A portable system-prompt override, PowerShell launcher, and prompt-bank regression harness for the Grok CLI.
Local Grok 4.5 calibration improved from 10/12 with the stock prompt to 12/12 with the override. Under the reference project's marker heuristic, the final three-level regression produced 360/360 passing cases: 348 batch-first-pass responses and 12 individually retried responses.
This benchmark checks failure/fallback markers. It is not a claim of human-verified semantic correctness for every open-ended response. The public runner additionally supports row-declared structural requirements.
Raw local configuration, credentials, sessions, logs, private prompt banks, and complete model outputs are intentionally excluded.
git clone https://github.com/qiuxinyuan321/supppergrok.git
Set-Location .\supppergrok
pwsh -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
. $PROFILECommands:
grok # SuppperGrok prompt
grok-stock # Original Grok prompt$prompt = Get-Content .\grok-4.5-unrestricted.md -Raw -Encoding UTF8
grok --system-prompt-override $promptpython .\scripts\run_prompt_bank.py `
--bank .\tests\prompt-bank.example.jsonl `
--level minimalSee README.md for configuration, privacy, and secret-scanning details.
The prompt structure and evaluation approach were adapted from MDX-Tom/gpt-5.6-instruct. See THIRD_PARTY_NOTICES.md.