projects are available on Claude's paid plans.
What goes in project knowledge vs custom instructions?
The rookie mistake is dumping everything into one pile. Facts live in knowledge, behavior rules live in instructions.
| Project knowledge | Custom instructions | |
|---|---|---|
| What it holds | reference material, facts | behavior rules |
| Examples | brand guide, spec, glossary, past decisions | role, tone, output format, constraints |
| When you change it | when the facts change | almost never |
| Answers the question | what Claude should know | how Claude should answer |
Simple test: if it's a fact (a product detail, a number, a decision), it goes in knowledge. If it's about how Claude behaves (keep it short, answer in a table, don't invent things), it goes in instructions.
How do I set up a Project that remembers my work?
Five steps, in order:
1. Create a project per recurring context. One client, one product, one book. Not a single "Work" project that everything falls into. 2. Write the custom instructions. Who Claude is in this project, the tone it answers in, the format, what it never does. Short and concrete, not a wall of text. 3. Load the project knowledge. The stable material Claude refers to: the product description, a style guide, a glossary, key past decisions. 4. Start a new chat per task, but inside the project. Context carries over on its own, while chats stay separate and easy to find later. 5. Keep the knowledge fresh. When a fact changes, update the file. A stale document gets answered confidently and wrong.
How should I scope my Projects?
One project, one recurring context. A client, a product, a book, a role.
- Too broad: an "Everything work" project mixing three clients and personal notes. Claude starts blending contexts and pulling the wrong one into an answer.
- Too narrow: a separate project for every small task. You lose the whole point, shared context never accumulates.
- The rule: if you keep pasting the same background into chat after chat, that background has earned its own project.
What makes a Project drift or forget?
- A dumping ground for knowledge. Twenty files uploaded "just in case" and Claude cites the wrong one. Keep only what's actually needed.
- Knowledge that never updates. An old spec gives old answers. Every couple of weeks, check what's still current.
- Facts in instructions, behavior in knowledge. Swap them and both work worse. Keep them separate.
- One giant project for everything. Contexts leak into each other. Several narrow projects beat one pile.
In short
A project is Claude's memory of your work. Put stable facts in knowledge, behavior in custom instructions, and keep one project per recurring context. Then a new chat starts with the work, not with "let me explain who I am first."
You learn to work with Claude hands-on at AGINE Academy - on your own tasks, not from notes. The starting block is open without signing up.
Questions
A normal chat only remembers itself: close it and you re-explain everything in the next one. A project holds context above the chats, so its knowledge and instructions load into every new chat inside it. You set it up once and stop restating who you are and what you're working on.
Stable reference material Claude should refer to constantly: a product description, a style guide, a glossary, key past decisions. Don't dump everything in, the less noise the more accurately Claude cites the right thing. And keep it current, or you'll get confident answers based on outdated docs.
One per recurring context: a client, a product, a book, a role. Avoid a single "everything" project, or the contexts start blending. A simple tell: if you keep pasting the same background into chat after chat, it's time to give it its own project.