Guide · Open source
How to use gpt-instruct: a guide for authorized safety testing
Updated · written with AI from the project's own documentation and checked against our catalogue data
gpt-instruct is a Python project from MDX-Tom that bundles prompt text files for Codex, OpenAI's coding agent, along with a script that installs them into a Codex configuration and regression tests that score how those prompts perform. The README describes it as a jailbreak prompt pack, meaning its prompts aim to get models to bypass their usual restrictions. This guide explains how the tool is organised, how to run its safe preview command, and why it is only appropriate for authorized research and controlled evaluation.
gpt-instructMDX-Tom/gpt-instruct · ★ 11.1k · MIT Install options & downloads Quick start
- 1Confirm your purpose. Make sure you have authorization to test in the environment you plan to use and that your goal is research or controlled evaluation.
- 2Install the prerequisites. Install Python 3.8 or newer and Git, and confirm that Codex is installed and configured.
- 3Clone the repository. Run git clone on the project URL and change into the gpt-instruct directory.
- 4Run the dry run. Run codex-instruct.py with --apply, a version flag and --dry-run to preview the change without writing configuration.
- 5Review and record. Read the preview output and the README's warnings, then log what you tested before deciding anything further.
What gpt-instruct contains
The repository provides prompt text files written as instructions for Codex, along with a Python script named codex-instruct.py that copies a chosen instruction file into a Codex configuration. It also includes regression tests that run the prompts against several test cases so results can be compared between versions.
The README states that the project is a jailbreak prompt pack and that it is intended to improve AI safety research. Treat it as a research artifact for studying how instruction files change model behaviour, not as a general productivity tool.
- Prompt text files for Codex (Markdown instruction files)
- codex-instruct.py, the installer script
- Regression tests and evaluation scripts for comparing prompt versions
- MIT licence, written in Python
Check the prerequisites before you start
The README lists three prerequisites: Python 3.8 or newer, Git, and a working Codex installation that is already configured. The project uses Codex's official configuration mechanism and, according to the README, does not modify binaries, intercept network traffic or tamper with processes.
- Python 3.8+
- Git
- Codex installed and configured
Clone the repository and preview with a dry run
Clone the repository, then run the installer in dry-run mode first. According to the README, the dry-run flag previews the change without writing any configuration, so you can see what would happen before anything touches your Codex setup.
The README also notes that a snapshot of the pre-deployment state is saved by the script, and that the --reset option only restores the instruction file the project manages, leaving provider, model and authentication settings alone.
- Run the dry run before any real deployment
- Use --codex-dir to point at a specific Codex home directory if yours is not the default
- Use --reset to remove only the instruction file this project manages
git clone https://github.com/MDX-Tom/gpt-instruct.git
cd gpt-instruct
python3 codex-instruct.py --apply --version gpt-5.6-v45 --dry-runUse it only for authorized testing
The README warns that using custom model instructions can put your account at risk, and it recommends a disposable account. It also says you should only use the tool in environments you are authorized to operate, and that you take the risk yourself.
If your goal is research, the practical value is the repeatable evaluation: running the same test cases against two instruction files and comparing results under the same model and reasoning settings. The README requires that scores only be compared when the model, reasoning level and method are the same.
- Use a disposable account, as the README recommends
- Run only in environments you have permission to test
- Compare results only between runs with identical model and reasoning settings
- Keep a record of each test run and the parent instruction file it came from
Understand the release status
The README describes a stable file and several release candidates, and says the project is under ongoing iteration. Release candidates are not presented as supported or finished products, so expect changes and do not treat any single score as a final result.
- The stable file is the default starting point named in the README
- Release candidates are labelled as pre-release
- Gate results in the README show that some pre-release versions did not pass every hard requirement
What people use it for
- Security researchers studying how custom instruction files affect a model's safety behaviour
- Developers running controlled A/B comparisons of Codex instruction files under identical settings
- Teams with authorized test environments that need repeatable evaluation scripts
Frequently asked questions
Is gpt-instruct free to use?
The repository is released under the MIT licence, as listed on its page. The README also states that the project is not used for any commercial activity.
What do I need to install gpt-instruct?
The README lists Python 3.8 or newer, Git, and a Codex installation that is already configured. The install method is clone-and-build, so you run the script from a cloned copy.
Does gpt-instruct bypass model restrictions?
The README describes it as a jailbreak prompt pack, meaning its prompts aim to get models to bypass their usual restrictions. It is intended for authorized research and evaluation only.
Is it safe to use on my main account?
The README warns that using custom model instructions can put accounts at risk and recommends using a disposable account. Use it at your own risk.
Are there alternatives to gpt-instruct?
The project's listing shows no comparable projects. For general Codex customisation, use Codex's own configuration options rather than a jailbreak prompt pack.
Why does my dry run show something unexpected?
The dry run is designed to preview changes without writing config. Check that your Codex home directory is correct, using --codex-dir if needed, and compare the output with the README before deploying anything.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
gpt-instruct suits security researchers and developers who need a repeatable harness for comparing Codex instruction files in environments they are authorized to test. It is a poor fit for everyday coding work, for anyone who needs a stable supported product, or for anyone who cannot accept the account risk the README describes.
Install options & downloads