The Claude Code Playbook

Why this matters

If you use AI the way you used Google, you ask a question, copy the answer, and paste it somewhere else. It helps, but you’re still the one doing the work.

Claude Code works differently. It operates inside a folder on your computer and connects to the tools your business already runs: accounting, customer relationship management (CRM), email, calendar, and files. It can read the numbers, draft the follow-ups, build the report, and save the file. On a schedule, it can do the same job again tomorrow without being asked.

That lets you hand work off instead of asking questions about it, and handing work off well is how you become the least important person in your business.

What you’re building over 30 days

• A collections list every morning, with the follow-ups already drafted
• A morning brief waiting before your team starts
• One internal tool your team uses
• A repeatable way to hand off big questions and pressure-test big decisions

You won’t write any code. You’ll describe the work, approve plans, and check results, which you already do with your team.

Part 1

Four things to know first

1. Where Claude Code runs

Claude Code started in the terminal, the plain text window developers use. You don’t need the terminal anymore. You can run Claude Code in three places:

• The Claude desktop app for Mac or Windows, in its Code tab. This is the easiest place to start if you’ve never used a terminal.
• The terminal, if you’re comfortable there.
• The cloud, through Claude Code on the web and cloud routines. These keep working when your laptop is closed, but they need a GitHub repository set up first, so treat them as a later step.

2. Tokens are the unit of usage

Claude reads and writes in tokens, about three-quarters of a word each. Everything counts: your instructions, every file and record Claude reads, and everything it writes back. On a subscription, tokens count toward your plan’s usage limits. On the pay-per-use API, tokens are the bill.

That’s why a vague request costs more than a specific one. “Look at my receivables” sends Claude exploring. “List invoices more than 30 days overdue, largest first” goes straight to the answer.

3. Connectors, MCP, and APIs

Most business software has an API (application programming interface): a documented door that lets other programs read and change its data. A connector is a ready-made link between Claude and one tool. Connectors are built on an open standard called the Model Context Protocol (MCP), and behind the scenes most of them use that tool’s API.

There are connectors for many of the tools small to medium sized businesses (SMBs) run, including QuickBooks Online, HubSpot, Stripe, Gmail, Google Calendar, and Google Drive. When a tool has no connector, Claude Code can often work with its API directly. Part 4 walks through connecting your software each way.

Either way, you do the signing in and create any keys yourself. Never paste a password or key into the chat.

4. How you pay

Prices before tax. Anthropic updates plans often, so confirm at claude.com/pricing.

The other way to pay is Anthropic’s own API: no subscription, and you pay for the tokens you use. It suits automated jobs that run at volume. You probably won’t need it on day one.

If you hit your plan’s limit, you can turn on usage credits in claude.ai under Settings, then Usage. They’re billed at standard API rates, and Adjust limit on the same page sets a monthly spend limit so you stay in control of the cost.

Part 2

Set up once

Allow 30 to 45 minutes for this part (our estimate). Connecting your software comes after the guardrails, in Part 4.

1. Get a paid plan

   Sign in at claude.ai and choose Pro or Max. The Free plan doesn’t include Claude Code.

2. Install Claude Code

   The easiest route is the Claude desktop app. Download it from claude.com/download, sign in with the same account, and open the Code tab.

   If you’d rather use the terminal, open Terminal on a Mac (press Cmd + Space, type Terminal, and press Return) or PowerShell on Windows, then paste the line for your computer:

   curl -fsSL https://claude.ai/install.sh | bash

   irm https://claude.ai/install.ps1 | iex

   Check it worked by typing claude --version. Then type claude and sign in through the browser window that opens.

3. Make a home base

   Create a folder called business-ops in your home folder. Keep it out of iCloud Drive, Dropbox, OneDrive, and Google Drive, because sync tools struggle with files that change constantly and can slow your whole computer down.

   In the desktop app, choose this folder when you start a session. In the terminal, run mkdir ~/business-ops, then cd ~/business-ops, then claude.

4. Write your business brief once

   Claude Code automatically reads a file named CLAUDE.md in your folder at the start of every session. Put your business in it once and you stop re-explaining yourself. It’s also the first piece of your business knowledge written down somewhere other than your head.

   Interview me so you can write a CLAUDE.md file for this folder. It's the brief you'll read at the start of every session, so it should explain how my business works.
   
   Ask one question at a time, about 12 in total, covering:
   - What we do, who we serve, where, and roughly how big we are
   - Our team: key roles, who owns what, and who approves spending
   - The tools we use for accounting, CRM, scheduling, email, and files
   - How we talk to customers: tone, words we use, and words we never use
   - Rules you must always follow (for example: never send, pay, or delete anything without my approval)
   - The numbers I watch, and what a good week looks like
   
   Then write CLAUDE.md in plain language, under two pages, and show it to me before you save it.

Part 3

Six guardrails before you connect anything that matters

Claude Code can act inside your business, which is what makes it useful and also what makes it risky. These six rules cost nothing to follow and guard against the mistakes that are hardest to undo.

1. Read first, write later

   Every new connection starts read-only. Claude drafts; a person sends, pays, deletes, or changes records. Loosen one permission at a time, after a week of clean runs, and try new builds on an export or a copy before they touch live data.

   Check your permission mode too. Terminal sessions on Pro, Max, and Team plans start in Auto mode, where Claude takes most actions without asking you first. While you set up a new system, switch to Manual mode so it asks before it acts: press Shift+Tab to cycle modes in the terminal, or use the permission mode picker in the desktop app.

2. Keep secrets out of the chat

   Never paste passwords, card numbers, banking details, or API keys into a conversation. Sign in through the connector, or save keys in a file on your computer as shown in Part 4.

3. Tie the numbers out

   For the first few runs, check Claude’s totals against the same report in the source system. If they don’t match, stop and fix the setup before you trust it.

4. Make failures loud

   Tell every scheduled system to say so at the top when a source fails or comes back empty, and never to report success on nothing.

   Field note

   A backup I set up once finished, reported success, and had saved zero files. Nothing in the output said so. I only caught it by checking the file count.

5. Test it the way others will see it

   Check anything a customer or employee opens while signed out and on a phone before you call it done. A page that looks right on your desktop can be broken on theirs.

6. Share the minimum

   Customer and employee personal information is covered by privacy law (in Canada, PIPEDA or your province’s equivalent). Connect only what the job needs, and never put an internal tool on the open internet without a login.

Part 4

Connect your software

Claude can only work with what it can reach. Try these three routes in order and stop at the first one that works for your tool. Whichever route you use, start with read-only access.

Route 1: A ready-made connector

This is the easiest route, and the place to start with tools like QuickBooks Online, HubSpot, Stripe, and Google Workspace.

1. Go to claude.ai/customize/connectors (Customize, then Connectors) and find your tool in the directory.
2. Connect it and sign in to that tool with your own account. Read what access it asks for before you approve it.
3. In Claude Code, type /mcp. Connectors from your Claude account show up in that list, marked as coming from claude.ai.

Route 2: A custom connector

Some software companies publish their own MCP server without appearing in the directory. Search their help pages for “MCP”, or ask their support team for an MCP server URL.

1. In claude.ai, go to Customize, then Connectors, click the + button, and choose Add custom connector.
2. Paste the MCP server URL, save it, and sign in when prompted.
3. Check it in Claude Code with /mcp, the same as Route 1.

Prefer the terminal? This command adds a remote MCP server for every folder you work in. Then type /mcp inside Claude Code to sign in.

claude mcp add --transport http --scope user [NAME] [MCP SERVER URL]

Route 3: The software’s own API

No connector at all? If the software has an API, Claude Code can write a small script that uses it. You create the key yourself, and Claude never needs to show it to you or anyone else.

1. In your software’s settings, look for API keys, developer access, or integrations, and create a key. If it offers read-only access, choose that.
2. Ask Claude Code to create a file named api-keys.env in your business-ops folder with an empty line for the key, such as ACCOUNTING_API_KEY=.
3. Open that file in a text editor (TextEdit on a Mac, Notepad on Windows), paste the key after the equals sign, and save it. Don’t paste the key into the chat.
4. Give Claude the prompt below.

Connect to [SOFTWARE] using its API. The key is saved in api-keys.env in this folder as [VARIABLE NAME]. Never print the key, copy it into another file, or include it in anything you save or show me.

1. Find [SOFTWARE]'s official API documentation and tell me, in plain language, what it lets us read and change.
2. Build the smallest script that reads [WHAT YOU NEED, E.G. OPEN INVOICES MORE THAN 30 DAYS PAST DUE]. Read-only: no creating, editing, or deleting.
3. Run it once and show me the first five results so I can check them against [SOFTWARE].
4. Save it as a reusable skill I can run by typing /[NAME].

If [SOFTWARE] has no public API, or needs a plan upgrade to use it, tell me before you build anything and suggest the simplest alternative.

The key lives on your computer, so this route works for jobs that run there. A cloud routine needs the key added to its environment settings instead.

Still stuck? Zapier’s connector reaches thousands of apps (check its pricing for your volume), or export a CSV (a spreadsheet file) into your folder and point Claude at it.

Test the connection

Whichever route you used, ask Claude Code:

What can you see in [INVOICING TOOL]? Read only. Don't change anything.
Give me our total overdue receivables and the five largest overdue invoices.

Compare the answer to the same report inside your software. If it matches, you’re ready for System 1.

Part 5

The five systems

Install them in this order. Each one builds the habits and connections the next one needs. For every system you get what it does, what it needs, how to measure it, the prompt to paste, and what to expect.

Replace anything highlighted in [BRACKETS] with your details. Everything else can be pasted as written. Setup times are our estimates; yours will depend on how many tools you connect.

You are my Collections Assistant.

About my business: [ONE SENTENCE: WHAT WE DO AND FOR WHOM]
Invoicing tool: [QUICKBOOKS ONLINE / JOBBER / SERVICETITAN / OTHER]
Typical invoice: $[AMOUNT]. Payment terms: [NET 15 / NET 30 / DUE ON RECEIPT]
Follow-ups are signed by: [NAME AND ROLE]

Build a collections routine that:
1. Reads every open invoice more than [7] days past due (read-only)
2. Sets aside anything disputed, on a payment plan, or contacted in the last seven days, and lists those separately with the reason
3. Ranks the rest by amount owed, then by days overdue
4. Drafts a short, friendly, firm email for each: invoice number, amount, due date, how to pay, and one clear next step. No threats, no guilt.
5. Firms up the tone at 30, 60, and 90 days, and flags anything over 90 days for a phone call from me
6. Saves collections-YYYY-MM-DD.csv with: customer, invoice, amount, days overdue, last contact, draft email
7. Starts its reply with three lines: total overdue, the three largest balances, and what changed since the last run

Rules: never send, edit, or delete anything in my invoicing tool. Drafts only. If you can't reach it, say so first and stop. If there's no connector for my tool, suggest the simplest alternative (such as a daily CSV export) before you build.

Save this as a reusable skill I can run by typing /collections.

Before you build anything, ask me whatever you need to get it right the first time.

You are my Morning Brief Analyst.

About my business: [ONE SENTENCE]
Who reads the brief: [ME / MY LEADERSHIP TEAM]. We start at [TIME].
Sources, two to four to start: [E.G. QUICKBOOKS ONLINE DEPOSITS, JOBS COMPLETED, NEW LEADS IN HUBSPOT, GOOGLE REVIEWS]

Build a brief that:
1. Pulls yesterday's numbers from each source (read-only)
2. Compares each number to the same day last week and to the four-week average
3. Flags anything outside its normal range and says by how much
4. Opens with the top three wins, top three problems, and top three actions for today, one line each
5. Ends with a sources line showing what worked. If a source fails or comes back empty, say so at the top. Never skip it silently.
6. Arrives as [AN EMAIL TO ME / A MESSAGE IN OUR TEAM CHANNEL] before [TIME]

Recommend whether this should run as a local scheduled task on my computer or as a cloud routine, and explain the trade-off in plain language.

Walk me through connecting each source, one at a time, before you build.

You are my Internal Tools Builder. I'm not a developer, so explain everything in plain language.

The tool: [ONE SENTENCE, E.G. "A QUOTE CALCULATOR OUR ESTIMATORS FILL IN ON SITE THAT PRODUCES A BRANDED PDF QUOTE"]
Who uses it, and how often: [ROLES AND FREQUENCY]
Systems it connects to: [LIST]

Before you write any code:
1. Tell me whether a tool I already use, or a free or low-cost product, already does this. If one does, say so and stop.
2. Ask me what you need to know.
3. Show me a one-page plan: what it does, what it looks like, where it runs, what it costs per month, and who can see the data.
4. Wait for my approval.

When you build:
- Anything with customer or financial data sits behind a login
- Keep it simple enough that someone else could maintain it
- Test it yourself, then walk me through testing it as a real user, including on a phone
- Finish with a one-page guide for my team and a note on how it's backed up

You are my Strategic Analyst, working unattended. I won't be available to answer questions, so make reasonable assumptions and list every one.

The question: [YOUR QUESTION]
Data you can use: [E.G. QUICKBOOKS ONLINE (READ-ONLY), HUBSPOT (READ-ONLY), THE FILES IN THIS FOLDER, PUBLIC WEB SOURCES]

Standards:
- Every outside number gets a source link. Anything you estimate is labelled as an estimate.
- Keep facts, estimates, and opinions clearly separate.
- If the data can't answer the question, say so directly instead of stretching it.
- Stay focused: no more than [20] web searches.

Save one report in this folder as [SHORT-NAME]-analysis-YYYY-MM-DD.md with:
1. The answer, in one sentence
2. A five-bullet summary
3. The analysis, as deep as it needs to be
4. Every assumption you made, so I can challenge them
5. What would change the answer
6. The next three actions if I say yes

You are the chair of my Decision Board.

The decision: [TWO OR THREE SENTENCES]
Amount at stake: $[AMOUNT]. Decision needed by: [DATE]
Useful files in this folder: [LIST]

Run five sub-agents in parallel. Each works alone and must not see the others' work:
1. Market: demand, competitors, pricing, and whether the timing is right
2. Money: a three-year profit and loss with low, base, and high cases, the cash needed, and months to break even
3. Operations: the people, equipment, systems, and time it takes, and what it pulls away from today's work
4. Risk: what could sink this (legal, financial, market, or execution), with a way to reduce each risk
5. Fit: whether this strengthens the core business or distracts from it, and what else this money and attention could do

Each agent: stay focused (no more than [10] web searches), list its three biggest uncertainties, and save its findings to decision-[TOPIC]/[ANGLE].md.

When all five are done:
- Make the call: GO, NO-GO, or NOT YET
- Give the three strongest reasons
- Name the one or two things that would change the answer
- List the next three actions if I say go
- Save the summary as decision-[TOPIC]/summary.md

Part 6

The 30-day rollout

Don’t install all five at once. Stack them, one a week.

Follow the rollout and you’ll have five systems running by day 30. Keep tuning them over the next two months so they fit the way your business runs.

Part 7

Keep usage and costs predictable

• Be specific

  Name the tool, the records, the date range, and the output you want. Specific requests use fewer tokens and get better answers.

• Start fresh when you switch jobs

  Type /clear. Everything left in a conversation gets carried into every new message.

• Watch the gauges

  /context shows how full the current conversation is. /usage shows how much of your plan’s limits you’ve used, plus your usage-credit spend if you’ve turned credits on. Long conversations are compacted automatically, but a fresh start with a clear request is cleaner.

• Match the model to the job

  Claude Code starts on Opus 5.5 on every paid plan. Type /model to switch.

  Model	Use it for	API price per million tokens (input / output)
  Haiku 4.5	Quick, simple, repetitive jobs	$1 / $5
  Sonnet 5	Everyday work, such as the daily Cash Collector and Morning Brief runs	$2 / $10
  Opus 5.5	The default: building tools, messy data, strategy, and Systems 3 to 5	$4 / $20
  Fable 5.1	The hardest, longest-running work, when the stakes justify it	$10 / $50

  On the API, Opus 5.5 costs twice as much as Sonnet 5 per token, and Fable 5.1 costs two and a half times as much as Opus 5.5. Anthropic’s pricing guide suggests Haiku for simple tasks, Sonnet for most production work, and Opus for the most complex reasoning. On Pro, Fable 5.1 is billed as usage credits rather than coming out of your plan. For simple jobs, you can also turn the effort level down with /effort.

• Cap parallel work

  Tell Claude how many agents to run and how many searches each can make. Without a cap, one big question can use a large share of your limit.

• Decide your ceiling before you need it

  Before you turn on usage credits, set a monthly spend limit in claude.ai under Settings, then Usage (click Adjust limit). For an API key, set the limit in the Claude Console. Then check /usage weekly.

Start this week

Pick one system, get it working, and add the next one once the first runs without you.

Version history

What changed

Claude changes quickly. This playbook carries a version number and the date of each update, so you can tell whether what you’re reading is current. If something here doesn’t match what you see, trust the app and tell us in the community.

• Version 2.1 | September 24, 2026Moved to The Business Growth Factor website. Same content, with a new closing invitation to the community.
• Version 2 | September 23, 2026Opus 5.5 became Claude Code’s default, so the model table and prices changed. Added Part 4, Connect your software. Added current scheduling details (Keep computer awake, catch-up runs, and cloud routines) and a monthly cap for usage credits.
• Version 1 | September 17, 2026First release.