chaseai-yt/grill-me-codex — explained in plain English
Analysis updated 2026-05-18
Force a clear plan out of yourself and Claude before starting a coding task, using structured questioning.
Get a second AI model's adversarial review of your plan before any code is written.
Keep a written record of plan revisions and the reasoning behind them in a project.
| chaseai-yt/grill-me-codex | mrxujiang/hicad | tornikegomareli/aurora | |
|---|---|---|---|
| Stars | 99 | 99 | 99 |
| Language | — | JavaScript | Swift |
| Setup difficulty | moderate | moderate | easy |
| Complexity | 2/5 | 3/5 | 2/5 |
| Audience | developer | vibe coder | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires installing and authenticating the OpenAI Codex CLI (version 0.130 or later) separately from Claude Code.
This repository contains a set of skills for Claude Code, the AI coding assistant, that add a two-step review process before any code gets written. The idea is to address two weak points in AI-assisted development: first, making sure the AI and the developer agree on what is being built, and second, having an independent AI from a different company evaluate the plan before work begins. The first step, called Act 1, has Claude ask you a series of questions until the plan is fully defined and all decision branches are resolved. A variant called grill-with-docs-codex also checks the plan against a project glossary and records architectural decisions as the conversation progresses. Once the plan is locked, Claude writes it to a file called PLAN.md. The second step, called Act 2, feeds that written plan to OpenAI Codex, which is a separate AI tool from a different provider. Codex reviews the plan and returns either an approval or a list of objections. If it requests revisions, Claude updates the plan and sends it back for another round. This loop continues until Codex approves or until a maximum number of rounds is reached (five by default). Codex is restricted to read-only mode throughout this process, meaning it can only evaluate and critique, never write or change files. A log file called PLAN-REVIEW-LOG.md records the full round-by-round exchange. The rationale the README gives for using a second AI from a different company is that the model doing the planning and the model grading the plan are the same system, which creates a blind spot. A different provider is more likely to catch structural problems in the plan. Installation involves copying the skill folders into Claude Code's local skills directory and having the Codex CLI installed and authenticated separately via an OpenAI or ChatGPT account. The skills are invoked using Claude Code's slash-command interface. Act 1 was built on top of earlier work by Matt Pocock, Act 2 and the packaging were added by Chase AI. The project is MIT licensed.
Claude Code skills that grill you to lock a build plan, then have OpenAI Codex adversarially review it before any code is written.
You can use, modify, and distribute this project freely, including commercially, as long as you keep the copyright notice.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.