Business OS Playbook

Your Role as Operator

You are the person who keeps the AI infrastructure alive and growing after the BenAI team leaves. Think of yourself as the gardener, not the architect. The architecture is already built. Your job is to tend it, grow it, and help your team use it every day.

What the Operator Does

What the Operator Does NOT Do

Time Commitment

During the first 2-4 weeks, expect to spend 2-4 hours per week as you learn the system, build your first few skills, and get comfortable with the vault structure. Once things stabilize, this drops to 1-2 hours per week. Most of that time will be spent building new skills (which is genuinely fun once you get the hang of it).

What Success Looks Like

System Architecture

Your AI infrastructure has a simple, layered design. Understanding how these pieces connect will make everything else in this playbook click.

The Vault: Your Company's Brain

The vault is simply a folder of markdown files, organized by topic. It contains everything your AI needs to know about your business: who you are, what you do, how you talk, who your customers are, and how your processes work. Think of it like a company wiki that the AI reads every time it starts a conversation.

The vault is typically managed through Obsidian (a free note-taking app), but any folder of .md files works. The key principle: the more complete your vault, the smarter your AI becomes.

CLAUDE.md: The Bridge

This is a single file at the root of your vault. It acts as a table of contents and instruction set. When Claude opens a session, it reads this file first and learns where everything is located, what rules to follow, and how to navigate the vault.

If the vault is the library, CLAUDE.md is the librarian. It tells Claude: "Product info is in the /products folder. The ICP doc is at /sales/icp.md. Always use a professional tone. Never mention competitor X by name."

Projects: Organized Workspaces

Projects are separate workspaces inside your vault, usually organized by department or function. For example: /marketing, /sales, /operations, /hr. Each project can have its own skills, its own context files, and its own rules. This keeps things clean as your vault grows.

Skills: Saved Processes

A skill is a repeatable process that Claude can execute on demand. Instead of explaining the same task every time, you save it as a skill and trigger it with a slash command (like /write-newsletter). Skills are the core of your AI infrastructure. Section 4 is entirely dedicated to building them.

Connectors: Bridges to Your Tools

Connectors link Claude to external tools your team already uses: Gmail, Slack, your CRM, Google Sheets, calendars, and more. They allow Claude to read from and write to these tools without you copy-pasting data back and forth. Section 8 covers connectors in depth.

Scheduled Tasks: Agents That Work Without Prompting

Scheduled tasks are agents that run on a timer. You set them up once, and they execute automatically. Examples: categorize incoming emails every morning, prepare call briefings before meetings, generate a weekly report every Friday. Section 7 walks through how to set these up.

Daily Operations

The day-to-day habits that keep your AI infrastructure humming. These are the three things you will do most often.

Starting a Session

Every time you open a new Claude conversation, the first thing to do is point Claude to your vault folder. This ensures Claude reads your CLAUDE.md and knows where everything is.

In Claude Desktop, you can also set your vault as the default project folder. This way Claude automatically loads context every time, and you never need the manual step.

Telling Claude to Remember Things

Whenever you discover something Claude should know permanently, tell it to save that information to the vault. This is one of the most important habits you can build: every correction, every preference, every rule should be saved.

Checking Your CLAUDE.md

CLAUDE.md is the single most important file in your vault. It is the file Claude reads first, and it determines how well Claude can navigate everything else. Make it a habit to review it at least once a month.

When to Update CLAUDE.md

• You add a new folder to the vault
• You rename or reorganize existing folders
• You add a new skill or command
• You add a new connector
• You want to change a global rule (tone, formatting, behavior)
• Claude consistently can't find a file it should know about

# Company AI Brain - [Your Company Name]

## Navigation
- /products/ -- Product descriptions, features, pricing
- /sales/ -- ICP, objection handling, pitch decks
- /marketing/ -- Content calendar, voice guide, brand assets
- /operations/ -- SOPs, vendor contacts, tool docs
- /team/ -- Org chart, roles, preferences per person

## Global Rules
- Always use a professional but conversational tone
- Never mention competitor names in external content
- Default to bullet points over long paragraphs
- When uncertain, ask before assuming

## Skills
- /write-newsletter -- Weekly newsletter from content calendar
- /prep-call -- Pre-meeting research and talking points
- /draft-proposal -- Client proposal from template

## Active Connectors
- Gmail (marketing@company.com)
- Google Calendar
- Slack (#marketing, #sales channels)
- HubSpot CRM

Building Skills

Skills are the core of your AI infrastructure. Every skill you build is a permanent time-saver for your entire team. This is the most important chapter in the playbook.

When to Build a Skill

Not every task needs a skill. Build one when:

• You do the same task more than twice. If it happened twice, it will happen again.
• The process has clear steps. Even if the steps are complex, they should be definable.
• Output quality matters. Skills ensure consistency. Don't bother for throwaway tasks.
• Multiple people need to do this same process. One skill serves the whole team.

The Manual-First Method Recommended

The best way to build a skill is to not try to build a skill. Instead, do the task once with Claude in a normal conversation, correcting it as you go. Then save the process.

Skill Building Framework (7 Steps)

Whether you use the manual-first method or build from scratch, every skill needs these seven components.

# Marketing Newsletter Writer
Trigger: /write-newsletter
Trigger (natural): When user asks to write, draft, or create a newsletter

# Connectors Required
- Google Calendar (to pull today's meetings)
- HubSpot CRM (to pull contact/company data)
- Gmail (to pull recent email threads with attendee)

# Process

## Step 1: Topic Selection
- Action: Read /marketing/content-calendar.md for this week's assigned topic
- Human input: Confirm topic or provide override
- Output: Confirmed newsletter topic

## Step 2: Angle Generation
- Action: Read /marketing/voice-guide.md and /marketing/audience.md
- Generate 5 unique angles for the topic
- Human input: Select preferred angle (or request more options)
- Output: Chosen angle

## Step 3: Outline
- Action: Create 3 outline options (short/medium/long format)
- Reference: /marketing/newsletter-examples/ for structure patterns
- Human input: Select outline and suggest modifications
- Output: Approved outline

## Step 4: Draft
- Action: Write full newsletter draft based on approved outline
- Reference: /marketing/voice-guide.md (MUST re-read before writing)
- Reference: /marketing/newsletter-examples/best-performers/
- Output: Complete draft with subject line, preview text, body

## Step 5: Review
- Action: Present final draft for approval
- Human input: Approve, request edits, or restart from a specific step
- Output: Final approved newsletter

# Reference Files
- MUST read /vault/context/voice-personality.md before generating any copy
- MUST read /vault/context/icp.md before any audience-targeting step
- Read /vault/skills/newsletter/examples/ for approved output examples
- Read /vault/context/brand-guidelines.md for formatting rules

# Progressive Updates

## Auto-Update Rules
- When user says "don't do X" or "never do X," automatically add
  to the Rules section of this skill
- When user says "always do X" or "make sure to X," automatically
  add to the Rules section
- When user corrects tone/style, update voice-personality.md reference

## Learning from Approved Outputs
- When user approves a final output, save it to
  /vault/skills/[skill-name]/examples/approved/
- Use approved examples as reference for future runs
- The examples folder is the most impactful reference for quality

## Version Notes
- After any rule update, add a dated note:
  "Added [date]: [rule description] (based on user feedback)"

Core Reference Files Every Company Should Have

These are the foundational context documents. If you are missing any of these, build them with Claude (more on that below).

Testing Skills (Evals)

Once a skill is built, you want to measure and improve its output quality systematically. This is called running evals.

A/B Testing

A/B testing is for skills that already work. You have a "version A" that produces decent output, and you want to see if "version B" is better.

Building Commands

Commands chain multiple skills together into multi-step workflows. Think of skills as individual tools and commands as the workbench that arranges them into a production line.

When to Use Commands

Build a command when you have a workflow that involves multiple skills in sequence. Examples:

• Content repurposing: Take a blog post, run the newsletter skill, then the LinkedIn skill, then the infographic skill
• Client onboarding: Run the welcome email skill, then the account setup checklist, then the kick-off meeting prep
• Weekly reporting: Pull analytics data, generate the report, create the executive summary, draft the distribution email

How to Build a Command

Key Principles

• Output flows forward. The output of skill 1 becomes the input (or context) for skill 2.
• Human checkpoints optional. You can add approval gates between steps, or let the whole chain run uninterrupted.
• Commands can be triggered with a slash command just like individual skills.
• Commands can be scheduled to run autonomously (see Section 7).

Building Plugins

Plugins are bundles of related skills, commands, agents, and connectors, packaged together for easy sharing and installation. Think of a plugin as a department-in-a-box.

When to Use Plugins

• Organizing by department: Bundle all marketing skills, commands, and connectors into a "Marketing Plugin"
• Sharing with team members: A new hire installs one plugin and gets everything they need
• Distributing across companies: If you create something useful, you can package it for others

How to Create a Plugin

How to Share Plugins

• Export as zip: The simplest method. Claude generates the plugin folder, you zip it, share via Slack/email/drive.
• Deploy via GitHub: For versioned sharing across teams. Push the plugin folder to a repo, and team members pull updates.

How to Install a Plugin

In Claude Desktop: Customize > + > Upload Plugin. Select the plugin folder or zip file. Claude reads the plugin manifest and installs all included skills, commands, and connector requirements.

Scheduled Tasks

Scheduled tasks are agents that work without being prompted. Set them up once, and they execute on a schedule: every morning, every week, after specific triggers. This is where your AI infrastructure starts working for you instead of the other way around.

Three Ways to Schedule

Best Use Cases with Examples

Scan my inbox for all emails received since yesterday.
Categorize each into: Urgent, Action Needed, FYI, or Skip.
Present a prioritized summary with sender, subject, and
recommended action for Urgent and Action Needed items.

Check my calendar for today's meetings. For each meeting:
1. Look up each attendee in CRM (if available)
2. Research their company and recent news
3. Prepare 3-5 talking points relevant to the meeting purpose
4. List potential questions or objections they might raise
5. Save as a call prep document for today

Check for new meeting transcripts since last run.
For each transcript:
1. Extract all action items with owner and deadline
2. Identify follow-up commitments made to external parties
3. Note key decisions made
4. Format as a structured summary with action items at top
5. Save to /vault/meetings/[date]-[meeting-name].md

Check Stripe for any failed payments in the last 24 hours.
For each failed payment:
1. Look up the customer name and email
2. Check payment history (first failure vs. recurring)
3. Draft a friendly follow-up email (different tone for
   first-time vs. repeated failures)
4. Save as Gmail draft -- do NOT send
5. Flag for my review in the morning summary

Scan these sources for trending topics from the past week:
- [Source 1 URL]
- [Source 2 URL]
- [Source 3 URL]

Read /vault/context/icp.md to understand our audience.
Generate 5 newsletter topic ideas that:
1. Are relevant to our ICP
2. Connect to current industry trends
3. Include a unique angle (not just news recap)
4. Come with a rough 3-point outline each
Save to /vault/marketing/newsletter-ideas/[date].md

Connectors & API Guide

Connectors link Claude to the tools your team uses every day. The right connector turns Claude from a smart chatbot into an integrated assistant that can read your email, update your CRM, and post to Slack.

Priority Order

When connecting a new tool, always try these options in order. The first available option is always the best choice.

How to Install an MCP Server

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "@package/mcp-server"],
      "env": {
        "API_KEY": "your-api-key-here"
      }
    }
  }
}

Essential Connectors by Department

Apify Setup (Step by Step)

Apify is one of the most useful connectors for marketing and sales teams. It provides web scrapers for YouTube, LinkedIn, Google, and hundreds of other sites.

How to Find if a Tool Has an MCP

• Google it: Search "[tool name] MCP server" (e.g., "Notion MCP server")
• Check directories: Visit smithery.ai or mcp.so for curated MCP listings
• Check the tool's own docs: Many SaaS companies now publish their own MCP servers
• Ask Claude: "Is there an MCP server for [tool name]?"
• Build your own: If no MCP exists, ask Claude to build one using the MCP Builder skill. You will need the tool's API documentation and an API key.

Context Management

Your vault is a living document. The more current and complete it is, the better your AI performs. Context management is the ongoing practice of keeping this brain up to date.

Adding New Context

• Create new markdown files in the appropriate vault folder
• Update CLAUDE.md if you added new folders or if the new file needs special routing
• Let Claude help: use the prompt below to create context docs efficiently

Keeping Context Current

The biggest risk to your AI infrastructure is stale context. Here are the triggers for updating:

Skills Should Point to Vault, Not Embed

Team Rollout

Getting your team on board is about reducing friction, not mandating adoption. Make it so easy and useful that people want to use it.

Adding a New Team Member

Permission Settings

• Connectors: Set commonly-used connectors (Gmail, Calendar, Slack) to "Always Allow" for trusted actions. This removes the approval prompt each time and makes the experience faster.
• Dispatch for Mobile: If team members need to trigger agents from their phone, set up Dispatch. This allows phone-based triggers that execute on the computer running Claude Desktop.

Troubleshooting

Quick reference for the most common issues and their fixes. 90% of problems fall into these categories.

Best Practices Cheat Sheet

The twenty rules that matter most. Print this page and keep it at your desk.