Codex GPT Pro Skill

A small Codex skill for asking ChatGPT Pro from a Codex session.

Codex can run local tools, edit files, inspect repos, and use browser-control plugins. It does not currently give users a direct "send this to ChatGPT Pro and wait" command. This skill adds that missing handoff.

When you ask Codex to "ask GPT Pro," "prompt GPT Pro," or "use ChatGPT Pro," the skill opens ChatGPT, selects Pro, sends your prompt, waits for the answer, and brings the result back to Codex.

Problem

Without this skill, using GPT Pro from Codex is a manual loop:

1. Leave Codex.
2. Open ChatGPT.
3. Check that the right account is signed in.
4. Pick Pro in the model selector.
5. Paste the prompt.
6. Wait through the long response.
7. Copy the answer back into Codex.

This skill automates that browser loop. It still respects the account boundary: you stay signed in through your browser, and Codex never handles your password, cookies, session tokens, one-time codes, or browser profile files.

What It Does

• Checks whether Browser Use / Playwright and/or Computer Use are available.
• Uses Browser Use / Playwright for the in-app browser route.
• Uses Computer Use for an already-authenticated default browser route.
• Asks for a first-run default route when both routes are available.
• Saves that route choice locally in gpt-pro/preferences.json.
• Opens https://chatgpt.com/.
• Starts a new chat unless you ask to continue the current one.
• Selects Pro / Research-grade intelligence.
• Enters your prompt exactly as requested.
• Waits until the Pro response is complete.
• Returns the latest assistant answer and the conversation URL.

What It Will Not Do

• It will not ask for your ChatGPT password, one-time code, recovery code, cookie, token, authorization header, localStorage, sessionStorage, or browser profile.
• It will not export or persist ChatGPT/OpenAI auth material.
• It will not bypass login, MFA, CAPTCHA, billing, rate-limit, or account-security prompts.
• It will not silently use another model if Pro is unavailable.
• It will not submit a prompt when you explicitly ask for a dry run.

Requirements

• Codex with Agent Skills support.
• A ChatGPT account with Pro access.
• At least one browser-control route available to Codex:
  • Browser Use / Playwright, for the in-app browser.
  • Computer Use, for your normal browser.
• A signed-in ChatGPT session in the browser route you want to use.

Install

Clone the repo:

cd ~/Downloads
git clone https://github.com/scasella/codex-gpt-pro.git

Then copy the gpt-pro skill folder into the skills directory your Codex build uses.

For Codex installs that use ~/.codex/skills:

mkdir -p ~/.codex/skills
cp -R ~/Downloads/codex-gpt-pro/gpt-pro ~/.codex/skills/

For Codex installs that use the newer Agent Skills location:

mkdir -p ~/.agents/skills
cp -R ~/Downloads/codex-gpt-pro/gpt-pro ~/.agents/skills/

For a repo-local install:

mkdir -p .agents/skills
cp -R ~/Downloads/codex-gpt-pro/gpt-pro .agents/skills/

Restart Codex, or start a new Codex session, after installing.

Development Install

If you plan to edit this skill, symlink the whole folder:

mkdir -p ~/.codex/skills
ln -s ~/Downloads/codex-gpt-pro/gpt-pro ~/.codex/skills/gpt-pro

Prefer a whole-folder symlink over a SKILL.md-only symlink. Some Codex versions have skipped single-file skill symlinks during discovery.

Usage

Explicit invocation:

Use $gpt-pro to ask GPT Pro: say hello.

Natural language invocation:

Ask GPT Pro to critique this rollout plan and wait for the result.

Prompt ChatGPT Pro with this research question and bring the final answer back here.

Force Computer Use for one request:

Use $gpt-pro via Computer Use to ask GPT Pro: what are the main risks in this design?

Force Browser Use / Playwright for one request:

Use $gpt-pro via Browser Use to ask GPT Pro: summarize this tradeoff.

Dry run without submitting:

Use $gpt-pro to open ChatGPT Pro and confirm Pro is selected, but do not submit a prompt.

First-Run Route Choice

If both routes are available and no preference exists, Codex asks which route to use by default:

• playwright: use the in-app browser / Browser Use route.
• computer-use: use the already-authenticated default browser route.

The choice is saved locally:

gpt-pro/preferences.json

That file is ignored by git. Delete it if you want Codex to ask again.

Route Tradeoffs

Browser Use / Playwright is usually easier for Codex to inspect. Use it when ChatGPT is already signed in inside the in-app browser.

Computer Use is useful when your normal browser is already signed in. It can be more fragile because it depends on window focus, zoom level, accessibility permissions, and visible UI state.

Verify Installation

After installing and restarting Codex, try:

Use $gpt-pro to ask GPT Pro: say hello.

Expected behavior:

1. Codex checks for Browser Use / Playwright and Computer Use.
2. If both are available and no local preference exists, Codex asks for a default route.
3. ChatGPT opens in the selected route.
4. Codex selects Pro.
5. Codex submits the prompt and waits for the answer.

To test without sending a prompt:

Use $gpt-pro to open ChatGPT Pro and confirm Pro is selected, but do not submit a prompt.

Update

cd ~/Downloads/codex-gpt-pro
git pull
rsync -a --exclude preferences.json gpt-pro/ ~/.codex/skills/gpt-pro/

Use ~/.agents/skills/gpt-pro instead if that is where your Codex install loads skills.

Uninstall

rm -rf ~/.codex/skills/gpt-pro

Or, for newer Agent Skills installs:

rm -rf ~/.agents/skills/gpt-pro

Restart Codex after uninstalling.

Repository Layout

codex-gpt-pro/
  README.md
  LICENSE
  .gitignore
  gpt-pro/
    SKILL.md
    agents/
      openai.yaml

The installable skill folder stays small on purpose. Setup notes, examples, and publishing guidance live in this README so Codex does not load them as skill instructions.

Publishing Notes

Suggested repository description:

Codex skill that lets Codex prompt ChatGPT Pro through Browser Use or Computer Use.

Suggested GitHub topics:

codex, codex-skill, agent-skills, chatgpt, chatgpt-pro, gpt-pro, computer-use, browser-use, playwright, openai

Before publishing:

• Confirm gpt-pro/preferences.json is not tracked.
• Confirm no screenshots, browser profiles, cookies, tokens, or local session files are present.
• Run the skill validator.
• Try a dry run that reaches the Pro composer without submitting a prompt.

Limitations

• This skill depends on the ChatGPT web UI and may need updates if the UI changes.
• It requires ChatGPT Pro access.
• It is not an API integration.
• It stops at private auth boundaries and asks the user to take over.

References

This repo follows the Agent Skills shape: one installable skill folder with SKILL.md, optional agents/openai.yaml, and distribution docs at the repo root.

• OpenAI Agent Skills documentation
• openai/skills
• huggingface/skills
• ComposioHQ/awesome-codex-skills
• skills-directory/skill-codex