Mad Build

mad-build skill

A structured Claude + Codex collaboration on a creative or technical artifact. Use when you want a second pair of eyes from a different model on something you're making.

When this skill is the right tool

• Writing code: "MAD-build an R script that does X."
• Drafting prose: "MAD-build a one-page summary of attention mechanisms."
• Designing a plan: "MAD-build a research design for testing Y."

When this skill is NOT the right tool

• For a stress-test / referee report on a paper or grant, use

mad-research.

• For a single ad-hoc Codex call, use codex-bridge.

How it works (four steps)

1. Claude scaffolds /ASSIGNMENT.md from the user's request.
2. Both agents draft independently (Claude writes claude_v1/, Codex

via codex exec writes codex_v1/). Neither reads the other.

1. Each reviews the other's draft → round1_feedback/. **Strict

rule:** each side reviews ONLY the other side's draft, never its own.

1. Each revises in response → round2_revisions/. **Do NOT skip

this step.** If no changes are needed, document v2 = v1 explicitly. Then Claude merges into final/ and writes README_DEBRIEF.md.

See prompts/mad_build_protocol.md for the full operational protocol.

Safety defaults

This skill uses --sandbox workspace-write by default (Codex needs to write into the session folder). The bypass flag is behind an explicit --unsafe-yolo opt-in. See helpers/invoke_codex.md.

Files

mad-build/
├── SKILL.md                       — this file
├── prompts/
│   └── mad_build_protocol.md      — the 4-step operational protocol
├── helpers/
│   ├── invoke_codex.md            — verified codex exec patterns
│   └── doctor.md                  — pre-flight checks
└── README.md                      — what to install and why

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tjhavranek
• Source: tjhavranek/mad-research
• License: MIT

Install and usage instructions live in the source repository linked above.