Source · repo

ARIS: Markdown skills for autonomous ML research with cross-model review

Summary

ARIS ("Auto-Research-In-Sleep") is a bundle of about 83 Markdown skills that run ML research workflows in Claude Code and Codex CLI: idea discovery, literature search, experiments, paper review, rebuttals, and slides or posters. Its main idea is a cross-model loop in which one model does the work and a different model (Codex by default) reviews it. MIT licence, about 19.5k stars, actively maintained (last commits 2026-11-19, many external PRs merged). date_published is the date of the latest commit, not of a release.

Key points

  • Install in Claude Code: claude plugin marketplace add wanshuiyin/Auto-claude-code-research-in-sleep, then claude plugin install aris@aris, run /aris:setup once, and restart. Codex has an equivalent codex plugin add aris@aris. These commands were read from a summary of the README, so check them against the README before running.
  • Commands: /research-pipeline "<direction>" for an end-to-end run, /rebuttal (paper, reviews, venue limits), /research-implement-feature on a base repo, and /paper-slides or /paper-poster-html after acceptance.
  • Reviewer: the review loop needs a second model, by default Codex via an MCP server or the bundled ARIS-Code CLI, so it needs OpenAI (or other) API access.
  • Good for: literature reviews, rebuttal drafting, and turning a paper into slides; the skills are plain Markdown and can be read or adapted one by one.
  • Caution: it is built to run experiments unattended, including over SSH. Use a sandbox or dedicated machine with limited credentials. The README is promotional; treat its claimed results with care.

Why it matters

A large, maintained example of the "executor plus independent reviewer model" pattern packaged as ordinary skills, which can be reused outside research (for example for code or document review loops).