How I Combined Claude Code and Codex to Produce an Accepted Conference Paper

Lu Gong, an AI algorithm veteran, describes his recent work on auto‑research and Vibe Researching within the medical AI domain, focusing on the MICCAI 2026 Challenge series.

He runs Claude Code and Codex in parallel on five competition tracks: lung‑airway segmentation (ATM26), brain ischemic‑stroke lesion segmentation (ISLES 2026), intracranial aneurysm detection (TopAnue 2026), intracranial vessel segmentation (TopBrain V2), and a prostate‑cancer diagnosis agent (CHIMERA).

Because handling five projects overwhelms human cognition, he conducts deep reviews only for ISLES 2026 and CHIMERA, delegating the remaining three entirely to Claude Code.

The workflow requires substantial GPU resources—at least two RTX 4090 cards—to avoid prohibitive experiment latency.

He previously documented a similar Claude‑Code‑only approach for a Kaggle competition, outlining the overall methodology.

For ISLES 2026, the task is 3‑D segmentation of T1‑weighted MRI ischemic‑stroke lesions across acute, sub‑acute, and chronic phases. Out of 369 registered teams, 60 submitted results; the main challenges are domain generalization, large lesion variability, multi‑lesion cases, and irregular shapes.

The pipeline starts with ChatGPT performing a preliminary deep‑research pass, after which all reports, PDFs, website information, and datasets are fed to Claude Code. Claude then runs a dynamic deep‑research workflow and produces a detailed competition plan.

With the plan ready, Lu sets up a training machine, establishes SSH access, logs experiments, and lets Claude Code connect to the GPU server to launch training—mirroring the earlier Kaggle workflow.

Claude Code then drafts a conference paper based on the experiments and results, compiling the manuscript with LaTeX into a PDF.

Codex subsequently reviews the draft, checking for factual inaccuracies, table and data correctness, and overly AI‑styled phrasing. After several human‑in‑the‑loop revisions, the paper is finalized.

The submission to OpenReview is accepted, earning a reviewer score of 7.

This case demonstrates the high ceiling of auto‑research: when researchers master Claude Code and Codex, they can dramatically accelerate scientific output, yet the author stresses that AI cannot replace human responsibility—manual review remains essential for publishable work.