AI Research Skills Library

AI Research Skills Library is an open-source, MIT-licensed collection of modular skill packages from Orchestra Research. It equips coding agents such as Claude Code, Cursor, Codex and Gemini CLI to carry out AI research work across the lifecycle, from ideation and experiments to paper writing. The README lists 98 skills in 23 categories, and the latest release shown is v1.7.2 (Qoder agent support).

What It Is

The library is a set of structured skill folders. Each folder has a SKILL.md quick reference, deep reference docs, and optional scripts and assets. A central autoresearch skill orchestrates the workflow with a two-loop architecture (inner optimization and outer synthesis) and routes to domain skills as needed. Those domain skills cover frameworks and tools such as Megatron-Core, vLLM, TRL, DeepSpeed, PEFT, FAISS, W&B and DSPy.

Skill Coverage

Categories include model architecture, tokenization, fine-tuning, mechanistic interpretability, data processing, post-training, safety and alignment, distributed training, optimization, evaluation, infrastructure, inference serving, agents, RAG, multimodal, prompt engineering, MLOps and observability. Ideation skills offer structured brainstorming frameworks. ML paper writing skills provide LaTeX templates for major venues, citation verification and academic plotting. Three Agent-Native Research Artifact skills compile research inputs into structured artifacts, record decisions and dead ends from sessions, and review research rigor.

Setup Path

Run the interactive installer with npx @orchestra-research/ai-research-skills. It detects installed coding agents, installs skills to ~/.orchestra/skills/ and symlinks them to each agent, with a copy fallback on Windows. It can also list, update and uninstall skills. Alternatively, an agent can be pointed at a welcome document to install and start autonomous research, or categories can be installed through the Claude Code plugin marketplace.

Demos

The repository links demo projects, including two papers the README says were produced by agents using the autoresearch skill, plus examples using NeMo Evaluator, GRPO/TRL, llama.cpp and FAISS.