How to Use Claude Projects for Work: Knowledge Base and Style Guide

Most people who use Claude for real work hit the same wall within a few weeks. Every new chat starts from zero, so you paste the same brief again, then the same tone notes, then the same "please don't write like a press release." Learning how to use Claude Projects for work solves this, but only if you set up two things properly. The first is a knowledge base that Claude treats as the source of truth. The second is a style instruction that tells it how to behave. This guide covers both and includes prompts you can copy.

We focus on the standard Projects in Claude chat, which most professionals use for writing, client work and documents. All facts below are current as of November 2026.

What a Claude Project actually does

A Project is a workspace. Every chat inside it can use three things:

• Project knowledge: files and text you upload once. Claude reads them as background in every chat in the project.
• Project instructions: rules for how Claude behaves and responds in that project.
• Project memory: each project keeps its own memory, separate from your other projects.

There's one thing a Project doesn't do. Chats inside a project can't see each other. If you made a decision in Monday's chat, Thursday's chat won't know about it unless you add it to the knowledge base. Also, according to Anthropic's help center, Claude can't see the project's name or description. Anything Claude needs to know has to go in the instructions or the files.

As of November 2026, Projects are available on every plan. The Free plan allows up to five projects, and paid plans (Pro starts at $20 a month billed monthly) list no cap, according to Claude's pricing page. On paid plans, Claude automatically turns on retrieval (RAG) mode when project knowledge gets close to the context window limit, so larger knowledge bases keep working. Sharing a project with colleagues is available on Team and Enterprise plans, with "Can view" or "Can edit" permissions.

A note on the name: since September 2026, Anthropic has also been testing a redesigned "Projects" in Claude Code. In that version, one conversation coordinates parallel cloud threads for long-running technical work (MarkTechPost). It's still in beta. This guide covers only the knowledge-and-instructions Projects in Claude chat.

Step 1: One project per recurring job

The most common setup mistake is creating one giant "Work" project. Knowledge from five clients and three content types gets mixed together, and the instructions turn into a compromise that fits none of them.

Build each project around a job you do at least weekly with the same context. That could be one client account, your newsletter, a recurring report or proposals for one service line. A quick test: could you hand a new colleague one folder and one page of rules and have them do this job? If so, it's a project.

Step 2: Build the knowledge base

Upload less than you think you need, and give each file one job. For most writing and client work, five files are enough:

File	What goes in	What stays out
context.md	Who you are, who reads your work, what you sell, current priorities	Background nobody uses anymore
facts.md	Prices, names, dates and numbers, with a "last updated" line at the top	Anything you haven't confirmed
style-guide.md	Voice rules, banned words, a few "write this, not that" pairs	Vague advice like "be clear"
samples.md	3–5 pieces you'd send as-is, each labeled with its format	Drafts you had to rewrite
formats.md	Templates for recurring outputs: weekly update, proposal outline, reply types	One-off formats

We follow a simple rule: facts go in knowledge, and behavior goes in instructions. A price belongs in facts.md, while "answer in under 150 words" belongs in the instructions. Give files plain names and mention them by name in your prompts, so Claude knows which source you mean.

Be honest about your samples. If your best examples are mediocre, Claude will copy that mediocrity faithfully.

Step 3: Write the style instruction

Project instructions cover behavior: role, sourcing rules, tone and output format. Keep them short, because they share the context window with your files and the conversation. Here's the template we start from. Replace the brackets with your details.

Role: You are the writing assistant for [your role] at [company]. You draft; I edit and send.

Sources:
- Treat the files in project knowledge as the source of truth. When you use a fact, name the file it came from.
- If the files don't cover something, say so and ask me. Don't fill gaps with guesses.
- If two files disagree, point out the conflict instead of choosing one.

Style:
- Follow style-guide.md. If my request conflicts with it, follow the request and flag the conflict.
- Plain US English. Short paragraphs. Put the main point first.
- No filler openers, no hype adjectives, no exclamation marks.
- Use the matching template from formats.md. If none applies, stay under 200 words.

Output:
- Give the draft first. Then add a short "Check before sending" list: every number, date, name or claim I should verify.

The "Check before sending" line matters more than it looks. It means every draft arrives with a list of things a person should verify, and those are the spots where AI errors usually hide. For more patterns, see our custom instructions examples for work. They're written for ChatGPT, but the same principles work in Claude.

Test the project before you trust it

Before the first real task, open a new chat in the project and run this:

Before we start: list the files in project knowledge you can use, with one line on what you'd use each for. Then tell me:
1. Where the files contradict each other.
2. What looks outdated or incomplete.
3. What you'd still need from me to write a [weekly client update] without guessing.

This usually turns up an outdated fact or two files that disagree. Fix them in the files, not in the chat. A fix made in the chat is lost when the chat ends.

A worked example: the weekly client update

Here's a typical task for a marketer or consultant: every Friday, you turn messy notes into a client status update. With a "Weekly update" template in formats.md and the style instruction above, the prompt can be short:

Write this week's status update for [client name] using the "Weekly update" template in formats.md.

My notes (raw):
- landing page v2 shipped Tuesday; conversion data after 14 days
- email sequence delayed: waiting on their legal review since Oct 28
- next week: second round of ad creative, call on Thursday

Keep it under 150 words. If anything in my notes is unclear, ask before drafting.

On the first run, the draft usually follows the template, comes in at the right length and sticks to your notes instead of inventing progress. Where it often falls short is tone around problems. Models tend to soften bad news ("a slight adjustment to the timeline"), but the client needs a plain statement of what's late, why, and who owns the next step.

Put the fix in the project instructions, not in the chat:

Delays and risks: state them in the first two lines. Say what is late, why, who owns the next step, and the new date. Don't soften them with words like "slight" or "minor".

This is how a project gets better over time. Whenever you make the same correction twice, move it into the instructions or a knowledge file, and it will apply to every future chat in that project. For editing what's left, our guide on making AI writing sound human covers the final pass.

Where Claude Projects break

• Stale facts lead to confident wrong answers. If last quarter's price is in facts.md, Claude will quote it. Put a date on the file and review it every month.
• Large knowledge bases lose precision. The Free plan has no RAG, so all knowledge has to fit in the context window. Paid plans switch to retrieval, but retrieval can miss a passage. Name the file and section you need in your prompt.
• Memory doesn't replace files. Project memory helps with continuity, but you can't review it the way you review a document. If a decision matters, write it into a file. To keep a client project separate from your other chats, turn off "Use account memory" on the project page.
• Numbers still need checking. Calculations and figures pulled from long documents deserve a second look. Our fact-checking checklist shows how.
• Some material shouldn't go into a project at all. Personal data, anything under NDA and trade secrets need a policy decision before anyone uploads them. Check your plan's data terms and your company's rules. Our article on what not to upload to AI explains where to draw the line.
• Shared projects drift. On Team plans, anyone with "Can edit" access can change the instructions, so agree on who owns them.

Monthly maintenance checklist

• facts.md has a "last updated" date within the past 30 days
• No two files give different answers to the same question
• Every correction you made twice this month is now in the instructions or a file
• Samples still match how you write today
• Instructions fit on one screen
• Decisions from old chats are summarized in a knowledge file
• The project still covers one job. If it has grown into two, split it

If your team works mostly in ChatGPT, the same structure applies there. See our guide to ChatGPT Projects for the differences.

Build it yourself or bring in help

If you're on your own or in a small team, set the project up yourself. You know the context better than anyone, and the setup fits into one working session plus a few minutes of upkeep each week. It gets harder when projects need to connect to your CRM, email or document workflows, or when several teams need the same setup with the same rules. At that point, our AI implementation and automation service can take it on. If you'd prefer to build an assistant for one recurring task yourself, step by step, the Build Your Own AI Assistant mini-course walks you through it.

Going further

Session 4 of our live practicum, AI System in Two Weeks, covers this topic in depth: persistent settings, projects and workspaces, and building your own assistant for a recurring task. You work on your own tasks during the session and finish with configured assistants, a map of your personal AI system and a 90-day plan. The full program lists what each session covers.