The practical setup guide — projects, file architecture, skills, agents, financial security, connectors, and team policy. Written by an owner, for owners.
Most business owners are prompting their way through AI. The ones pulling ahead are building systems around it.
The gap between a business that uses Claude well and one that doesn't isn't the quality of their prompts. It's whether they've built the scaffolding that makes Claude work consistently, at scale, across a team.
Without structure, every team member starts from scratch in every conversation — re-explaining your brand, re-describing your customers, re-stating your tone of voice. Claude is equally capable in conversation one and conversation one hundred. But without persistent context built into projects and knowledge files, every session costs you setup time that compounds across a team.
This guide covers what that scaffolding looks like in practice: how to organise projects, where files should live, how to give Claude standing knowledge about your business, how to handle sensitive financial data safely, and how to build the first automations that actually save meaningful time.
On a personal Claude.ai account, your conversations may be used for model training. The Claude Team plan changes this: your data is not used for training, nothing is retained after the session ends, and you get admin controls over all users. Everything in this guide assumes the Team plan. If any team member is still on a personal account, fix that first.
A project structure that reflects who should see what. Shared projects for the team. Role-specific projects for sensitive functions. Private projects for the owner.
Your brand voice, product knowledge, and processes written as structured documents. Every project draws from them. Update once — it improves everywhere simultaneously.
A simple traffic light system the whole team follows. Financial data and sensitive files handled with a per-session pattern that gives Claude access without storing it permanently.
A written AI usage policy. A quarterly review. Human approval before anything AI-generated goes to a customer. Simple, but most businesses skip it entirely.
Claude projects are persistent workspaces with their own instructions, knowledge files, and conversation history. Getting this structure right is the single highest-leverage thing you can do.
A project is a dedicated role. Its system prompt tells Claude what it is in this context. Its knowledge files give it the information it needs. Every conversation in the project inherits both — so your team is always working with a Claude that already knows your business.
Projects appear as a flat list. There is no way to group them into folders. With multiple functions, customers, and team members, this list grows quickly. A strict naming prefix convention is the fix — it creates visual grouping and makes the list scannable. Set the convention on day one and enforce it. Changing it later when you have 15 projects is painful.
Use a consistent prefix per tier. Example: CO — Brand & Voice · CO — Customer Service for shared projects. SALES — [Client name] for accounts. No prefix for private and restricted projects — they don't appear in the Team tab anyway. Agree the convention on day one. Everyone uses it.
Your files are Claude's long-term memory. Two storage locations with different rules. Getting this distinction right is the most important architectural decision in your Claude setup.
There are two places a file can live. One makes it permanently available to everyone in the project. The other gives Claude access for one session only, controlled by your cloud storage permissions. Sensitive data always uses the second.
Claude reads this at the start of every conversation in the project. Anyone with project access can see it. Use for skill files, brand guidelines, product knowledge — non-sensitive, always-needed context only.
The file lives in your cloud storage with its own access permissions. Uploaded at the start of a session when needed. Gone from Claude when the session ends. Never embedded in the project. Financial data, margins, sensitive pricing — always here.
Create one root folder — /company-claude/ or similar — in SharePoint, Google Drive, or Dropbox. Build this structure inside it:
| File type | Project Files tab | Cloud storage only |
|---|---|---|
| Brand voice, tone of voice rules | ✓ Always here | — |
| Product knowledge, specs, FAQs | ✓ Always here | — |
| Sales playbook, objection handling | ✓ Always here | — |
| Margin data, account-level pricing | ✗ Never here | ✓ Per session |
| Sales pipeline, deal values | ✗ Never here | ✓ Per session |
| Supplier-specific data, price lists | ✗ Never here | ✓ Per session |
| P&L, cashflow, management accounts | ✗ Never here | ✓ Per session |
| Customer personal data (GDPR) | ✗ Never here | ✗ Not in Claude |
Every team member can create private projects and attach their own files. This is fine for personal workflows. It becomes a problem when someone downloads a copy of product-kb.md or brand-voice.md and attaches it locally — that copy will drift out of date silently. When you update the source in cloud storage, their private copy doesn't update. The rule: if it describes how your company works, it lives in cloud storage and is always referenced from there. If it describes how one person works, it can be private.
Write your knowledge once. Make it available in every project. Stop re-explaining your business to an AI that can't remember between sessions.
A skill file is a Markdown (.md) document that gives Claude deep, standing knowledge about one topic. Think of it as the briefing document you'd hand a brilliant new contractor on their first day — written once, applied every time.
1. Purpose line — one sentence: what this file is and when Claude should apply it.
2. Core rules — 5–10 specific things Claude must always do or never do. Specific beats vague every time.
3. Before / after examples — 3–5 pairs showing the wrong version and the right version side by side. This is the single most powerful thing you can include.
4. Reference data — facts, lists, templates Claude needs to do the job. What would a new hire look up in week one?
5. Last updated date — one line at the bottom. Revisit every 6 months.
"Be professional and friendly"
"Write in our brand voice"
"Keep responses concise"
"Be helpful to customers"
"Never use exclamation marks in emails"
"Lead with the outcome, not the feature"
"Max 20 words per sentence in customer copy"
"Never use 'innovative' — use specific proof"
A separate category from company skill files. Private to you. Stored differently. Used when Claude needs to write as you specifically — not as the company.
Your personal voice file is distinct from the company skill files in Section 04. It captures how you write — your sentence rhythm, your level of formality, your characteristic phrases, what you never say. It belongs to you, not to the business. It should never be in the shared /skills/ folder.
Emails, LinkedIn posts, proposal intros, anything that sounds like you at your best. Recent is better than old.
Add anything it missed. Remove anything wrong. Save as your-voice.md to a private location — not the shared skills folder.
Strategy, CFO Advisor — these always write in your voice. For shared projects (Presentation Builder, Content), upload it per session when you specifically need it.
Per-session upload is not just for sensitive financial data — it is also the right pattern for personal context that belongs to one person but occasionally needs to inform a shared project. Drag your-voice.md into the conversation at the start of that session. It is available for that conversation only. Other team members using the same project in their own sessions will not have it — which is correct. The file stays private; the output is yours.
Your voice evolves. A voice file written today will feel slightly off in 18 months. Schedule a 30-minute refresh — paste new writing samples, run the same prompt, compare to the existing file, update what's changed.
Claude can be an exceptional financial thinking partner. The setup is simple. The one rule is that actual financial figures are never stored permanently in any project.
The risk is not Claude — it is the project files. A file permanently attached to a project is readable by anyone with access to that project. Keep the framework permanent. Keep the numbers per session. That is the entire security model.
Do not add it to the Team workspace. It should appear under "Your projects" only. If a finance manager also needs access, they create their own identically-configured private project.
Something like: "You are an experienced CFO advisor. I will share financial data for analysis. Be direct. Identify risks, flag weak assumptions, and always clarify what additional information you would want before making a real recommendation."
This contains your KPI definitions and measurement framework — no actual numbers. It makes Claude immediately familiar with how you run your business without any financial exposure.
Drop in the PDF or spreadsheet at the start of the conversation. When the session ends, it is gone from Claude's context. Never use permanent project attachments for live financials.
Financial analysis, scenario models, board summaries — save to your authorised-access-only cloud folder, not the general outputs folder the team can see.
Upload management accounts. Ask Claude to identify the top variances, likely causes, and questions to raise with your accountant before the next meeting.
"If revenue drops 20% next quarter, where do we have flexibility?" Claude runs the logic. You provide the numbers. Faster than a spreadsheet for structured thinking.
Turn raw financial data into a clear narrative for a board update, bank review, or investor brief — in your voice, with the right level of detail for the audience.
"Act as a CFO reviewing this plan. What are the three weakest assumptions?" Structured critical thinking on demand, before you commit to anything.
Claude's financial analysis is excellent for internal decision-making, stress-testing, and communication preparation. For decisions with legal or tax implications — acquisitions, pension changes, statutory filings — validate with your accountant or solicitor. Claude gets you 80% of the way there, faster. The professional provides the final 20% with legal accountability.
Three categories. Every team member learns them in five minutes. Most data security failures happen because people don't know the rules — not because they're careless.
| Status | Category | What this means | Examples |
|---|---|---|---|
| Safe | Public or product knowledge | Already public, or shareable with any customer. Safe as permanent project files in any shared project. | Product specs, brand guidelines, website copy, FAQs |
| Safe | Internal operational content | Internal processes and templates with no personal data and no financial figures. Safe for shared projects. | Sales scripts, process docs, proposal templates, onboarding checklists |
| Caution | Anonymised business data | Fine with names and personal details removed. Use a role-specific project, not a shared team project. | Meeting notes (anonymised), market analysis, internal strategy drafts |
| Caution | Customer account data | Only in a role-specific or account-specific project. Never in a shared team project. Anonymise where possible. | Account history, pricing agreements, customer correspondence |
| Never* | Financial statements & figures | Private project only. Uploaded per session. Never attached permanently. Outputs saved to authorised-access-only folder. | P&L, cashflow, management accounts, payroll, customer-level margins |
| Never | Personal data (GDPR / CCPA) | Any data covered by privacy law. Without a Data Processing Agreement, using this in Claude is a compliance risk. | Employee HR records, home addresses, personal emails, ID documents |
| Never | Passwords & credentials | No exceptions. No workarounds. Not even for "just a second." | System passwords, API keys, banking logins, SSO tokens |
* Financial figures are permitted in the private CFO Advisor project via per-session upload only — see Section 06.
Claude can connect to your email and cloud storage. The capability is real. The behaviour is less obvious than most people expect. Read this before enabling anything.
When you enable a connector — Microsoft 365, Google Workspace, or similar — Claude authenticates using your credentials. It can only reach what you personally have permission to access. It does not bypass your cloud storage folder permissions. The connector sees what you see, nothing more.
The risk is not Claude breaking into locked files. The risk is what project you are in when the connector is active. Any content Claude surfaces from your email or cloud storage appears in the conversation — and all conversation history is visible to everyone with access to that project.
No connector needed. Open the relevant project. Copy the email thread or document you want Claude to work with. Paste it into the conversation. You control exactly what context goes in. Outputs save to cloud storage as normal. This is safer, more predictable, and available immediately with no configuration.
An agent is Claude taking a sequence of actions to complete a goal. Not answering a question — doing a job. Here is how to think about it and when to build one.
The difference between using Claude and building an agent: using Claude is asking a question and getting an answer. An agent is giving Claude a goal, tools, and letting it figure out the steps. Build your project and skill file foundation first. Agents are what you add on top once that foundation is solid.
Paste a transcript → Claude extracts decisions, action items, owners, deadlines → structured summary ready to paste into your project management tool.
Paste a customer email → Claude classifies it → drafts an on-brand response → flags if it needs human review before sending. Review the draft, send or edit.
Paste a blog post → Claude produces a LinkedIn post, a 3-email newsletter sequence, and 5 social captions — all in your brand voice, in one pass.
One page. Everyone reads it and signs it. You review it every 90 days. This is what responsible AI adoption looks like for a small business.
Below is a template you can adapt. Replace [Company] with your name. The rules should work for most businesses without changes — but read them carefully and adjust anything that doesn't fit how you operate.
[Company] — AI Usage Policy
Version 1.0 · All staff · Review every 90 days
Claude is a secure AI assistant used for writing, research, analysis, customer service, and business advisory. It is accessed via Claude Desktop on the Team plan. Our data is not used to train Claude's models. Conversations do not persist after a session ends.
I have read and understood this policy. Name: _______________________ Date: ___________
The habits that separate consistent results from inconsistent ones. Not advanced techniques — fundamentals most business owners skip in the first 90 days.
Most people include one or two of these. The difference in output quality when you include all four is immediate and significant.
Show Claude what right looks like. Paste an example of a good email. Say "write it like this." Examples outperform descriptions every time.
For creative or strategic work, ask for 3 versions. You will nearly always pick parts of multiple — but you make a better decision than from one option.
If the first output isn't right, tell Claude exactly what to change. "Make the opening more direct." Refine the conversation — don't start over.
For complex decisions or analysis, add this phrase. It measurably improves output quality on difficult tasks by making Claude show its reasoning.
Claude can be confidently wrong about specific facts, dates, and technical details. Always verify anything you will publish or quote to a client.
When a prompt consistently produces excellent results, save it in your skill files or a prompt library document. Your best prompts are assets.
Don't try to do everything at once. This sequence gets you from zero to structured in three months without disrupting the day-to-day.
The businesses that stall on AI adoption usually make one of two mistakes: they do too much too fast and lose control, or they stay in experiment mode indefinitely and never build real systems. This sequence threads the needle.
Foundation before skills. Skills before agents. Agents before connectors. Every layer depends on the one below it being solid. An agent built on a poorly-configured project with vague skill files will produce inconsistent results at best. At worst, it will automate your chaos.
The businesses that get the most from Claude in year one are not the ones who moved fastest. They are the ones who built the foundation properly, let the team get comfortable, and then expanded deliberately. The compound effect of good skill files and well-configured projects grows quietly over time — and after six months the gap between structured and unstructured AI adoption inside a business is significant.
Write your brand-voice.md file and circulate the team usage policy. Those two things — knowing how to sound and knowing the rules — immediately change how your team uses Claude. Everything else can wait a week. Do those two things first.
Share freely · Adapt for your business · No attribution required
Accurate as of April 2026. Check claude.ai for current plan details.
Claude for Business Owners
2026 Edition · Practical Setup Guide
If you found this guide useful, I'd really appreciate a follow.
I share practical AI and business content on both platforms.