Learning to code with AI.

Goal: Build a tiny internal tool with an AI coding assistant, review it properly, and know where your skills end. Who it is for: Business people who want to build small tools for their own team, and junior developers new to coding agents. Time: 30 minutes, plus about an hour if you complete the build.

An AI coding assistant can turn a paragraph of plain English into a working script. That moves the hard part rather than removing it: deciding exactly what to build, checking what was built, and keeping it safe.

What to learn before you let AI write code

Learn to read code, undo changes, run tests and write a spec, because you will approve every change the assistant makes.

Which assistant to learn with

The habits transfer between tools, so pick one and stay with it. Prices are as listed on the dates shown.

Our guides to Claude Code and Cursor cover plans and how each handles your code.

Exercise: build an overdue invoice report

Illustrative scenario: a small business wants a Monday list of overdue invoices from its accounting software’s CSV export. Use made-up data only.

1. Set up a safe folder. Create an empty project folder, install Python, and run git init so every change can be undone.
2. Make fake data with traps. Create invoices.csv with 12 made-up rows: day-first dates such as 03/04/2026, amounts written as “$1,250.00”, one credit note with a negative amount, one paid invoice, one blank due date and one impossible date (31/09/2026).
3. Write the spec. Adapt the one below. It states the input, output, rules and what to do when unsure.
4. Ask for a plan first. Use Plan Mode in Claude Code; in other tools, tell the assistant to restate the task, list assumptions and wait. Fix misunderstandings before any code exists.
5. Let it build, with tests. Approve file edits one at a time. Ask for a test for every rule in the spec, then run the tests yourself.
6. Check three rows by hand. Count days overdue on a calendar and compare. Look hardest at 03/04/2026: if the report treats it as 4 March, the date handling is wrong.
7. Review and commit. Read the full diff, run the review below, and commit with a message saying what the tool does and who checked it.

Build a small command-line tool in Python, standard library only.

Purpose: list overdue invoices from a CSV export so I can chase them.
Input: invoices.csv with columns Invoice, Customer, Due date, Amount, Status.
Dates are New Zealand format, day first: 03/04/2026 is 3 April 2026.
Amounts may include "$" and commas. Credit notes have negative amounts.

Output: a new file overdue.csv with Invoice, Customer, Due date, Amount
and Days overdue, most overdue first. Print the total amount overdue.

Rules:
- Never change invoices.csv.
- Skip rows where Status is Paid or Voided, and skip credit notes.
- If a date or amount cannot be read, do not guess. Write the row to
  problems.csv with the reason.
- Let me set "today" for testing, so results do not change daily.

Before writing code, restate the task, list your assumptions and
questions, and name the files you will create. Wait for my OK.
Then write the code and a test for every rule above.

In the code, look for an explicit date format like this rather than a library that guesses:

from datetime import datetime

def parse_nz_date(raw):
    """Parse a day-first NZ date such as 03/04/2026 (3 April 2026)."""
    return datetime.strptime(raw.strip(), "%d/%m/%Y").date()

assert parse_nz_date("03/04/2026").month == 4

The impossible 31/09/2026 raises an error here, which your spec turns into a row in problems.csv instead of a guess.

How to review what the assistant wrote

Run the seven-check routine in evaluating AI generated code, then add three checks specific to coding agents:

• Read the diff, not the summary. git diff shows every change; an agent’s summary of its own work can leave edits out.
• Run the tests yourself. Confirm they check results rather than just print them.
• Ask what is missing. “Which rules in my spec have no test?” regularly finds gaps.

Security basics: secrets, permissions and data

• Secrets. OWASP’s secrets management guidance notes that many organisations have keys and passwords “hardcoded within the source code in plaintext”. Keep keys in an environment file listed in .gitignore, never paste them into a prompt, and rotate any key that has been committed or shared.
• Permissions. Claude Code’s security documentation, read on 20 October 2026, says interactive sessions start in auto mode, where a separate classifier model reviews actions instead of you. While learning, switch to manual mode: it starts read-only and asks before editing files, running tests or executing commands. Use the equivalent setting in other tools, and read every command before approving it; see understanding AI tool permissions.
• Data. In manual mode Claude Code asks before reading outside the folder it started in, so anything inside that folder is open to it. Keep real customer files out; the Privacy Act 2020 applies to personal information sent to an AI provider.
• Untrusted content. Files, web pages and issue text can carry instructions aimed at the agent. See recognising prompt injection.

When to stop and bring in an engineer

Stop when a mistake would reach a customer, a ledger or someone’s personal information before a person could catch it.

Hand over your spec, fake data, tests and git history. Our custom AI development team takes prototypes like this to production.

What World of ClaudeCraft shows about building at scale

World of ClaudeCraft is Levy Street’s free, open-source (MIT) browser MMO. Its first playable version was built in about 48 hours in June 2026 with the Claude Fable 5 coding model. Since then roughly 60 contributors have pushed about 19,900 commits, and it had 2,300+ GitHub stars and 730+ forks in October 2026. The first version was the fast part; what followed was thousands of changes that had to be understood and maintained, the same shape as your invoice tool at far larger scale. Our write-up on what building games with AI taught us covers the engineering.

Free resources worth your time

• Claude Code 101: free, 12 lessons, about 1.5 hours.
• OpenAI Academy: free paths including Get Started with Codex.
• GitHub Copilot Fundamentals: a Microsoft Learn path of about 5 hours; Microsoft says Learn training is free.
• GitHub Skills: interactive courses including Introduction to Git and Getting Started with GitHub Copilot.
• CS50 Python: ten weeks of material, free through OpenCourseWare, including unit tests and file handling.

To compare paid, university and micro-credential options, see AI courses in New Zealand.

Check your work

• I can explain what each function in my tool does.
• My project is in git, and I know how to undo the last change.
• I wrote a spec, and the assistant planned before it coded.
• There is a test for every rule in my spec, and I ran them myself.
• I checked three results by hand, including a day-first date.
• No keys or passwords appear in my code, prompts or git history.
• The project folder holds only fake data.
• I know which signal would make me hand this tool to an engineer.

Next lesson: Understanding AI tool permissions.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.