Claude Projects and Skills for a small team: how to set them up

This guide explains how to set up Claude Projects and Skills for a small team. Projects hold the shared context for a piece of work. Skills hold the steps for a task you repeat. If you get both right before you connect Claude to your business systems, those later connections have something reliable to work from.

Three terms matter. A Project is a self-contained workspace with its own chat history and knowledge base. You upload documents to it and set project instructions. A Skill is a folder of instructions, scripts and resources that Claude loads when a task matches it. Custom instructions are your general preferences, and they apply to everything Claude does for you.

The difference is scope. Context you upload to a Project exists only in that Project. A Skill, once created, is available in any conversation, including regular chats and chats inside Projects. That split gives you a simple rule. Put what the work is about in a Project, and put how the work is done in a Skill.

The Claude Help Center puts it directly. Projects provide static background knowledge that is always loaded when you start a chat inside them. Skills activate when needed and work everywhere across Claude. Skills use what the help centre calls progressive disclosure: Claude decides which Skills are relevant and loads only the information the task needs, which helps prevent context window overload.

The table shows why the two work best together. Skills also combine automatically when a task needs more than one area of expertise, so you do not have to name each one. Custom instructions sit above both. Keep them for preferences that genuinely apply to everything, such as British spelling.

It is tempting to create one large Project called "Company" and upload every document the business owns. Claude then reads the price list while drafting a job advert, and nobody is sure which file is current. A Project per process keeps the background tight and makes ownership obvious.

A process here means a piece of work with a clear start, finish and owner. Examples are answering customer enquiries, preparing quotes, onboarding a new starter and month-end reporting. Each gets its own Project, its own instructions and its own knowledge files. If two processes need the same document, such as your terms of business, upload it to both and make one person responsible for updating both copies.

Keep the number small at first. Free accounts can create a maximum of five Projects, and on paid plans more Projects simply means more files to keep current. Two or three Projects covering your busiest processes will teach you more than a dozen half-built ones.

One change is on the way. The help centre article on Projects describes a new version in beta, rolling out in stages starting with Claude Code, in which a Project becomes one conversation that Claude splits into parallel threads. Existing Projects keep working as they do today, so nothing here needs to wait for it.

Treat a Skill as the procedure you would hand a capable new colleague. It says when to use it, what to check, what order to work in and what the finished output looks like. Simple Skills are written in Markdown and need no coding. For more advanced work you can attach executable scripts.

Every Skill has a name and a description, and these decide when it loads. Claude matches the task in front of it against them. So the description should read like a trigger: "Use when replying to a customer enquiry about prices or delivery times." A vague description produces a Skill that never fires.

Put the detail in the body. Skills can contain far more detail than custom instructions, and in Claude Code a Skill's body loads only when it is used, so long reference material costs almost nothing until you need it. A useful body covers:

• the trigger, repeated in one line at the top;
• the steps, in order;
• the rules that must not be broken, such as never promising a delivery date the terms do not support;
• the output format, with one good example;
• what to do when information is missing.

You do not have to start from a blank page. Describe the task to Claude and it will build and structure it into a properly formatted skill file, which you then edit. In Claude Code the file is called SKILL.md. If you upload a Skill to claude.ai, keep the frontmatter to the fields the Agent Skills specification allows; the Claude Code documentation warns that any other field makes the upload fail with a hard error.

Take a small team that answers customer enquiries by email. Here is the setup, using the settings named in Anthropic's own pages.

1. Turn on Skills. Go to Settings > Capabilities > Skills and check that code execution and file creation are on, since Skills run in Claude's sandboxed computing environment. Toggle on some of the pre-built example Skills to see how they behave.
2. Create the Project. Name it "Customer enquiries". Upload the current price list, delivery terms and your tone-of-voice guide to its knowledge base. Add project instructions: who the customers are, what you sell, and that replies go out under the sender's name.
3. Share it. On a Team or Enterprise plan, share the Project with everyone who answers enquiries. Give Can edit to the process owner and Can view to the rest, so they can chat in the Project but not change its instructions or knowledge.
4. Write the Skill. Create a Skill called "enquiry-reply" with the description "Use when drafting a reply to a customer enquiry about price, stock or delivery." The body lists the steps: identify the product, check the price list, state the delivery terms, flag anything the files do not cover, draft in the house tone.
5. Provision it. As organisation Owner, provision the Skill for all users and set it as enabled by default, so it appears in every team member's skills list.

Now paste a past enquiry into a chat inside the Project, with names, email addresses, phone numbers and order references removed first. The Project supplies the prices and terms; the Skill supplies the method. Because the Skill works everywhere, the same method applies if someone drafts a reply outside the Project, although without the Project's files it should flag more gaps.

A Project is only as good as its knowledge files. An out-of-date price list is worse than none, because Claude will use it with confidence. So every Project needs a named owner, and every file needs a date.

Use the permission levels to enforce this. Members with Can edit can modify project instructions and knowledge, add and remove members and update member settings. Members with Can view can see the contents, knowledge and instructions and chat in the Project, but cannot edit it. In a small team, one person with Can edit per Project, plus a deputy for holidays, keeps changes traceable.

Organisation Owners also control sharing as a whole. They can turn project sharing off for the organisation, and existing shares stay in place when they do. Make that choice deliberately rather than leaving it to chance.

A simple maintenance routine

• Put a version date in every file name.
• Replace old files rather than adding new ones beside them.
• Update the Project the day an underlying fact changes, such as a price rise.
• Give each Skill an owner too, usually the same person as the Project.

Size is rarely the constraint. On paid plans, when project knowledge approaches context limits, Claude switches on a mode the help centre calls RAG and expands capacity by up to 10x. The constraint that matters is whether the files are right.

Test with real work, not invented prompts. Collect past requests where you know what a good answer looked like, including awkward ones: a discontinued product, a request outside your terms, a customer asking two things at once. Remove names, email addresses, phone numbers and order references before you paste them in, because the test works just as well without them.

First, check that the Skill loads. In Claude's thinking, look for "Reading [your skill's name]". If it is not there, the task did not match the Skill's name or description, so adjust one of them, or mention the Skill by name, and try again. In Claude Code you can also call a Skill directly with /skill-name.

Then check the output against the procedure. Confirm that it used the current price list, that it flagged gaps instead of guessing, and that it followed the format. Fix each failure at its cause: a missing step goes into the Skill, a missing fact goes into the Project.

At Svennis we keep every past request that once failed as a standing test set and rerun it whenever a Skill or knowledge file changes, because a fix to one step often breaks another.

Finally, check where your Skills come from. Skills can include executable scripts, so install only those written by you or a source you trust. On the Enterprise plan, organisations can turn on skill scanning to check uploaded skills and plugins for malicious content.

Use this table to assign the work. Each step has one owner and a clear test of when it is finished.

Note that sharing Projects and provisioning Skills for everyone both require a Team or Enterprise plan. On individual plans each person keeps their own Projects and Skills. That is fine for a trial, but it does not give you one shared version of a process.

Projects and Skills are the foundation for the next stage: connecting Claude to the systems where your data lives. MCP (Model Context Protocol) is the protocol that connects Claude to external services and data sources. The help centre describes the division of labour clearly. MCP connections give Claude access to tools, while Skills teach Claude how to use those tools effectively.

That is the reason to write and test Skills first. A connection to Zoho CRM without a written procedure lets Claude read and change records with no agreed method. With a tested Skill in place, the connection follows the steps your team has already checked. For a fuller example of a procedure working on live data, see building a quote generator with Claude that uses your CRM prices.

Two points from Anthropic's Claude for Small Business announcement help here. Existing tool permissions carry over: if an employee cannot see something in QuickBooks or Drive today, they cannot see it through Claude. The package also includes 15 skills built on the repeatable tasks owners said slow them down most, which makes a useful model for writing your own.

Claude for Small Business is switched on inside Claude Cowork. If that is where your team works, read how to use Claude Cowork for your business alongside this guide.

UK figures show the gap this setup addresses. A government announcement from January 2026 puts AI adoption at 24.3% among micro businesses with fewer than 10 employees, against 44.4% for large businesses with 250 or more. Only 21% of UK workers feel confident using AI at work, and 46% of small businesses told the FSB they do not yet have the skills or knowledge to use it well.

Projects and Skills lower that barrier because they capture knowledge once. A new starter does not need to be good at prompting. They need access to the right Project and a Skill that already contains the method. Training still helps: every adult in the UK is eligible for free, benchmarked AI courses, some taking under 20 minutes. Claude Academy also lists an Introduction to agent skills course of 6 lessons over about an hour, and an AI Fluency course for small businesses.

Data protection is the other concern. In a survey Anthropic ran with small business owners, half named data security as their single biggest hesitation about AI. Anthropic states that it does not train on your data by default on its Team and Enterprise plans, which is one reason to keep shared business files on those plans rather than on personal accounts. Before you upload customer data to any Project, work through our GDPR checklist for Claude.

You can set this up in stages without outside help. Work through the list below in order.

1. Pick the one process that takes most of your team's time and name its owner.
2. Create its Project, upload dated knowledge files and write the project instructions.
3. Ask Claude to draft a Skill for the task it repeats most, then edit the name and description until they state the trigger plainly.
4. Test against real past requests. Fix the Skill when the method fails and the Project when a fact is missing.
5. Share the Project with Can view for the team, provision the Skill and set a review date.
6. Only then plan connections to live systems.

Once one process runs reliably, the next step is to let Claude work with your live records. Our guide to connecting Claude to Zoho CRM with MCP covers that setup, and Claude for small business inside your tools explains where it fits in day-to-day work.