gpt-instruct is a versioned library of Codex prompt variants with a reproducible evaluation toolchain. A Python script applies a selected prompt version to the local Codex configuration, and a reset command restores the prior state.
Project overview
It pairs versioned Codex prompt releases with regression and scoring scripts whose pass/fail hard gates decide whether the next evaluation suite runs, giving a reproducible way to test prompt changes instead of relying on first-pass execution.
Project type
Prompt Engineering
Deployment
Refer to project documentation
License
MIT
Best for
AI engineers and prompt researchers who want versioned, reproducible Codex prompt variants with an evaluation toolchain.
Users of the Codex CLI who want a scriptable way to apply and reset prompt versions in its official configuration mechanism.
Key capabilities
A Python script applies a selected released prompt version to the local Codex configuration (e.g. --apply --version gpt-5.6-v45); a --reset command restores the model_instructions_file managed by this project.
Regression and scoring scripts run per model line (e.g. run_gpt56_sol_issue_regression.py with an explicit --model flag), and pass/fail hard gates determine whether the C suite runs.
Optional JB-A/JB-B suites use frozen official prompts with a fixed upstream judge model, together_ai/meta-llama/Llama-3-70b-chat-hf.
Limitations and risks
Both pre-release versions failed the B hard gates, so the C suite was not run.
Official JailbreakBench judging is an optional module and requires a Together API key, relying on a remote third-party service.
Coding is optional for use, so some comfort with the command line and Python scripts helps.
Getting started
git clone https://github.com/MDX-Tom/gpt-instruct.git, cd gpt-instruct, then run python3 codex-instruct.py --apply --version gpt-5.6-v45. Setup difficulty is documented as easy.