Alex: If you have a mess of Python scripts, scheduled tasks, and Excel workbooks holding your monthly close together, and you want an AI that can actually read the whole folder without touching production data, this is the tool. This is AI Applied, lesson 1.
Jordan: We are talking about Claude Code. It used to be just an experimental terminal preview, but now it is a full product with a command-line interface, IDE extensions, desktop tools, and background agent routines.
Alex: Right, and the models underneath it just had a big shakeup. What is running under the hood right now?
Jordan: Sonnet 4.5 is officially phased out. The current lineup is Claude Sonnet 5.5, Opus 5.5, Haiku 5.5, and Fable 5.1. For our work, Sonnet 5.5 is the sweet spot of speed and smarts. Opus is for massive agentic deep dives, and Haiku is for cheap, high-volume tasks.
Alex: The mental model here matters. This is not just a chat window where you copy and paste snippets.
Jordan: Exactly. Think of it as tool-mediated agency. You point Claude Code at a real project directory on your machine. It reads files, reasons about the whole architecture, runs shell commands, and edits code. But the critical piece is permissions. By default, it asks before it does anything. You can also run it in plan mode, which strictly explores and drafts plans without touching a single file.
Alex: Let us walk through the exact workflow Brian should use today. He has a real-estate finance project with rent reconciliation scripts and journal-upload templates. How does he set this up safely?
Jordan: Step one, open your terminal and navigate straight into that project folder. Step two, launch Claude Code specifically using the Sonnet model and set the permission mode flag to plan. That ensures read-only behavior right out of the gate.
Alex: What does he actually tell it once it boots up?
Jordan: Frame it like an internal controls review. Tell Claude: you are reviewing this folder for close readiness and audit evidence. Map the data flow from source files to outputs. Flag where bad rent figures, missing leases, or duplicate entries could sneak in. Ask it for five test cases using fake data, what steps need logging for audit proof, and a written implementation plan. Explicitly instruct it not to edit files.
Alex: That turns it from a generic code bot into an analyst reviewing your spreadsheet pipeline. Now, what is the gotcha? What bites people here?
Jordan: Two big ones. First is billing. If you have an Anthropic API key saved in your environment variables, Claude Code might default to burning metered API credits instead of using your monthly Pro or Max subscription limits. Check your login state before you run heavy sessions.
Alex: And the second gotcha is security and permissions.
Jordan: Never use the flag that bypasses all permissions on your primary work laptop. If it runs wild, it has full system access. And never point it at folders containing live tenant payroll, banking data, or real customer identities. Work on copies with sanitized data. Claude can help draft and test your automations, but it should never be both the preparer and the approver of a live journal entry. Keep segregation of duties intact.
Alex: Quick check time. First question: Brian needs Claude Code to inspect a rent-reconciliation folder without risking accidental edits. What mode does he start in, and what should that first prompt focus on?
Jordan: Plan mode. The prompt should explicitly forbid file edits and ask for data flows, edge-case risks, fake-data test cases, and an audit trail plan.
Alex: Second check: your Claude Code terminal suddenly starts running up paid API bills instead of using your Max plan. What is the culprit?
Jordan: You almost certainly have an Anthropic API key set in your system environment variables. Clear that or check your authentication status inside the tool so it recognizes your subscription login.
Alex: Love it. Now, what is the homework?
Jordan: Take twenty minutes today. Pick one unattended Python loop from your scheduled tasks, something low-risk that does not write to production storage. Copy that script into a scratch folder alongside a dummy input file. Launch Claude Code using Sonnet in plan mode, and paste an audit prompt asking for failure modes, missing logs, and idempotency risks. While it runs, notice whether Claude naturally spots business control flaws or just tries to refactor code style. Steering it toward evidence and risk is the whole game.
Alex: Tomorrow, we look across the aisle at OpenAI to see where Codex and ChatGPT agents fit into this exact same stack. See you then.