Claude Projects Tutorial: How to Use Custom Instructions and Knowledge Files

Trackline Academy · Updated July 30, 2026 · All articles

A Project in Claude is a workspace that groups related chats together and gives all of them two shared resources: custom instructions (standing rules for how Claude should behave: its role, tone, and non-negotiables) and a knowledge base (reference files Claude can draw on in every conversation). You set both up once, and every chat you start inside the Project arrives already knowing your context.

That one sentence solves the most common frustration with AI chat tools: every new conversation starts from zero. You explain your business, paste your style guide, describe your audience. Then close the tab, and tomorrow's chat knows none of it. Projects end the re-explaining loop. This tutorial walks through setting one up, writing custom instructions that actually work (with a copy-paste template), structuring a knowledge base, and knowing when a Project is worth it versus just opening a chat.

The problem Projects solve

If you use Claude for anything recurring (a newsletter, a client, a course you teach, a job search), you have almost certainly caught yourself re-pasting the same background for the third time. Each individual chat is a fresh start, which is fine for one-off questions and terrible for ongoing work.

A Project fixes this by moving your context up one level. Instead of living inside a single conversation, your standing rules and reference material live at the project level, where every chat inside the workspace can use them. Think of it as setting up Claude's desk for one area of your work, once, instead of clearing and restacking it every morning.

The two halves of a Project

1. Custom instructions: the standing rules

Custom instructions are directions that silently apply to every chat in the Project. You never repeat them; they are just always in effect. Good instructions usually cover five things:

• Role: who Claude is acting as. "You are my grant-writing partner for a small nonprofit."
• Audience: who the output is for. This quietly steers vocabulary and depth.
• Tone and style: concrete, testable adjectives, not vibes. "Under 150 words, active voice, no exclamation marks" beats "professional but friendly."
• Standing rules (the non-negotiables): always cite which knowledge file a claim comes from, never invent statistics, ask before assuming.
• Format defaults: the shape you want 80% of the time, with an escape hatch: "…unless I ask otherwise."

2. The knowledge base: the reference shelf

The knowledge base is where you upload the material Claude should always have on hand: a style guide, a project brief, a product list, an FAQ, sample documents in your voice, background data. Claude uses these files as context in every chat inside the Project, so you stop re-attaching the same documents over and over.

Setting up your first Project, step by step

Exact button labels shift as the product evolves, so follow what's on screen, but the shape of the flow is durable:

1. Create the Project. Go to your Projects area and make a new one. Give it a name and short description. Small but useful fact: the name and description are just for your own organization: Claude doesn't read them as context. Anything Claude needs must go in the instructions or knowledge base.
2. Write the custom instructions. Open the instructions panel and start with a rough draft covering role, tone, and one or two standing rules. Don't aim for perfect. You'll refine it (more on that below).
3. Add one knowledge file. Upload a real document you own: a past newsletter, a style note, a bio, an FAQ. One good file is enough to feel the difference.
4. Start a chat inside the Project. Give it a real task that should use your file ("Draft this month's intro in our usual voice") and watch it arrive pre-loaded.
5. Start a second chat. Different task, same Project. Confirm the instructions and knowledge still apply without you re-pasting anything. That's the payoff moment.

A copy-paste custom-instructions template

The fastest way past the blank box is a labeled structure. Paste this into your Project's custom instructions and fill in the brackets:

Notice how every line is checkable. You can look at an output and immediately see whether the intro is under 120 words or whether a claim was sourced. Vague instructions ("be helpful and engaging") give you nothing to verify and nothing to fix.

Write instructions by subtraction, not perfection

The best custom instructions aren't drafted in one heroic sitting: they're grown. Ship a rough four-line version, run a couple of real tasks, and then add one precise rule per failure mode. Output too long? Add a word cap. Wrong voice? Add "always match the tone of the uploaded style guide." Ignored your file? Add the sourcing rule. Three quick iterations reliably beat one exhaustive attempt, because the rules end up targeting problems you actually had rather than problems you imagined.

Structuring the knowledge base

A knowledge base is a small, curated context source, and quality beats volume every time. Four habits that pay off:

• Curate, don't dump. A focused brief and one excellent example outperform twenty loosely related PDFs.
• Include a "gold" example. One artifact you'd call perfect teaches voice and format better than any pile of adjectives in the instructions.
• Add a do-not-do file. Explicit anti-patterns (phrases to avoid, claims never to make) are powerful and almost always forgotten.
• Name files for scanning. A convention like brief_, style_, example_, data_ makes the base self-documenting when you come back in three months.

And prune periodically. A lean, current knowledge base outperforms an ever-growing one: stale files can steer Claude toward last year's facts.

When a Project beats a long chat (and when it doesn't)

Not everything needs a Project. The test is simple: will you reuse the same context more than two or three times? If yes, it has earned a Project. If not, a plain chat is faster.

• Just chat: a one-off question, a quick fact-check, a single rewrite.
• Make a Project: you keep re-pasting the same background; it's a recurring area of work (a client, a newsletter, a course); or many chats should share one style and one set of rules.

A long-running single chat is the worst of both worlds: it accumulates context, but it becomes a junk drawer: hard to search, easy to derail, and everything in it is trapped in that one thread. Most people settle into a handful of Projects, one per ongoing area of their work or life: Monthly Newsletter, Client A, Job Search. Claude shows up with the right context, and you can actually find past work again.

Projects, memory, and privacy: the boundaries that matter

Separate from Projects, Claude also has memory features: at a high level, it can carry forward useful context across your conversations, learning your role and preferences over time. Two boundaries are worth internalizing, especially for sensitive work:

• Each Project is a contained workspace. Its knowledge files and chat history stay separate from your other Projects: a confidential client Project is a safer home than a pile of loose chats.
• Project chats get their own memory space. What Claude learns inside one Project doesn't leak into another or into your general profile.

You stay in control: settings typically let you view, edit, pause, or reset what Claude remembers, and there's usually a way to start a private session that isn't saved to memory. These features evolve, so check the current interface before relying on them for anything sensitive.

One useful rule of thumb from the boundary between the two: use the knowledge base for facts and rules that must be right (explicit and controlled by you), and lean on memory for ambient personalization (implicit and learned), verifying it when stakes are high.

FAQ

Do chats inside a Project share what I said in other chats?

No. Chats in a Project share the custom instructions and knowledge base, but what you type in one chat is not automatically visible to the next. If a fact needs to persist across conversations, put it in the instructions or upload it as a knowledge file.

What's the difference between custom instructions and the knowledge base?

Custom instructions are for standing behavior: role, tone, and rules like "never invent statistics." The knowledge base is for reference material: documents, data, and examples Claude should draw on. A style rule goes in instructions; the style guide itself goes in the knowledge base.

Do I need a paid plan to use Projects?

Projects are available on free claude.ai accounts, though free plans cap how many you can create. Since limits change, check the current plan details, but you don't need an enterprise subscription to start using Projects today.

Related guides

• How to Use Claude AI: A Beginner's Guide
• How to Write Better Claude Prompts
• Claude Artifacts: Build by Describing
• What Course Is Right for Me?