Orchestra-Research/AI-Research-SKILLs

What it solves

AI researchers often spend more time managing infrastructure and debugging tools than testing scientific hypotheses. This library provides a standardized set of "skills" that allow AI agents to autonomously handle the entire research lifecycle—from brainstorming and literature surveys to executing experiments and writing the final paper.

How it works

The project implements a two-loop architecture (inner optimization and outer synthesis) via an autoresearch orchestration layer. This central layer routes tasks to 98 specialized domain skills. Each skill contains expert-level guidance, production-ready workflows, and code examples for specific frameworks like Megatron-LM, vLLM, and TRL. These skills can be installed into coding agents (such as Claude Code, Cursor, or Gemini CLI) to give them the technical expertise required to conduct research.

Who it’s for

AI agents and the human researchers who orchestrate them, specifically those looking to automate the technical engineering aspects of machine learning research.

Highlights

• Comprehensive Coverage: 98 skills across 23 categories, including fine-tuning, distributed training, mechanistic interpretability, and post-training.
• Agent-Native Installation: Includes an interactive installer that auto-detects coding agents and symlinks skills directly to them.
• End-to-End Workflow: Supports the full pipeline from ideation and data processing to evaluation and LaTeX paper writing.
• Research-Grade Knowledge: Documentation is sourced from official repositories, GitHub issues, and production workflows.