The parallel track — an alternative reasoning LLM and agentic coder
Everywhere the course says "Claude / ChatGPT," this is the other half of that choice — not a lesser option, a genuine alternative.
AGENTS.md too.Offered as an alternative to Claude/Claude Code from setup through referee — not used for the Claude-Code-specific skill/subagent exercises (D1·2–D1·5).
The tool setup lab installs “Claude Code / Codex” side by side — pick one or try both. The same AGENTS.md ground rules apply either way.
The schedule lists “Claude/ChatGPT, optional Gemini” for the hypothesis loop, and “Claude/ChatGPT” for design & identification. The generate–critique–refine loop and the adversarial design review work the same way with either model.
“ChatGPT Data Analysis” is the listed alternative to Claude Code for running the analysis and finding the pre-trend confound (D2·3), and for writing up the naive vs. robust estimate (D2·4).
Paired with Paperpal/Writefull in the schedule (“Claude/ChatGPT + Paperpal/Writefull”) for the writing session — ChatGPT drafts, a dedicated writing tool polishes.
Works as an alternative critical-reviewer persona for the referee pass, same checklist logic (causal language, pre-trends, citations, numbers, disclosure) run through a different model.
A design memo or referee pass reviewed by only one model can share that model's blind spots. Running the same adversarial question through both products is the cross-tool version of D1·5's “do two independent checks agree” pattern — applied across products, not just subagents.
The course's rules — no invented citations, no bare causal claims, name the identification strategy, disclose every AI-assisted step — aren't features of any one product. Your own AGENTS.md for ChatGPT/Codex should say the same things Claude's does.
See the home page for the full pipeline — every stage where this tool appears is listed as a Claude/ChatGPT choice, never ChatGPT-only.