How Voice Changed My Life: Add Codex Without Giving Up Claude

I already valued working with Claude. Codex Voice added something different.

I could talk while I was working. I could ask what was happening, move between tasks and ask for a simple explanation without stopping to write a formal prompt. It felt less like operating software and more like having a calm colleague beside me.

This guide is for people who already use Claude and want to add Codex Voice. It is not a Claude-versus-Codex contest. Keep using Claude for the work you value. Add Codex when voice, local navigation and task coordination help.

Who this is for

This guide is for managers, founders and organisational leaders who work with information but are not technical specialists.

You already use Claude or Claude Code. You are new to Codex. You want to speak to your work, find your way around local folders and keep several tasks in view.

You do not need to know Git commands, terminals, branches or worktrees. Codex can check the technical state and explain it in ordinary language.

Before you start

You need:

your existing Claude setup;
a supported ChatGPT plan and workspace with Codex Voice available;
the ChatGPT desktop app;
permission to use one harmless local folder;
a folder with no passwords, customer records or sensitive personal data;
and about 30 minutes.

OpenAI says availability depends on plan, rollout and workspace settings. Its current desktop guide covers macOS, Windows and Linux. Check the official page rather than assuming every account has the same controls.

What you will learn

By the end, you should be able to:

install and open the ChatGPT desktop app;
open a local folder in Codex;
start a Codex Voice task safely;
tell Codex which company, account, project and task you mean;
continue the right task or deliberately create a separate one;
show Codex the visible Claude task safely when that helps;
ask Codex to find, explain and coordinate local work;
keep Claude doing the work you already value;
prevent both tools changing the same thing at the same time;
leave a plain-language handoff when you switch tools.

The simple idea

Claude and Codex can both work with the same folders and information.

That is useful. It also creates one practical risk: both tools may change the same file or task at the same time.

The rule is not "never share a folder". The rule is:

Do not let both tools change the same thing at the same time without a clear handoff.

1You choose one outcome

2Codex checks the folder and unfinished work

3Codex or Claude owns the next change

4The active tool explains and verifies its work

5Plain-language handoff

6The other tool may continue

One tool owns the active change, verifies it and leaves a plain-language handoff for the other.

The editable source for this diagram is Mermaid plain text. See why that works for people and agents.

Step 1: Decide what you are adding

Do not start by moving all your Claude work.

Choose one useful addition. For example:

speak to Codex while you review a project;
ask Codex to explain a folder or recent change;
ask what is blocked across several tasks;
let Codex prepare a handoff for Claude;
or use Codex to make one small, approved update.

Write down what Claude should continue to own. That may be research, writing, analysis, coding or a familiar project workflow.

Step 2: Install and open Codex

Use OpenAI's official ChatGPT desktop download page.

Download the app for your computer.
Open it and sign in with the correct ChatGPT account.
Choose Codex.
Open the harmless folder you selected.
Choose Ask for approval for your first exercise.

OpenAI describes Ask for approval as the normal starting point. Codex can work inside the chosen folder and pauses before reaching outside that boundary.

Do not choose full access for this exercise. More access does not make a first lesson better.

Step 3: Ask Codex to check the room

Before Voice starts, paste this:

Codex may not know the company or account from the folder alone. That is useful. Tell it the missing information. Do not let it guess.

Step 4: Start a new Voice task

OpenAI's current instructions say a task must begin in Voice mode if you want a Voice conversation, rather than ordinary dictation.

Open a new, empty Codex task.
Select Start new voice chat before sending a message.
Allow microphone access if you are comfortable doing so.
Choose a voice.
Keep the visible conversation open.

If you use screen context, inspect the front window first. OpenAI warns that an app snapshot may include accessible text outside the visible part of the window. Do not show it a screen containing information you would not deliberately share.

Step 5: Name the project and task

Speak naturally. You do not need to supply an internal ID, computer path, branch name or Git command.

Say:

In this guide I use task because that is how the Codex app describes a work conversation. You may also hear people call it a thread or chat. The important thing is its clear human-readable name.

A project keeps related files and context together. Separate tasks inside that project keep distinct outcomes understandable. For example, one task can prepare a client brief while another checks the source documents. They can use the same approved project files without mixing every conversation into one transcript.

Ask Codex to resolve the names for you. If two projects or tasks have similar names, it should explain the ambiguity and ask which one you mean before acting. It should not ask a beginner for hidden IDs or technical names unless there is no safer way to distinguish them.

Show Codex the Claude task safely

Sometimes the easiest handoff is visual. Attach a cropped screenshot of the relevant Claude task. On macOS, where ChatGPT Voice screen context is available and allowed by your workspace, you can instead bring that task to the front and ask Codex to look at the visible window.

OpenAI says screen context can take an appshot of the frontmost window. It may include accessible text outside the visible scroll area, and an organisation may disable it. Treat it as optional. If it is unavailable, attach a screenshot or paste a short handoff.

Before any edit, Codex should:

summarise what it can actually see;
identify the likely project, task and unfinished work;
state what it cannot confirm from the image;
ask you to confirm the project and task;
inspect the approved local files before treating the screenshot as current.

A screenshot does not reveal hidden files, the full conversation history or work that happened after it was taken. Crop or hide unrelated conversations, credentials, personal data and client information before sharing it.

Step 6: Say which room you are in

Begin naturally:

Then ask:

The six useful facts are:

machine;
company;
account;
task;
folder or project;
current line of work.

You do not need to use the technical names. Codex can perform the checks and translate the result.

Step 7: Use Voice to navigate and understand

Now ask ordinary questions:

What is this project trying to achieve?
Which files or documents matter most?
What changed most recently?
Is anything blocked or unfinished?
Explain that as if I am new to this area.
What decision do you need from me?
What should Claude know before it continues?

Voice is particularly good when you know the business outcome but not the technical language. You can keep asking until the explanation is useful.

Step 8: Give one tool ownership

Before anything changes, say which tool owns the next piece of work.

For example:

Or:

Both tools may read the same folder. Only one should own the active change.

When not to run both

Stop one tool before using the other when:

both may edit the same file, document or task;
you do not know whether Claude has unfinished work;
changes exist that nobody can explain;
the company, account or folder is unclear;
the folder contains secrets or sensitive data that are not approved;
the work could publish, deploy, delete, spend money or contact people;
or you cannot tell which tool currently owns the outcome.

If you are unsure, ask both tools to review without editing. Then choose one.

Step 9: Ask for a handoff

When Codex finishes, ask:

Paste that handoff into Claude and say:

What Codex checks for you (optional)

Behind the plain-language check, Codex may inspect:

the exact project folder;
whether files have unfinished changes;
which saved line of work is active;
where the shared copy of the project lives;
and the exact differences since the last saved point.

Developers may call these the repository, working tree, branch, remote and diff. You do not need to operate them yourself. You do need Codex to explain whether the work is clean, unfinished or likely to collide.

Hands-on exercise: one explanation and one handoff

Use a harmless folder you already use with Claude.

Make sure Claude is not currently changing it.
Install and open the ChatGPT desktop app.
Open the folder in Codex with Ask for approval.
Paste the workspace-check prompt.
Name the existing project and task you want Codex to use.
Ask Codex to confirm the project, task and existing work.
Start or resume the Voice conversation.
State the company, account, project, task and folder.
Ask Codex to explain the project without editing.
Ask for a handoff to Claude.
Give the handoff to Claude and ask it to verify without editing.

The exercise is complete when both tools describe the same project and neither has changed a file.

Verification checklist

[ ] You signed into the intended ChatGPT account.
[ ] You opened the intended company folder.
[ ] Codex checked for unfinished work before editing.
[ ] Voice started in a new task.
[ ] You stated company, account, project and task.
[ ] Codex confirmed the named project and task before acting.
[ ] A separate task was created only for a genuinely separate outcome.
[ ] Any screenshot or screen context was cropped, confirmed and treated as partial context.
[ ] Only one tool owned the active change.
[ ] Sensitive screens and files stayed out of scope.
[ ] Codex explained what it did in ordinary language.
[ ] Codex verified its result.
[ ] The handoff told Claude what changed and what remained unfinished.
[ ] No authority to publish, deploy, delete, spend or contact was implied.

Common mistakes and safe recovery

Mistake	Why it matters	Safe recovery
Opening the wrong company folder	The conversation starts with the wrong information	End Voice, close the folder and reopen the correct company project
Letting both tools edit	One change can overwrite or confuse another	Ask both to stop, preserve the current work and explain what each changed
Assuming Codex knows the company	A folder name is not reliable authority	State company, account, project and task at the start
Continuing the wrong task	Related work gets mixed into the wrong conversation	Ask Codex to confirm the named project, task and visible unfinished work before acting
Treating a screenshot as the full record	Hidden files or later work may contradict the image	Confirm what the image shows, state uncertainty and inspect the approved project files
Starting with full access	A small mistake can reach much further	Return to Ask for approval and use a harmless folder
Showing a sensitive screen	Screen context may include more text than is visible	End Voice, turn off screen context and review what was shared
Switching tools without a handoff	The second tool works from an incomplete story	Stop edits and ask the first tool for the plain-language handoff

Do not fix a collision by deleting unexplained work. Ask Codex or Claude to preserve it, explain it and propose a safe recovery.

Ready-to-copy Codex prompts

Start with the right project and task

Paste this

Please do this in the [task name] task inside the [project name] project. Before you start, confirm that you are in the right project and task, and tell me what existing work you can see. Work only in the approved local folder for [company and project]. I already use Claude here. Resolve the named project and task for me. If the name is ambiguous, ask me which one I mean. Do not ask me for IDs, paths, branch names or Git commands unless there is no safer way to distinguish the work. Before changing anything, inspect and preserve all existing work, including any unfinished changes made by Claude. Do not overwrite, delete or hide anything. Tell me if Codex and Claude would collide on the same file or task. Keep this work local and private. Do not publish, upload, share, deploy, spend money or contact anyone. Explain in ordinary language what you plan to change. Wait for my approval before editing. After the approved change, verify the result and leave a clear handoff for Claude covering what changed, what you checked, what remains unfinished and the safest next step.

Continue an existing task

Create a separate task only when you mean to

Ask for a status update

Ask for a plain-language handoff

Show the Claude task visually

Do this now

Choose one harmless folder you already use with Claude.
Open it in Codex with Ask for approval.
Paste the project-and-task prompt.
Ask Codex to confirm the named project and task.
Ask Codex to explain the folder without changing it.
Ask for a handoff to Claude.

That is enough for your first session.

Your next step

Repeat the exercise with one small, reversible change. Keep Claude for the work you value. Add Codex where Voice and coordination make your working day easier.

Then continue with: