From Spreadsheet to Live Dashboard by Lunch: Inside the Builder Track

Claude Code for business means pointing Anthropic's agentic coding tool at your own numbers. You describe the report or dashboard you want in plain English, and the agent plans the approach, writes the code and checks that it works while you steer and approve the plan. Going from a spreadsheet to a first live dashboard in one morning is a realistic target if you keep the path short: a CLAUDE.md file with the rules, one data source connected read-only, a plan you approve before any code changes, and a security review before real data goes anywhere near the agent.

That is the path the Builder Track laid out at the AI Business Retreat on September 15, 2026, at The Island in Spooner, Wisconsin. Here is what it covered and how to follow it with your own data.

What the Builder Track covered, session by session#

The retreat ran three tracks side by side: Leadership, Sales and Marketing, and Builder. The Builder Track was the hands-on one, aimed at developers, technical leads and IT directors. The bootcamp page widened that to "hands-on builders, analysts, and accidental developers." The track was taught by Tim Holum, Coulee Tech's software developer, who builds the production agent systems the rest of the retreat talks about.

The two morning sessions are the part you can repeat at your own desk. For which numbers belong on the screen, see our guide to building a one-screen business dashboard with AI. This article covers how the screen gets built.

Claude Code for business, step by step: the morning bootcamp#

What Claude Code is#

Anthropic describes Claude Code (opens in a new tab) as an agentic coding tool that reads your codebase, edits files, runs commands and integrates with your development tools. It runs in a terminal, in code editors such as VS Code, in a desktop app and in the browser. Most of those require a Claude subscription or an Anthropic Console account. Unlike a chat window, it works directly in a project folder instead of handing you code to paste.

The bootcamp's promise was "AI writes the code, you steer." The session page set the bar at comfort with spreadsheets and a willingness to read what the AI writes. You do not have to be a programmer. You do have to review.

The CLAUDE.md file#

The bootcamp opened with setup and "the CLAUDE.md file that changes everything." A CLAUDE.md file (opens in a new tab) is a plain-text markdown file in your project folder that Claude reads at the start of every session. Typing /init generates a starter version. Anthropic suggests keeping each file under 200 lines and notes that specific, concise instructions are followed more consistently.

For a first dashboard project, a useful CLAUDE.md is short and concrete:

Project: weekly sales dashboard
- Source data: exports/sales-2026.csv. Read-only. Never edit, move or overwrite it.
- Output: one web page with four charts: revenue by week, open quotes,
  top 10 customers, overdue invoices.
- Every number on the page must trace to a column or query we can point to.
- Ask before installing anything or connecting to anything outside this folder.
- Plain-language labels. No jargon.

One caution from the same documentation: CLAUDE.md is context, not enforced configuration. Claude treats it as instructions, not a lock. If something must never happen, block it with a deny rule in Claude Code's settings, which the software enforces no matter what Claude decides.

Plans, permissions and checkpoints#

"Letting the agent run: plans, permissions, checkpoints" was the bootcamp's third topic.

• Plans. In plan mode (opens in a new tab), Claude researches and proposes changes without editing your files, and edits stay blocked until you approve the plan. Shift+Tab cycles between modes in the terminal. Start every real-data task here.
• Permissions. As of September 2026, recent versions of Claude Code start terminal and VS Code sessions in auto mode, where a second model reviews actions in place of you. For a first project on real data, switch to Manual mode, where Claude Code starts with read-only permissions and asks before it edits files or runs commands. Check the mode indicator before you begin. Deny rules block an action in every mode.
• Checkpoints. Claude Code saves a checkpoint before each prompt you send, and /rewind (or pressing Esc twice) rolls file edits back. The limits matter: checkpoints do not track (opens in a new tab) files changed by shell commands, and they are not a replacement for version control such as Git. They only cover files Claude changes with its own editing tools, so they cannot reverse a changed database row or a sent email. That is the whole case for read-only access.

Your first live metric before lunch#

The bootcamp's target was "first live metrics on the board before the lunch bell." Here is a version you can run at your own desk:

1. Pick one export you already trust, such as last quarter's sales from QuickBooks or your job system, and copy it into a new project folder.
2. Write three weekly questions it should answer. Keep them boring: revenue this week, open quotes, jobs past due.
3. Start in plan mode and ask for a one-page dashboard that answers those three questions from that file.
4. Read the plan. Approve it only when it names the file, the calculations and where the page will run.
5. Check every number against the spreadsheet by hand. If a figure is off, fix the calculation before you add anything else.
6. Only then point it at a live source, read-only.

Picture a 30-person distributor whose office manager rebuilds a Friday report from three exports. Those exports and three questions are a complete first project.

What "wire AI into your systems" means in plain English#

The second session took the next step: connecting an agent to the systems a business actually runs, "spreadsheets, a QuickBooks-style ledger, email, a job database," using MCP and plain APIs.

MCP, the Model Context Protocol (opens in a new tab), is an open-source standard for connecting AI applications to outside systems. Its documentation compares it to a USB-C port: one standard plug instead of a custom cable for every device. Anthropic open-sourced MCP in November 2024 (opens in a new tab), and the MCP site lists support in Claude, ChatGPT, Visual Studio Code and Cursor, among others.

In plain English, an MCP server is a small adapter between the AI and one system. It offers the agent a fixed set of actions there, such as looking up an invoice or searching the job list, and the agent can only do what the adapter offers. A plain API does a similar job when no MCP server exists; it just takes more code to wire.

The session closed on "the connector recipe you'll reuse for every system at work." Here is that recipe in six steps:

1. Name the question first. "Which jobs are over budget?" tells you which system and which fields you need. "Connect everything" does not.
2. Choose the adapter. Anthropic's Claude Code security guidance (opens in a new tab) encourages writing your own MCP servers or using ones from providers you trust, and says Anthropic does not security-audit or manage any MCP server. In practice, that means the software vendor's own server or one you build.
3. Create a read-only login for the agent. The Claude Code MCP documentation (opens in a new tab) says to use a read-only database user so the queries Claude runs cannot modify data. Some vendors build this in: Intuit's open-source QuickBooks Online MCP server (opens in a new tab) has settings that switch off its create, update and delete tools, and it supports a sandbox company for testing.
4. Test on a sandbox or a copy before the real database.
5. Keep it local until it earns trust. Claude Code adds MCP servers at local scope by default, meaning only you, in one project. Share the connection with the team after it has been reviewed.
6. Write down what it can touch. One line per system: which data, read or write, whose login, who approved it.

Security review before real data#

The session page put it plainly: "Read-only first: scopes and guardrails set as you connect." Permissions expand only as trust is earned. Before an AI agent touches live business data, run this review:

• Its own login, narrow scope. A dedicated read-only account, never an owner's or an administrator's, limited to the tables, folders or mailbox the question needs.
• Trusted adapters only. Claude Code's documentation warns that MCP servers which fetch outside content can expose you to prompt injection, where hidden text tries to give the agent new orders.
• Hard blocks, not polite requests. Anything that must never happen goes in a deny rule, not just in CLAUDE.md.
• A security pass on the code. Claude Code's /security-review command checks your pending changes (opens in a new tab) for security vulnerabilities. Run it, then have a person who can read code look at the result.
• Secrets out, history in. Passwords and API keys never go into CLAUDE.md, the code or the chat. Working versions get committed to Git, because checkpoints are not version control.
• A person checks the numbers. For the first few weeks, someone compares the dashboard to the source system every week.
• Your policy covers it. Your AI use policy should say what data can go into AI tools and who approves new connections. If you do not have one, start with our free AI acceptable use policy template.

If the data is regulated, such as patient records, payment card data or customer financial information, bring in whoever owns compliance before you connect it. This is general guidance, not legal advice.

Claude Code, Cursor or Codex: which AI coding tool fits?#

Our Cursor vs Claude Code vs Codex comparison lays out the trade-offs: Cursor suits developers who want AI built into their everyday code editor, Claude Code suits teams comfortable handing an agent larger tasks and codebase-wide changes, and Codex fits organizations already using ChatGPT and OpenAI's other tools. For the management view, Claude Code for business leaders explains why a developer or technical lead still belongs in the loop.

The tool matters less than the habits: a written brief, read-only access, a plan you approve, and a person who checks the output.

Frequently asked questions#

Want it wired to your own systems?#

Rather not do the wiring yourself? Talk to us about custom AI integration. Our custom software and system integration services cover API development, custom dashboards and process automation around the software you already run, and RapidDashboard turns plain-language prompts into live dashboards and reports from your own systems.