Coding with AI, even if you are not a programmer

Start here: a formula is a set of instructions#

If you have ever left a note for someone covering your shift ("count the till, lock the back door, text me the total"), you have written a program. Code is a set of instructions written so precisely that a computer can follow them without asking what you meant.

A spreadsheet formula such as =SUM(C2:C20) is code. So is a script, a small program that does one job, such as renaming 200 photos. So is an automation, a rule that runs by itself, such as "every Monday, email me last week's total".

AI assistants are good at writing all three. Here is the one new part. When you wrote the shift note, you knew what you meant. When an AI writes the instructions, you have to find out whether they mean what you meant. That checking is what this module teaches, and you do not need to be a programmer to do it.

Our running example. Gloria runs a small home bakery that sells to five local cafes. Every Monday she updates a sales sheet: which cafe ordered, the invoice amount, when it went out and when it was paid. She wants AI help with two jobs: a column that works out late fees, and later a small script that emails payment reminders. Everything in her sheet is made up, and a pencil is all you need to follow along.

Describe the job: inputs, output, an example, edge cases#

An AI fills every gap in your request with a guess. A wrong guess produces a formula that works perfectly on a rule you never had. Four parts describe almost any formula or script:

An edge case is an unusual input at the limit of a rule: a blank cell, a zero, a value exactly on a boundary. Most mistakes in code live there, and it is the part people forget to describe. Gloria's second prompt:

I use Google Sheets. Write a formula for cell E2 that I can copy down the column.
Inputs: B2 is the invoice date, C2 is the amount in dollars, D2 is the date it was paid.
Output: the late fee in dollars, or 0 if there is no fee.
Rule: payment is due within 14 days of the invoice date. If it is paid later than that,
the fee is 2% of the amount, but never less than $5.
Example: an invoice for $400 paid 24 days after the invoice date has a fee of $8.
Edge case: paid exactly 14 days after the invoice date is on time, so no fee.
Give me the formula, then explain each part in plain words.

It names the app, because formulas differ a little between apps, and it asks for an explanation, which you will need soon. It still has a gap Gloria does not see; the next section's test finds it.

Run it, do not trust it: a worked example#

Here is the formula Gloria got back:

=IF(D2-B2>14, MAX(C2*0.02, 5), 0)

IF chooses between two answers: if the test is true, the first; otherwise the second. Spreadsheets store each date as a count of days, so D2-B2 is the number of days it took to pay. MAX gives the larger of its inputs, so MAX(C2*0.02, 5) is "2% of the amount or $5, whichever is bigger". In words: if it took more than 14 days, charge that; otherwise charge 0.

It looks right, and that is exactly the code that needs testing: nothing about it looks wrong.

A test row is a row whose right answer you already worked out by hand. You choose them on purpose: a few ordinary rows plus the edge cases. Gloria made five:

Here is what the formula gave when we ran it in a real spreadsheet:

Gloria described the missing edge case in the same chat. Unpaid invoices have a blank paid date, and they must get the fee once they are more than 14 days old. The new version:

=IF(IF(D2="", TODAY(), D2)-B2>14, MAX(C2*0.02, 5), 0)

Only one piece changed. IF(D2="", TODAY(), D2) means: if the paid date is blank, use today's date instead. She then re-ran all five rows, not just Hilltop: 0, 8, 5, 12, 0. Every row matched. Re-run every test after a fix, because a change that repairs one row can quietly break another.

Notice that the right answers came from Gloria's pencil. Answers from the AI would carry the same misunderstanding as its formula, and every row would "pass".

Read every line you can, and ask about the rest#

Testing covers the rows you tried. Reading covers the rows you did not, and it does not need fluency.

• Read every part you can: numbers, column letters, words you recognize.
• Ask about every part you cannot. "What does D2="" do?" is a good question. If a part cannot be explained in plain words, treat that as a warning.
• Explain it back. Say what you think it does in your own words, then ask whether you are right.
• Get a second opinion from something that did not write it. A fresh chat does not remember writing the code, so it has nothing to defend.

Here is a spreadsheet formula and the rule it should follow.
Formula for E2: =IF(IF(D2="", TODAY(), D2)-B2>14, MAX(C2*0.02, 5), 0)
B2 is the invoice date, C2 the amount in dollars, D2 the date paid (blank if unpaid).
Rule: due within 14 days; if late, the fee is 2% of the amount, never less than $5.
Do not rewrite it yet. List inputs that would make it give a wrong or confusing answer,
and for each one give me a test row I can check by hand.

Four ways AI code goes wrong#

Gloria's formula failed in a classic way. Three others are worth knowing by name.

Researchers have measured two of these: security holes and invented names. In one study, an AI coding tool completed 89 small tasks chosen because they are easy to get wrong in ways an attacker could use. About 40% of the 1,689 programs it produced had such a weakness. A second study gave 47 people security-related coding tasks, some with an AI assistant and some without. Those with the assistant wrote less secure code overall and were more likely to believe it was secure. On one task, 3% of people with the assistant wrote a secure answer, against 21% without. These were small groups: 33 people with the assistant and 14 without. The danger is not only the code; it is the confidence.

A third study generated 576,000 code samples from 16 AI models and checked every package name the models suggested. On average, at least 5.2% of package names from commercial models did not exist. For open-source models, which anyone can download and run, it was at least 21.7%. Anyone can publish a package under a name AIs tend to invent. Whoever installs it after that is running a stranger's code on their own computer.

Keep passwords and keys out of prompts and code#

Gloria's second job is a script that emails each cafe a polite reminder when an invoice is overdue. The AI's draft looks complete, and near the top is a line like this:

password = "type-your-email-password-here"

A secret is anything that lets someone act as you, such as a password or an API key. An API key is a long code that lets a program use an online service on your account. Inside the code, a secret travels wherever the code goes: chat history, a shared file, a screenshot, a public code site.

• Never paste a real password or key into a chat. Replace it with a placeholder such as PASSWORD_HERE first. The AI does not need the real one to help.
• Keep secrets out of the code. Ask: "Change this so the password is read from an environment variable." An environment variable is a named value stored on your computer, outside the file. In the Python language the line becomes:

import os
password = os.environ["BAKERY_MAIL_PASSWORD"]

• If a secret was pasted or shared, change it now. Deleting the message does not undo copies that may already exist.
• Use made-up data. Gloria tests with invented cafes and her own email address, not her customer list. Rows with the same shape test the code just as well.

Anything you paste into a chat leaves your computer and may be stored. Check the service's privacy policy, and at work, your employer's policy on what you may share.

If you are learning to code: write it first#

Gloria does not want to become a programmer, so letting the AI write her formula was the right call. If your goal is to learn to code, the trade is different: when the AI writes the code, it also does the part that teaches you.

1. Write your own attempt first, even if it is wrong or unfinished.
2. Run it and read the error message.
3. Ask for a hint, not the answer: Here is my code and the error I get. Do not give me corrected code. Ask me one question that points me toward the bug.
4. Ask for a review once it works: "What would a careful reviewer change, and why?"

Module 5 covers learning with AI in full. Before you use AI on graded coding work, check your school's or course's policy; the rules differ from class to class.

When the AI can edit files or run commands#

Some AI tools can open your files, change them, or run commands. A command is an instruction typed to the computer itself, such as "delete this folder". A tool that acts for you like this is often called an agent. Module 4 covers agents in full. For coding work, four rules do most of the protecting.

Module 4 tells our worst moment with an agent: a helper our AI named for "read" that the computer took as "delete", and four record files gone.

What was new in this module#

1. Formulas, scripts and automations are all code.
2. Describe the job: inputs, output, the rule with an example, edge cases.
3. Test on rows whose answers you worked out yourself: a blank, a boundary, something odd.
4. A blank cell counts as zero in spreadsheet arithmetic.
5. After any fix, re-run every test.
6. Read what you can, ask about the rest, and ask a fresh chat what would break it.
7. Four failures: wrong at the edges, invented names, out-of-date ways, security holes.
8. Secrets never go into prompts, shared code or screenshots. If one leaks, change it.
9. Learning to code: run your own attempt and read the error before you ask for a hint.
10. When a tool reports "none found", check the real file before you let it act.

Key terms#

Code

Instructions precise enough for a computer to follow. A spreadsheet formula counts.

Edge case

An unusual input at the limit of a rule: a blank cell, a zero, a value exactly on a boundary.

Test row

An input whose right answer you worked out yourself.

Package

A bundle of ready-made code that someone publishes for others to install.

Secret

Anything that lets someone act as you, such as a password or an API key.

API key

A long code that lets a program use an online service on your account.

Environment variable

A named value stored on your computer, outside your code, so a secret is not written in the file.

Agent

An AI tool that takes actions for you, such as editing files or running commands.

Least permission

Giving a tool access to only what its job needs.

Check yourself#

Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.

I can#

Tick each one only when you could do it tomorrow without this page open.

• I can describe a formula or script job in four parts.
• I can build test rows, including a blank and a boundary, and work out their answers by hand.
• I can find which row a formula gets wrong and explain why in plain words.
• I can name four common ways AI code goes wrong and how to catch each one.
• I can keep passwords and API keys out of prompts, shared code and screenshots.
• I can set up a backup and least permission before an AI tool edits my files.

Frequently asked questions#

Sources#

Pearce, H., Ahmad, B., Tan, B., Dolan-Gavitt, B. and Karri, R. (2022). Asleep at the Keyboard? Assessing the Security of GitHub Copilot’s Code Contributions. IEEE Symposium on Security and Privacy 2022 (preprint arXiv:2108.09293). 89 scenarios, 1,689 programs, about 40% vulnerable.

Perry, N., Srivastava, M., Kumar, D. and Boneh, D. (2023). Do Users Write More Insecure Code with AI Assistants? ACM Conference on Computer and Communications Security (CCS 2023), doi:10.1145/3576915.3623157. 47 participants; message-signing task: 3% secure with an AI assistant, 21% without.

Spracklen, J., Wijewickrama, R., Sakib, A. H. M. N., Maiti, A., Viswanath, B. and Jadliwala, M. (2025). We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs. USENIX Security Symposium 2025 (preprint arXiv:2406.10279). 576,000 samples, 16 models; at least 5.2% (commercial) and 21.7% (open-source) of suggested packages did not exist.

Microsoft Support. IF function (Excel). Checked October 2026.

Microsoft Support. TODAY function and MAX function (Excel). Checked October 2026.

Google. IF function, MAX function and TODAY function, Google Docs Editors Help (Google Sheets). Checked October 2026.