AI Research Skills 🧩 Skill Open source

What it does

AI Research Skills, from Orchestra Research, is a large skill library for machine learning research and engineering. It aims to let a coding agent take a research project from literature survey and idea generation through experiments to a written paper. A central "autoresearch" skill runs the process in two loops, one for inner optimisation and one for outer synthesis. It hands work to the domain skills as needed, and it can keep running through Claude Code's /loop.

What is inside

At the time of writing there are 98 skills in 23 categories. Each one is built around a specific framework, with code examples, troubleshooting notes and reference files:

• Research workflow: autoresearch, ideation, ML paper writing with LaTeX templates, and academic plotting
• Training: model architectures (LitGPT, Mamba, NanoGPT), fine-tuning (Axolotl, PEFT, Unsloth), post-training (TRL, GRPO, OpenRLHF, verl) and distributed training (DeepSpeed, FSDP, Megatron-Core)
• Serving and optimisation: vLLM, SGLang, TensorRT-LLM, llama.cpp, Flash Attention and quantisation formats
• Evaluation and safety: lm-eval-harness, LlamaGuard, NeMo Guardrails
• Applications: RAG vector stores, agent frameworks, prompt tooling, multimodal models, MLOps and interpretability libraries

Works with

The interactive installer detects Claude Code, OpenCode, Cursor, Gemini CLI, Hermes Agent and Qoder. It also offers a Claude Code plugin marketplace organised by category. The skills use the standard SKILL.md layout, so other skills-compatible agents such as Codex can load them too.

How to install

npx @orchestra-research/ai-research-skills

In Claude Code, by category:

/plugin marketplace add orchestra-research/AI-research-SKILLs
/plugin install fine-tuning@ai-research-skills

Maintenance and safety

MIT licensed. The last repository update was in June 2026, so check framework versions against current releases. The installer writes skills to ~/.orchestra/skills/ and symlinks them into each agent. The README also suggests telling an agent to fetch and follow a remote welcome document. Read that document yourself before you let an agent act on it. Autonomous research runs can also start long, costly GPU jobs.

Who should use it

ML engineers and researchers who use coding agents for training, fine-tuning, evaluation or paper writing. It is less useful if you mainly build web or mobile apps.