Run investment due diligence by feeding a research question through the ARIS adversarial AI loop and saving findings to a persistent wiki for future reference.
Conduct a medical literature review with automatic adversarial critique between two different AI models to catch gaps and biases in the analysis.
Apply the 75+ pre-built research skills from the ARIS library to structured tasks like market research, journalism source verification, or engineering post-mortems.
Use the plan-draft-review-iterate workflow for any knowledge-work domain with Claude and GPT alternating as critic and author.
Requires API keys for at least two AI providers (e.g. OpenAI and Anthropic) to run the adversarial review loop between different models.
This repository is an extension of a project called ARIS, which stands for Auto Research in Sleep. The original ARIS system was built to automate academic research: it takes a research question, drafts work using one AI model, then has a different AI model from a different company critique that work, and iterates until the output is solid. This repository, ARIS-Anything, explores applying that same loop to non-academic tasks. The five-step loop at the center is: plan the question, draft an answer, review it adversarially using a separate AI (for example, having Claude review something GPT wrote, or vice versa), iterate on the feedback, and then save the findings to a persistent wiki so future work builds on what was already learned rather than starting from scratch. The Chinese terms in the description capture the idea: the original ARIS covered academic science (one specific shape of inquiry), while this repo explores all other structured investigation. The domains the README maps out include investment due diligence, legal case research, medical literature reviews, market and product research, self-directed learning, journalism source verification, and engineering incident post-mortems. Each of these gets matched to specific named workflows from the main ARIS library, which contains 75 or more pre-built research skills. The main ARIS repo the project builds on already has around 10,000 GitHub stars and was featured as the top paper on HuggingFace Daily Papers. ARIS-Anything is positioned as the next expansion: same methodology, different application areas. Several domain-specific sibling repos are planned, including ones for long-form video generation, product requirement documents, design critique, and a benchmarking harness for measuring research agent performance. These are described as scheduled rather than promised, with code to follow. The codebase runs across seven or more platforms including Claude Code, Codex CLI, and Cursor. The README is also available in Chinese. This is primarily a conceptual and workflow repository: the value is in the structured methodology and the named skill library, not in a deployable application.
← wanshuiyin on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.