The beginner's guide to Codex for business users
You have a messy export and a job a small script could do, but you're not a developer. Codex can write the script; your job is to check it didn't quietly lose a row on the way.
What does your job involve?
Six jobs where a coding agent helps someone who isn't a developer. Each card says where to start, what access it needs, a prompt to use and the check that matters.
Open a task to see where to start, the access to check, a prompt to use and what to check in the result.
Build a small business calculator Codex → practice project
Make a small reusable tool with visible limits.
- Where to start
- In the ChatGPT desktop app, open an empty practice folder in Codex and ask for a plan before any files are written.
- Access to check
- Codex edits inside the folder you opened; deployment is a separate step.
- Bring this
- A checked formula, units, input bounds and known-answer cases.
Try this in Codex
Plan an offline calculator for [rule] in this practice folder. Show inputs, units and invalid-value handling. After I agree the plan, create the files and run checks for [known cases]. No live data, external services or deployment. Report output paths and the results of each check.Why this request helps: A checked rule and boundary cases give you a way to inspect generated software without judging it by appearance.
Check before using it: Open the tool, try normal/zero/missing/invalid values and compare with independent arithmetic.
If this route is missing: Prepare a formula specification for a technical colleague.
Create a CSV helper Codex → source files
Automate a repeated file operation you can verify.
- Where to start
- Put a practice copy of the CSV in an empty folder and open that folder in Codex.
- Access to check
- Edits inside the practice folder only; say no to network access.
- Bring this
- CSV, field definitions, matching rule and expected rows.
Try this in Codex
Build a local CSV helper for [job] using only the CSV in this folder. Keep the original unchanged, state your assumptions about each field, flag missing or duplicate keys and write a separate output. Count rows read, written and flagged. Report which files changed. No upload or connection to business systems.Why this request helps: Separate outputs and known-answer rows make a transformation easier to audit.
Check before using it: Compare row counts, headers and a few known records; read the flagged rows as well as the output.
If this route is missing: Ask for a step-by-step spreadsheet method instead.
Prototype a page or small dashboard Codex → local prototype
Make a reviewable example before committing to a full system.
- Where to start
- Open an empty practice folder in Codex and describe one small page and the question it answers.
- Access to check
- Local project/tools; hosting and production data are separate.
- Bring this
- Audience, sample data, layout and the real user question.
Try this in Codex
Create a local prototype for [purpose] using these illustrative records. Make the question and next action clear, with readable mobile layout and keyboard controls. Do not add authentication, analytics, external calls or live transactions. Explain what is mocked and how to preview it.Why this request helps: A bounded prototype lets you review the experience before assuming the missing production systems exist.
Check before using it: Open the page on a phone and with a keyboard; check data definitions and what it shows when there is no data.
If this route is missing: Ask for a static wireframe or implementation brief.
Understand proposed software changes Codex → project inspection
Ask better questions about work your technical team proposes.
- Where to start
- Open the project folder in Codex and ask for an explanation before any edits.
- Access to check
- Read access to that project; reading it gives no permission to change it.
- Bring this
- Problem, relevant project and business question.
Try this in Codex
Read this project and explain how [behaviour] currently works. Cite file paths and distinguish observed code, inference and unanswered questions. Do not edit files yet. Propose a small change and explain acceptance checks and risks for a business reviewer.Why this request helps: Read-first evidence separates a grounded plan from a confident description of an unseen system.
Check before using it: Open the cited files and confirm scope with the technical owner.
If this route is missing: Give a chat the written specification and ask for review questions.
Review the files an agent changed Codex → changed-file review
Know what changed and what completion actually means.
- Where to start
- After a small task, open the list of changed files and the differences Codex shows for each.
- Access to check
- Project/version access and compatible viewer for the output.
- Bring this
- Original requirement, changed-file list and stated check results.
Try this in Codex
Explain each changed file against the requested outcome. Show the checks you ran and any failures or checks not run. Identify unrelated changes, remaining assumptions and what I need to check by hand. Do not describe a local preview as a published release.Why this request helps: The requirement gives the diff a purpose; actual checks distinguish evidence from narration.
Check before using it: Read the differences and run the output yourself before accepting it.
If this route is missing: Ask the technical owner for the list of changes and a short note of the tests run.
Make a repeatable task specification Codex → project instructions / skill
Reuse a checked process with explicit inputs.
- Where to start
- Start from a method that has already worked, and test it with fresh practice inputs.
- Access to check
- Project instructions or skills in your setup; a saved instruction adds no new access.
- Bring this
- Stable method, variable inputs and known acceptance cases.
Try this in Codex
Prepare reusable instructions for [small job]. Separate the stable rule from changing facts and name the required inputs. Include how to handle missing or conflicting inputs, and a review checklist. Do not include private records or say these instructions have been tested.Why this request helps: A reusable method should preserve the rule rather than yesterday’s conclusion.
Check before using it: Run the normal and missing-input examples yourself and read the results.
If this route is missing: Save the instruction in a shared document and paste it in each time.
Know what you have
Codex: the feature map
Use this map to distinguish a reply, an editable output and an action in another app. Availability depends on the signed-in account, enabled features and the specific operation.
| Feature | Useful for | Where | Access / limit |
|---|---|---|---|
| Desktop project | Review files and small outputs | ChatGPT desktop app › open a folder | The folder you opened and the permissions you agree |
| CLI / IDE | Developer workflows | Terminal or editor extension | Technical setup and command controls |
| Files and local tools | CSV helpers, calculators and prototypes | The practice folder | Practice copies and a viewer for the output |
| Connected tools | Supported service context or actions | Plugins and tools | Sign-in and what each action can change |
| Publication | Move a checked change to real users | Separate repository or hosting step | Local output is not a release |
Find out what your account can use
- Download the ChatGPT desktop app, sign in with your work account and open Codex on a practice folder, or use the setup your technical team gives you.
- Check the folder Codex shows; naming a folder in a prompt doesn't enforce permissions.
- Keep a copy of any source file, and say no to network or wider file access a small job doesn't need.
- Ask for a plan first and read the changed files afterwards.
If your organisation has not approved project access, ask an ordinary chat to prepare a specification and acceptance examples instead.
Ask your administrator:
Which ChatGPT account should I use for Codex at work? Please confirm whether Codex is enabled for our workspace, which permission settings we use, which information I may put in a practice folder, and who agrees wider access. I want to try [specific task] on a practice folder first.Official sources
What you'll make
a small mileage summary built by Codex from a practice CSV, with a row count that proves nothing was dropped.
What you'll need
a ChatGPT account (Codex is included from the Free plan up, with different usage limits), the ChatGPT desktop app or the setup your technical team gives you, and a new empty folder. About twenty minutes the first time.
Example result.
| Van | Valid rows | Miles |
|---|---|---|
| VAN-1 | 3 | 125 |
| VAN-2 | 2 | 126 |
| Flagged | 1 | 6 Oct, VAN-2: “fifty-one” is not a number |
Rows read: 6. Rows summarised: 5. Rows flagged: 1.
1. Get Codex open
Download the ChatGPT desktop app for macOS, Windows or Linux, open it and sign in with your work ChatGPT account. Codex lives inside it: choose where to work by opening a folder. If your organisation uses the command line, an editor extension or a cloud setup instead, use that and follow the rest of this guide unchanged.
Make a new folder called codex-practice and put nothing else in it. Save the fictional mileage.csv from the worked example there, then open the folder in Codex.
Why this works: an empty folder with one practice file means anything Codex changes is easy to see, and nothing real is in reach.
What you should see: Codex showing codex-practice as the folder it’s working in, with mileage.csv inside.
2. Ask for a plan before it changes anything
In this folder, plan a small script that reads mileage.csv and
writes summary.csv with total miles per van.
Do not change mileage.csv.
Count the rows read, summarised and flagged, and list any row
whose miles value is not a whole number, without guessing a value.
No network access and no files outside this folder.
Show me the plan and the checks you will run before writing anything.Why this works: a plan is cheap to correct. Asking for row counts and flagged rows up front gives the script a way to admit what it couldn’t read.
What you should see: a short plan naming the files it will create, how it handles a bad value and the checks it will run. You don’t need to understand the code to judge whether that’s the job you asked for.
3. Let it make the change inside the folder
If the plan matches, tell Codex to go ahead. Approve edits inside codex-practice only. If it asks for network access, wider file access or to install something, say no and ask why; a small local job shouldn’t need any of them. Naming a folder in a prompt doesn’t enforce anything, so check the folder Codex shows rather than trusting the sentence.
Why this works: each permission is a separate decision. Saying no to the ones a practice job doesn’t need keeps a mistake small.
What you should see: a list of the files it created or changed, and the checks it says it ran.
4. Open the output and count the rows
Open summary.csv in Excel or Sheets and mileage.csv beside it. Count the rows in the source, then check the summary accounts for every one — summarised or flagged. Then change one number in a copy of the source, ask Codex to run the script again, and see whether the total moves.
Why this works: a confident “done” message is not a check. The row count catches a silent drop; changing a value catches a total that was typed in rather than calculated.
What you should see: rows read = rows summarised + rows flagged, and a total that changes when the input does.
Check the result
mileage.csvis unchanged.- Rows read equals rows summarised plus rows flagged.
- Each van’s total matches your own sum of its valid rows.
- Changing one value changes the total.
The usual failure: the script skips a row it can’t read and still reports success. The totals look tidy, so the only sign is a row count that doesn’t add up, or no row count at all.
The summary does not account for every row in mileage.csv.
Count rows read, summarised and flagged, and list each flagged row
with its date, van and the value you could not read.
Do not convert or guess any value. Calculate totals from the file,
then test it again with one changed value and report what changed.Worked example: the van that lost a day
Fictional example. The practice file:
date,van,miles
2026-10-05,VAN-1,42
2026-10-05,VAN-2,67
2026-10-06,VAN-1,38
2026-10-06,VAN-2,fifty-one
2026-10-07,VAN-1,45
2026-10-07,VAN-2,59Codex’s first run reported “Processed mileage.csv: VAN-1 125 miles, VAN-2 126 miles. All checks passed.”
- The totals were right and the job was wrong. 42 + 38 + 45 = 125 and 67 + 59 = 126, but VAN-2 drove on 6 October too. The script dropped “fifty-one” without a word, and the report had no row count.
- The repair made the gap visible. After the repair prompt, the summary said 6 rows read, 5 summarised, 1 flagged, matching the table at the top of this page. Codex didn’t turn “fifty-one” into 51; whoever keeps the log decides that.
- The changed-value test caught nothing new, which is the point. Changing 59 to 60 in a copy moved VAN-2 to 127, so the total is calculated, not typed in.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Codex or Claude Code?
Both are coding agents that can read a project, edit files and run commands. If your organisation already has one, start there. If you’re choosing, compare them on the same small job with the same checks: readable changes, honest row counts, and a clean way to undo. The questions that matter — where it runs, what it may change and who pays for long tasks — are set out in Choose ChatGPT or Claude for your business.
From a local file to a live website
Local output, a repository commit and a live website are different milestones. Codex finishing a file on your laptop means the file exists; it doesn’t mean anyone else can see it, or that it’s safe to publish. If a job ends at a prototype, report it as a prototype with its limits. Moving it into a shared repository or onto a website is a separate, deliberate step, usually for someone who can review the change and undo it.
Next jobs, and what saved instructions don't do
Once the practice job works, try a CSV helper, a business calculator or reviewing the files Codex changed. Project instructions or skills can save a method you like, such as “always report row counts”. They change how Codex works, not what it can reach, and they don’t prove the method was tested on today’s file.