Step 2 · Build the foundation
Build your company knowledge base
Most people open Claude and start typing. That works for one-off questions. It doesn't work for running a business.
Create a file named CLAUDE.md in the folder where you keep your work documents. This is the first file Claude reads every time it opens that folder. Think of it as a cheat sheet for a new hire; the way you'd brief someone on their first day.
Include these sections:
- Who we are: one paragraph on what you do, who you serve, and what makes you different
- How we talk: tone, formality, words you always use, words you never use
- Services and pricing: what you offer, typical ranges, how you structure proposals
- Our team: who does what
- Important rules: things Claude should never say or do; things that are always true
A good CLAUDE.md takes 15 minutes to write. The difference is immediate. Not perfect; but the gap between generic output and output that sounds like your company shrinks from the first conversation.
The same prompt with and without a CLAUDE.md produces completely different results. One sounds like a machine. The other sounds like your best salesperson on a good day.
Step 3 · Define the details
Define your brand voice and context
The 15-minute CLAUDE.md gets you started. A mature knowledge base gets you output you can use without editing.
Go deeper on these four areas:
- Brand voice and communication style. How do you talk to clients versus prospects versus your team? Do you lead with data or stories? Formal or casual? What does an email from your company sound like? This is the section most people underestimate.
- Services, pricing, and sales process. What do you sell? How is it structured? How do you handle objections? When Claude knows your sales playbook, it can draft proposals, follow-ups, and objection responses that match how your team actually sells.
- Client and industry knowledge. Who are your typical clients? What industries? What pain points do they have? When Claude knows your clients, it writes to them; not at them.
- What not to do. This section is often the most valuable. “Never promise delivery dates without checking with the project manager.” “Never mention competitor names.” “Never discount below 15% without approval.” The “don'ts” prevent the kind of mistakes that damage client relationships.
Writing a CLAUDE.md forces you to put on paper things your company has never written down. That documentation is valuable far beyond AI.
Step 4 · Connect your stack
Connect your tools and team
Your business doesn't live in one folder. Your clients are in your CRM. Your conversations are in Slack and email. Your calendar holds your schedule. In this step, you give Claude access to the tools where your business actually runs.
Start with one integration; usually Slack or email, where you spend the most time. The process takes about 15 minutes:
- In Claude's settings, find the integrations or MCP section
- Select the tool; Claude will walk you through authorization
- Choose what Claude can access; start narrow, maybe one or two channels
- Test it; ask Claude to summarize recent messages
- Do something useful; ask it to flag what needs your attention
Available connections include Slack, Gmail, Outlook, Google Calendar, Google Drive, HubSpot, Asana, and more. Each integration is a separate authorization. You can connect Slack without connecting email. You can grant read access without write access. You can revoke access anytime.
When you connect any tool, Claude gets read access by default. Sending messages, creating records, and modifying files require explicit write access, and Claude asks for approval before taking action. You stay in control; Claude proposes, you approve.
One sentence, five tools, ten minutes of prep work; done in seconds. “Check my calendar for tomorrow. For each meeting, pull up the contact in HubSpot, check open issues in Slack, and create a one-page briefing.”
Step 5 · Automate the repeat work
Build reusable skills and workflows
Integrations give Claude access to tools. Skills teach it how to do specific types of work.
A skill is like a recipe. It tells Claude: “When someone asks you to write a proposal, follow these steps, use this format, check these files, and produce this output.” Without skills, you explain what you want every time. With skills, you say “write a proposal for this client” and Claude follows the playbook.
Pre-built skills are available in a public library; free, open source, and ready to install. The real power comes when you customize them for your business.
Common skills to build first:
- Proposal skill: reads the client brief, pulls your pricing, uses your template, produces a ready-to-review draft
- Meeting prep skill: reads your calendar, pulls background on attendees, checks recent communication, and builds a one-page briefing
- Client follow-up skill: analyzes your communication style, references previous conversations, and drafts next steps
Installing a pre-built skill takes a few clicks. Customizing it to match your proposal format, pricing structure, and follow-up cadence requires editing the skill configuration. It's not programming, but it's not clicking a button either. Building a custom skill from scratch; for a workflow unique to your company; is where most non-technical users need structured guidance.
The gap between “I know what I want Claude to do” and “I can write the instructions that make it do it reliably” is wider than it looks.
Step 6 · Protect what matters
Scale with guardrails
Claude Code is a powerful tool. The difference between “this is incredible” and “this just sent the wrong thing to a client” comes down to the guardrails you put in place. This step covers the decisions that separate experimentation from a system your team can rely on.
Permission levels. Start with Read Only. Claude can look at files and data but can't change anything. Let it prove itself as a reader and analyst before you let it take action.
- Level 2 is Create with Approval. Claude drafts new files and proposes changes, but asks for your approval before saving or sending. Most business owners settle here for daily work.
- Level 3 is Autonomous Action. Claude takes actions without asking. Only appropriate for tasks where you've verified consistent reliability and the consequences of a mistake are low. Saving a meeting prep document to your folder? Low risk. Sending a follow-up email to a client? Keep approval on.
Model selection. Match the model to the task. Opus is the heavyweight; use it for complex analysis and important client deliverables. Sonnet is the everyday workhorse; emails, summaries, drafts, meeting prep. Haiku is the quick and light option; reformatting, lookups, short drafts. Don't use Opus for everything.
Data framework.
- Green light: marketing materials, general SOPs, proposals with client names redacted, internal drafts, research, and scheduling.
- Yellow light: client-specific information, financial data, employee information, and anything under NDA. Use local mode, keep permissions tight, and understand your plan's data policies.
- Red light: passwords, API keys, Social Security numbers, credit card numbers, health records, and anything where a leak would create legal liability. No AI tool should have access to red light data.
Team consistency. When you scale beyond yourself, the real questions emerge. Are five people on your team getting the same quality output? Are they all using the same brand voice and pricing? Without shared CLAUDE.md files and shared skills, every person's Claude is a different AI with different instructions. Set up shared standards, role-specific permissions, and a clear owner who keeps documentation current.
The real power is compounding. When Claude has your business context, is connected to your tools, follows custom skills, and remembers your preferences; the output quality isn't five times better. It's qualitatively different.