AI Research Skills

// Overview

AI Research Skills is a library of agent skills for machine-learning research and engineering. Each skill is deep, framework-specific guidance — how to actually use Megatron-LM, vLLM, TRL, Axolotl, LLaMA-Factory, DeepSpeed, Flash Attention and dozens of others — written from official repositories, real GitHub issues and production workflows, so an agent follows a framework's documented path instead of improvising from whatever it half-remembers.

The library spans 98 skills in 23 categories: ideation and ML paper writing, model architecture, tokenization, fine-tuning, post-training, distributed training, optimization, inference, data processing, evaluation, safety and alignment, agents, RAG, multimodal, prompt engineering, MLOps, observability, infrastructure and mechanistic interpretability. Above them sits an `autoresearch` skill — an orchestration layer with a two-loop architecture that runs the research lifecycle end to end and routes to the domain skills as it needs them.

Installation is agent-agnostic. An interactive installer auto-detects the coding agents on your machine (Claude Code, OpenCode, Cursor, Gemini CLI, Qoder and others), installs skills into `~/.orchestra/skills/` and symlinks them into each agent — falling back to copies on Windows — and can install everything, a quickstart bundle, a category or a single skill. Categories are also published as a Claude Code plugin marketplace, so you can pull in just `fine-tuning` or `post-training` if that is all you need.

// What it does

• 98 skills across 23 categories covering the full research lifecycle, from literature survey to paper writing
• An `autoresearch` orchestration skill with a two-loop architecture that routes to domain skills as needed
• Framework-level depth — Megatron-LM, DeepSpeed, FSDP, Ray Train, TRL, GRPO, verl, vLLM, TensorRT-LLM, llama.cpp, SGLang, PEFT, Unsloth and more
• Mechanistic-interpretability skills (TransformerLens, SAELens, pyvene, nnsight) alongside training and serving
• Interactive installer that detects installed agents and symlinks skills into each (copies on Windows)
• Install everything, a quickstart bundle, one category, or individual skills — plus `list` and `update` commands
• Also distributed as a Claude Code plugin marketplace, installable by category

// Getting started

The library installs through npx; there is nothing to clone. Skills land in ~/.orchestra/skills/ and are linked into each detected agent.

npx @orchestra-research/ai-research-skills

Read https://www.orchestra-research.com/ai-research-skills/welcome.md and follow the instructions to install and use AI Research Skills.

/plugin marketplace add orchestra-research/AI-research-SKILLs

/plugin install fine-tuning@ai-research-skills
/plugin install post-training@ai-research-skills
/plugin install inference-serving@ai-research-skills

npx @orchestra-research/ai-research-skills list
npx @orchestra-research/ai-research-skills update

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Reach for it when an agent has to drive a training or post-training framework correctly rather than approximately
• Reach for it to give a research agent a full lifecycle path — ideation, experiments, evaluation, paper — in one install
• Reach for it when you want only one slice of expertise, such as fine-tuning or inference serving, as a plugin
• Reach for it to standardise how a research team's agents use shared infrastructure like Megatron, vLLM or SkyPilot

// How AI Research Skills compares

AI Research Skills alongside other open-source agent skills & plugins tools AI/TLDR tracks, ranked by GitHub stars.