This guide is for researchers and students who use the ChatGPT app and have never opened a terminal. If you have no preference between the two assistants, take the Claude Code guide: it has been checked more closely.

It has three parts. The basics (steps 1 and 2): a terminal, Python, and Codex itself. ai-architecture (steps 3 to 5): the main setup, a set of planning and continuity skills for Codex. Provenote (steps 6 to 9): an optional chapter for research work.

The nine steps: 1 Before you start · 2 Install Codex · 3 Get ai-architecture · 4 Add its skills to a project · 5 Check that Codex sees them · 6 Get Provenote · 7 Add the Codex adapter · 8 Check 1: discovery · 9 Check 2: the helper.

Part 1 · The basics

Step 1: Before you start

A terminal is a window where you give your computer instructions by typing short lines called commands. You type or paste one line at a time and press Return (Mac) or Enter (Windows and Linux). To paste: Cmd + V on a Mac; Ctrl + V or a right-click in Windows PowerShell; Ctrl + Shift + V on Linux. To open the terminal:

I have a Mac

Press Command + Space, type Terminal, press Return. A window opens (white, or black in dark mode). Its last line ends with % and shows your own name.

Open the terminal on a MacPress Command and Space, type Terminal, press Return. A white window opens with a line that ends in a percent sign.Open the terminal on a Mac1Press Command + SpaceCommand+Space2Type Terminal, press ReturnTerminal3A small window opensyou@Mac ~ %Yours shows your own name.The line ends with %Illustration, not a screenshot

I have a Windows PC

Click Start, type PowerShell, press Enter. A dark window opens. Its last line starts with PS, ends with > and shows your own name.

Open the terminal on WindowsClick Start, type PowerShell, press Enter. A dark window opens with a line that starts with PS and ends in a greater-than sign.Open the terminal on Windows1Click the Start buttonThe Start button, at the bottom2Type PowerShell, press EnterPowerShell3A small window opensPS C:\Users\you>Yours shows your own name.The line ends with >Illustration, not a screenshot

I have a Linux PC

Press Ctrl + Alt + T, or search for Terminal in your application menu. A terminal window opens.

Open the terminal on LinuxPress Control, Alt and T, or search for Terminal in your application menu. A terminal window opens with a line where you type.Open the terminal on Linux1Press Ctrl + Alt + TCtrl+Alt+T2Or search for TerminalTerminal3A small window opensyou@computerYours shows your own name.A cursor blinks where you typeIllustration, not a screenshot

Now check what you need:

  • An account that can use Codex, and its cost. Plans and prices change, so this guide quotes none. Read the official page first: https://developers.openai.com/codex/cli
  • Python 3.11 or newer for the helper programs. Python is a free tool they run on; ai-architecture's continuity skill and Provenote's helpers use it. In the terminal window, run python3 --version (macOS or Linux) or py -3 --version (Windows; if that is not recognised, try python --version). Remember which one worked; step 9 uses it. It should print Python 3.11 or newer. If not, install it. On a Mac, use the macOS page, then close and reopen the terminal window. On Windows, python.org now leads with the Python install manager: download it from the Windows page, open the file and click Install. Then open a new PowerShell window and run py install 3.11. Python may prompt you to add a folder to your PATH (the list of places the terminal looks for programs); the Python documentation calls this optional. Check with python --version. If python is not found, open Windows Settings, search for Manage app execution aliases and check the Python entries there.
  • Linux (if Python is missing or too old): install Python 3 with your package manager: sudo apt update && sudo apt install python3 (Debian or Ubuntu), sudo dnf install python3 (Fedora, which already has Python 3) or sudo pacman -S python (Arch). sudo runs a line with administrator rights, so it asks for your own computer password; while you type it, usually nothing appears on the screen. Type it and press Enter. If a line asks whether to continue, type y and press Enter. On Ubuntu 22.04, apt gives Python 3.10, which is below 3.11 and cannot reach 3.11 this way. If python3 --version still shows 3.10 after the install line, you can still do the copy steps in Part 2, which do not need Python; the Python-based helpers (step 9 and the installer in 7c) may not run.
  • A project folder for Codex to work in. Step 4 tells you which; if you skip Part 2, step 7 uses the example folder instead.

Done when the terminal is open and Python prints 3.11 or newer. (On Ubuntu 22.04 that is not possible; see the Linux bullet above for what you can still do.)

Stuck? See Stuck? Ask for help.

Step 2: Install Codex

Codex runs in the terminal. The official page is the authority: https://developers.openai.com/codex/cli. If the line below differs from that page, follow the official page.

The install lines below download OpenAI's installer and run it. Only use lines like this from an official page.

I have a Mac

  1. Open Terminal (step 1).
  2. Paste this line and press Return. Wait until you can type again.

From the official Codex CLI page (macOS/Linux, standalone installer): https://developers.openai.com/codex/cli

In the terminal
curl -fsSL https://chatgpt.com/codex/install.sh | sh
  1. If the installer's last lines mention PATH and show a line to run, copy that exact line, paste it, and press Return. Then close the Terminal window completely, open a new one, and check. This check comes from the Codex CLI source (https://github.com/openai/codex), not the official page, and prints a line like codex-cli 0.159.2.
In the terminal
codex --version

Done when a version number is printed.

Stuck? "command not found" or an error. Do not run the installer again and again. Copy the exact text and see Stuck? Ask for help.

I have a Windows PC

Paste this line into PowerShell and press Enter. It is the Windows line from the Codex README (https://github.com/openai/codex); the Windows tab of the official page should show the same command.

In the terminal
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

If the official page asks you to install Node.js or npm first, stop and ask for help.

Then close every PowerShell window, open a new one, and check. This check comes from the Codex CLI source (https://github.com/openai/codex), not the official page, and prints a line like codex-cli 0.159.2.

In the terminal
codex --version

Done when a version number is printed.

Stuck? "not recognized" or an error. Copy the exact text and see Stuck? Ask for help.

Option: WSL (click or tap to open)

OpenAI's page says to use the native Windows sandbox by default, so the route above is the normal one. WSL (a small Linux system inside Windows) is an option when you want Linux-native tooling for Codex itself. It needs WSL version 2 (which is what wsl --install sets up); Codex no longer supports version 1. Setting it up: open PowerShell as administrator (right-click Windows PowerShell, Run as administrator; if Windows asks "Do you want to allow this app to make changes?", choose Yes), run this line, restart, open Ubuntu from the Start menu and create a user name and password. You can close the administrator window after the restart; you only needed it for wsl --install. Windows 10 version 2004 or later, or Windows 11, is required.

In the terminal
wsl --install

Then, in the Ubuntu window, paste this line (OpenAI's macOS/Linux line) and press Enter. Wait until you can type again.

In the terminal
curl -fsSL https://chatgpt.com/codex/install.sh | sh

If you see curl: command not found, run sudo apt install curl and then paste the line again. Check with codex --version. OpenAI says Windows folders such as /mnt/c can be slower, and advises keeping your project folders under your Linux home folder (for example ~/code/my-app).

Steps 3 to 9 of this guide are not written for WSL. They copy folders in File Explorer and drag folders into the terminal, and this guide has not been tested with WSL. So with WSL you can stop after this step, or you can do steps 3 to 9 on native Windows, using the route above. You can always ask for help.

I have a Linux PC

  1. Open the terminal (step 1).
  2. Paste this line and press Enter. Wait until you can type again.

From the official Codex CLI page (macOS/Linux, standalone installer): https://developers.openai.com/codex/cli

In the terminal
curl -fsSL https://chatgpt.com/codex/install.sh | sh

On Debian or Ubuntu, if you see curl: command not found, run sudo apt install curl and then paste the line again.

  1. If the installer's last lines mention PATH and show a line to run, copy that exact line, paste it, and press Enter. Then close the terminal window completely, open a new one, and check. This check comes from the Codex CLI source (https://github.com/openai/codex), not the official page, and prints a line like codex-cli 0.159.2.
In the terminal
codex --version

Done when a version number is printed.

Stuck? "command not found" or an error. Do not run the installer again and again. Copy the exact text and see Stuck? Ask for help.

You sign in to Codex later, in step 5, once the project folder exists (or in step 8, if you skip Part 2).

Part 2 · ai-architecture

ai-architecture is a collection of skills kept on GitHub. Its planning skill turns a bigger job into a written plan, and its continuity skill helps the next session pick up where the last one stopped. You copy two folders into a project. A skill is a set of instructions the assistant follows for one kind of job.

  • Written for: ai-architecture v0.4.2-preview, read but not run by this guide's author (on GitHub). Its own page says it was built and used on one Linux machine and that other operating systems are untested, so this part is untested as a fresh install on any system. Whether your Codex finds the copied skills by itself is also untested.
  • Codex gets these skills only. There are no /ai-architecture: commands and no blocking hooks in Codex.

Step 3: Get ai-architecture

3a. Download it. Open the v0.4.2-preview page linked above in your web browser. Click the green Code button, then Download ZIP. You do not need Git for this route. A ZIP file is a folder packed into one file; unzipping unpacks it. It usually lands in your Downloads folder.

Done when a ZIP file from ai-architecture is in your Downloads folder.

3b. Unzip it. Your file window is called Finder on a Mac and File Explorer on Windows. On Linux, use your file manager.

I have a Mac

Open Finder, click Downloads, double-click the ZIP file. (Some browsers have already unzipped it.)

I have a Windows PC

Open File Explorer, click Downloads. Do not just double-click the ZIP: that only lets you peek inside. Right-click the ZIP file, choose Extract All…, then click Extract. Windows may have made a folder inside another one: open them until you see the folder skills directly inside.

I have a Linux PC

Open your file manager and click Downloads. Double-click the ZIP file, or right-click it and choose Extract. (Some browsers have already unzipped it.)

In the steps below, "the ai-architecture folder" means the unzipped folder.

Done when you can open the ai-architecture folder, then skills, and you see the folders planning and continuity.

Stuck? See Stuck? Ask for help.

Step 4: Add its skills to a project

You will copy the whole folders skills/planning and skills/continuity into a folder named .agents/skills inside your project. It does not change your global configuration. For real work, use your own project folder. To try it first, make a new empty folder, for example my-project in your Documents folder; that name is this guide's own suggestion.

4a. Choose the project folder. Make the new folder, or pick your project.

Done when the folder exists and you know where it is.

4b. Show hidden files. On a Mac or Linux PC, names that start with a dot are hidden, so you would not see your .agents folder. Turn hidden files on in your file window:

  • Mac (Finder): open your project folder and press Command + Shift + . (the full stop key). Press it again to hide them.
  • Linux (Files): open your project folder and press Ctrl + H.
  • Windows (File Explorer): click View → Show → Hidden items. (Windows 10: click the View tab and tick Hidden items.)

Done when the setting is on. (You will see it working in 4c.)

4c. Check that .agents is not already there. Open your project folder. If it already contains .agents/skills/planning or .agents/skills/continuity, stop: keep what is there and ask for help if you want both. Do not copy over a folder you have not inspected.

Done when neither folder exists yet.

4d. Create the folder .agents in your project. Make a new folder and name it .agents, with the dot. On a Mac, Finder may warn that names beginning with a dot are reserved and will be hidden: choose the button that keeps the dot. (If .agents is already there, use it.)

Done when .agents is shown inside your project folder.

4e. Create skills inside .agents. Open .agents, make a new folder, name it skills. (If it is already there, use it.)

Done when .agents/skills exists.

4f. Copy planning. Copy the whole folder skills/planning from the ai-architecture folder, with everything inside it, into .agents/skills.

Done when .agents/skills contains planning.

4g. Copy continuity. Copy the whole folder skills/continuity the same way.

Done when .agents/skills contains continuity.

I have a Windows PC

ai-architecture has no installer and no Windows copy instructions. Copying in File Explorer is our adaptation of its general rule, and it is untested. Its own copy instructions are written for a POSIX shell (the kind of terminal Linux and Mac use), so they do not apply here.

4h. Compare with the listing. The result should look like this:

Folder layout
<project>/.agents/skills/
  planning/
  continuity/

Done when the folders look exactly like the listing above. Do not copy a folder named skills into .agents/skills: that nests it twice.

Stuck? A folder is missing or nested twice. Copy the exact folder listing and see Stuck? Ask for help.

Step 5: Check that Codex sees them

5a. Open the terminal in the project folder. In the terminal, type cd (the letters c, d and a space). In your file window, drag your project folder icon into the terminal window. This writes out the folder's path, its address on your computer. Press Return/Enter.

Done when the line you can type on now mentions your project folder.

5b. Start Codex and sign in. Type codex and press Return/Enter. The first time, choose Sign in with ChatGPT (or another sign-in method it offers), as the official page describes.

Codex may ask more questions before you can type, for example whether to trust this folder. Use the arrow keys and Enter to choose; for your project folder, pick Trust and continue. The wording may differ by version. Later, Codex may ask you to approve a command before it runs it. Read what it wants to run, and approve it only if it matches what you asked for.

Done when Codex shows a prompt where you can type a message.

Stuck? A sign-in loop or a plan error. Copy the exact text and see Stuck? Ask for help.

5c. Ask Codex which skills it has. You are now inside Codex, no longer in the normal terminal. Type this line, or type $ to see the skills you can mention.

Inside Codex
/skills

You should see: planning and continuity in the list.

Done when planning and continuity appear in the list.

Stuck? One is missing → close the terminal window, open a new one, go back into the project folder as in 5a (type cd, drag the folder in, press Return/Enter), then type codex again. Still missing → check the folder listing in step 4, then see Stuck? Ask for help

You are done with the main setup. The rest is for research work. If that is not you, you can stop here.

Part 3 · Provenote (optional, for research work)

Provenote is a free add-on that turns papers into notes and checks every quote in your notes against the page you cite. You only need this part if you work with papers and notes. It does not depend on Part 2. Steps 1 and 2 come first. If you skip Part 2, you sign in to Codex in step 8b.

Step 6: Get Provenote

You need the Provenote download, and one folder in it.

6a. Download it. In your web browser, open Provenote's front page on GitHub. Click the green Code button, then Download ZIP. You do not need Git for this route. A ZIP file is a folder packed into one file; unzipping unpacks it. The file is called provenote-main.zip and usually lands in your Downloads folder.

Done when provenote-main.zip is in your Downloads folder.

6b. Unzip it. Your file window is called Finder on a Mac and File Explorer on Windows. On Linux, use your file manager.

I have a Mac

Open Finder, click Downloads, double-click provenote-main.zip. (Some browsers have already unzipped it.)

I have a Windows PC

Open File Explorer, click Downloads. Do not just double-click the ZIP: that only lets you peek inside. Right-click provenote-main.zip, choose Extract All…, then click Extract. Windows may have made a provenote-main folder inside another one: open them until you see the folders docs and examples directly inside.

I have a Linux PC

Open your file manager and click Downloads. Double-click provenote-main.zip, or right-click it and choose Extract. (Some browsers have already unzipped it.)

In the steps below, "the extracted folder" means that folder (provenote-main), so docs/codex/SKILL.md is provenote-main/docs/codex/SKILL.md. "Download ZIP" gives you the latest version of the branch, which may be newer than the commit this guide was tested with (see About this guide).

Download the ZIP, then find the example folderOn the GitHub page the green Code button opens a menu with Download ZIP. After unzipping you have a folder provenote-main that contains examples, which contains basic.Download, unzip, find the folder1On GitHub: Code, then Download ZIP<> Code ▾Download ZIP← the green Code button2Unzip it. You get these folders:provenote-mainexamplesbasicused in step 8On Windows there may be a secondprovenote-main inside the first.Illustration, not a screenshot

Done when you can open provenote-main, then examples, then basic.

Stuck? See Stuck? Ask for help.

Step 7: Add the Codex adapter by copying files

This follows INSTALL.md. It installs the adapter, one self-contained Codex skill, into one project. It does not change AGENTS.md, your global configuration, or any Claude plugin. A skill is a set of instructions the assistant follows for one kind of job.

For this guide, use provenote-main/examples/basic as your project. It is a demo only: it sits inside the download, which has its own AGENTS.md at the top and is not a Git repository, so Codex may treat it a little differently from a real project. For real work, repeat this step in your own project folder.

You will make a folder named .agents (it starts with a dot) and copy files into it. The result should look like this:

Before and after: the folder gets an .agents folderBefore: examples/basic holds its usual folders. After: it also holds a folder .agents, which holds skills, which holds provenote, which holds SKILL.md and the folders bin, templates, agents and workflows.Before and afterBEFOREexamples/basic00_admin01_notes… the other foldersAFTERexamples/basic.agentsnewskillsprovenoteSKILL.mdbin/templates/agents/workflows/(00_admin, 01_notes … are still there)Only .agents and what is inside it is new.Nothing is nested twice.Illustration, not a screenshot

Do the sub-steps one at a time. Each ends with a ✅ you can check.

7a. Show hidden files. If you did this in step 4b, it is still on: skip to 7b. On a Mac or Linux PC, names that start with a dot are hidden, so you would not see your .agents folder. Turn hidden files on in your file window:

  • Mac (Finder): open the basic folder and press Command + Shift + . (the full stop key). Press it again to hide them.
  • Linux (Files): open the basic folder and press Ctrl + H.
  • Windows (File Explorer): click View → Show → Hidden items. (Windows 10: click the View tab and tick Hidden items.)

Done when the setting is on. (You will see it working in 7b.)

7b. Check that .agents is not already there. Open provenote-main/examples/basic. If it already contains .agents/skills/provenote, stop: back it up and compare it; do not overwrite anything you have not inspected.

Done when there is no .agents folder yet.

7c. Create the folder .agents inside examples/basic. Make a new folder and name it .agents, with the dot. On a Mac, Finder may warn that names beginning with a dot are reserved and will be hidden: choose the button that keeps the dot.

Done when .agents is shown inside basic.

Your file window will not accept the name .agents? (click or tap to open)

Then a small installer from Provenote can build the folders for you. It uses the Python from step 1, makes the same layout as steps 7c to 7k, and refuses to overwrite a folder that differs. The lines below are copied from INSTALL.md.

In the terminal, first go into provenote-main: type cd, drag the provenote-main folder into the window, and press Return/Enter. Then type the line for your computer up to and including --project and a space. Instead of typing the example path after it, drag the basic folder into the window. Press Return/Enter. On Windows, write py -3 in place of python if that is the one that worked in step 1.

In the terminal
python3 scripts/install-codex.py --project /absolute/path/to/project
In the terminal
python .\scripts\install-codex.py --project C:\path\to\project

Then skip to 7l and compare the result.

7d. Create skills inside .agents. Open .agents, make a new folder, name it skills.

Done when examples/basic/.agents/skills exists.

7e. Create provenote inside skills. Open skills, make a new folder, name it provenote. This is the "final folder"; the rest of the copying goes into it.

Done when examples/basic/.agents/skills/provenote exists and is empty.

7f. Copy SKILL.md. Copy the file provenote-main/docs/codex/SKILL.md into the final provenote folder. Keep the name SKILL.md. Windows may show it as just SKILL; that is the same file.

Done when provenote contains SKILL.md.

7g. Copy bin. Copy the folder bin from provenote-main/plugins/provenote/ into the final provenote folder.

Done when provenote contains bin.

7h. Copy templates. Copy the folder templates from the same plugins/provenote/ folder into the final provenote folder.

Done when provenote contains templates.

7i. Copy agents. Copy the folder agents from the same plugins/provenote/ folder into the final provenote folder.

Done when provenote contains agents.

7j. Copy skills (the whole folder). Copy the whole folder provenote-main/plugins/provenote/skills into the final provenote folder.

Done when provenote contains a folder named skills.

7k. Rename that copy to workflows. Rename only the copy inside .agents/skills/provenote/; the original skills folder in plugins/provenote/ keeps its name.

Done when provenote contains workflows, and the original plugins/provenote/skills is still called skills.

7l. Compare with the picture. The final folder should look exactly like this listing:

Text
<project>/.agents/skills/provenote/
  SKILL.md
  bin/
  templates/
  agents/
  workflows/

Done when the folder looks exactly like the listing above and like the "After" part of the picture. Do not copy a folder named provenote into the final provenote folder: that nests it twice.

Stuck? A folder is missing or nested twice. Copy the exact folder listing and see Stuck? Ask for help. If the installer from 7c printed an error, copy that too.

Step 8: Check 1: discovery

8a. Open the terminal in the example folder. If you are still inside Codex, close the terminal window and open a new one (step 1). Type cd (the letters c, d and a space). In your file window, open provenote-main, then examples, and drag the basic folder icon into the terminal window. Press Return/Enter.

Done when the line you can type on now mentions basic.

8b. Start Codex. Type codex and press Return/Enter. If you did step 5, you are already signed in. If not, choose Sign in with ChatGPT (or another sign-in method it offers), as the official page describes. Codex may ask whether to trust this folder; for this demo folder, pick Trust and continue (the wording may differ by version). Approve a command only if it matches what you asked for.

Done when Codex shows a prompt where you can type a message.

8c. Ask for the workflows. You are now inside Codex, no longer in the normal terminal. Type the line below including the $ at the start; it belongs to the message, so do not skip it.

Inside Codex
$provenote show me the available Provenote workflows; do not run helpers yet.

Done when Codex lists Provenote workflows. This proves discovery only: Codex can find the skill. Whether the helper programs run is Check 2.

Stuck? Nothing is listed. Close the terminal window, open a new one, go back to 8a, and try once more. Still nothing: see Stuck? Ask for help. The Provenote authors have not tested live Codex discovery (adapters.md), so this check may fail on your Codex version even when the files are right.

Step 9: Check 2: the installed adapter helper

This runs the helper from the installed adapter folder, $provenoteKit/bin, on the bundled example, so you see real PASS and FAIL lines. It follows codex/SKILL.md. First leave Codex: close the terminal window and open a new one (step 1). In that normal terminal window, go into your project again as in 8a (type cd, then drag the basic folder into the window and press Return/Enter). The first line below sets provenoteKit to the installed adapter folder inside the folder you are in ($PWD stands for the folder you are in); type it exactly as shown. Here $provenoteKit is a variable name, unrelated to the $ in step 8c.

I have a Mac or a Linux PC

In the terminal
provenoteKit="$PWD/.agents/skills/provenote"
python3 "$provenoteKit/bin/verify-quotes" --strict 01_notes/lit/*.md
python3 "$provenoteKit/bin/verify-quotes" --strict 01_notes/negative/*.md

I have a Windows PC

PowerShell does not expand *.md, so the note paths are written out.

In the terminal
$provenoteKit = "$PWD\.agents\skills\provenote"
& python "$provenoteKit\bin\verify-quotes" --strict 01_notes/lit/lei2023financial.md
& python "$provenoteKit\bin\verify-quotes" --strict 01_notes/lit/kireyev2026pricebench.md
& python "$provenoteKit\bin\verify-quotes" --strict 01_notes/negative/lei2023financial.md

If py -3 was the one that worked in step 1, use & py -3 instead of & python.

You should see (observed on Linux through a local copy of the adapter, not in a live Codex session): 14 lines starting PASS. On a Mac or Linux PC they come from the second line; on Windows the second and third lines print 6 and then 8 of them. The terminal does not show an exit code, so look only at the PASS and FAIL lines.

Example output: quote check passesLines starting with PASS, one for each quote that was found. Three of the 14 are shown.Part of the window’s output, shortened:PASS kireyev2026pricebench.md:  quote 1 (p. 1)PASS kireyev2026pricebench.md:  quote 3 (p. 4)…PASS lei2023financial.md:  quote 6 (p. 11)Run 1 of 2: every line starts with PASS.You will see 14 PASS lines in all.Example: what yours should roughly look like

The last command prints one line, FAIL lei2023financial.md: quote 1 (p. 1): text not found in source. That failure is expected: the note contains an invented quote on purpose.

Example output: one expected FAIL lineOne line starting with FAIL: a quote that was not found in the source. This is expected.Part of the window’s output, shortened:FAIL lei2023financial.md:  quote 1 (p. 1): text not found  in sourceRun 2 of 2: one FAIL line. It is expected.The note holds an invented quote on purpose.Example: what yours should roughly look like

Done when you see 14 PASS, then the one expected FAIL.

Stuck? A different count of PASS lines, no FAIL, or a Python error. Do not trust the result. Copy the exact output and see Stuck? Ask for help.

Using the ChatGPT desktop app

Codex also runs inside the ChatGPT desktop app. This guide has not tried it; this note comes from OpenAI's own pages.

  • The app is available for Mac (Apple Silicon), Windows (from the Microsoft Store) and Linux (in preview). Sign in with your ChatGPT account.
  • Open a folder and choose your project's folder.
  • Skills in .agents/skills can be picked in the app by typing @.
  • The permission control sits below the message box. Its choices are Ask for approval, Approve for me, Full access and Custom (config.toml). Choose Ask for approval, and follow the same rule as in step 5: read what it wants to run, and approve it only if it matches what you asked for.

On Windows, OpenAI's pages say the app shares settings with Codex installed natively on Windows (not with Codex inside WSL). For Mac and Linux they do not say.

What is NOT the same as Claude Code

  • Copied roles are instructions, not imported permissions. The files in agents/ and workflows/ tell Codex how to behave. They do not carry over Claude's tool restrictions, and they do not make Codex read-only where it otherwise could write.
  • Serial review is non-independent. If Codex has no separate reviewer available, it runs the reading and review steps one after another in the same session. Treat that result as non-independent: the same model checked its own work.
  • Live Codex discovery is untested. The adapter is tested as a local copy; whether your Codex finds the skill by itself is not (see adapters.md).

The same limit as the Claude Code route applies: a passing quote check means the words are on that page of the text copy. It does not show that your claim is true (see What Provenote does NOT do).

Stuck? Ask for help

Do not reinstall again and again, and do not change settings to "fix" a message. Copy the help message below into a chat with ChatGPT or Codex, or into a new GitHub issue (that needs a free GitHub account). Fill in the blanks with the exact error text, never a summary, and leave blank whatever you do not know. Before you send it, remove your name, your email address, and anything that looks like a key or token:

Text
I am following the Codex setup guide for ai-architecture (v0.4.2-preview) and Provenote (0.2.1, commit 373ca03).
Step: <step number and name>
Exact error text (copy and paste, unedited): <paste here>
My system: <macOS, Windows or Linux, and the version; on Windows, PowerShell or WSL>
Versions: codex --version = <...>; Python version = <...>
What I expected: <the "Done when" line from the guide>
Please tell me the smallest next thing to check. Do not install or change anything yet.

Report problems, and what worked, under Issues on the ai-architecture GitHub page or Provenote's front page on GitHub.

About this guide

Written with Claude; reviewed by the Provenote author (review pending)

  • Date: 2026-10-27
  • Written for: ai-architecture v0.4.2-preview (read, not run, by this guide's author). Tested with: Provenote 0.2.1 at commit 373ca03; Codex CLI 0.159.2 (observed on Linux, through a local copy of the adapter)
  • Status: untested on real Mac/Windows/Linux desktops. Everything here was written from the projects' own documents and OpenAI's official Codex pages. The only run was the quote check in step 9, on Linux, through a local copy of the adapter; no live Codex session and no real Mac, Windows or Linux desktop was used.

Later