ARIS (Auto-Research-In-Sleep) is an open-source research automation project with over 11,900 GitHub stars that lets Claude Code run the full ML research lifecycle while the user sleeps. This article dissects its architecture: a pure-Markdown system of 79 slash-command skills covering literature review, idea discovery, experiment execution, paper writing, review loops, rebuttals, and talk preparation. Its core innovation is cross-model adversarial review—Claude Code executes while GPT-5.5, accessed via Codex MCP, acts as a critical reviewer, mirroring double-blind peer review. The post covers its assurance-contract grading system (draft to submission), six named workflows including a six-stage Idea Discovery pipeline, installation in three commands, and a full port of 13 portable skills to the WorkBuddy platform, with adaptation strategies replacing Codex MCP review with structured self-check lists and slash commands with natural-language triggers. Roughly 40% of skills (GPU queueing, cross-model loops) were not portable. Includes installation instructions and design lessons: Markdown as the LLM-native interface, two-model博弈 beats single-model reasoning, and symlink-based updates.
ARIS Deep Dive: 79 AI Research Skills That Write Papers While You Sleep
Summary
ARIS (Auto-Research-In-Sleep) is an open-source research automation project with over 11,900 GitHub stars that lets Claude Code run the full ML research lifecycle while the user sleeps. This article dissects its architecture: a pure-Markdown system of 79 slash-command skills covering literature review, idea discovery, experiment execution, paper writing, review loops, rebuttals, and talk preparation. Its core innovation is cross-model adversarial review—Claude Code executes while GPT-5.5, accessed via Codex MCP, acts as a critical reviewer, mirroring double-blind peer review. The post covers its assurance-contract grading system (draft to submission), six named workflows including a six-stage Idea Discovery pipeline, installation in three commands, and a full port of 13 portable skills to the WorkBuddy platform, with adaptation strategies replacing Codex MCP review with structured self-check lists and slash commands with natural-language triggers. Roughly 40% of skills (GPU queueing, cross-model loops) were not portable. Includes installation instructions and design lessons: Markdown as the LLM-native interface, two-model博弈 beats single-model reasoning, and symlink-based updates.
This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981137