How to use AI-Research-SKILLs: install and run it with your coding agent

What you need before installing

The skills are meant to be loaded into an AI coding agent, so you need one installed. The README names Claude Code, Codex and Gemini CLI as examples. The installer also detects other agents, including Hermes Agent, OpenCode, Qoder and Cursor.

The installer runs through npx, so you need Node.js. The README does not state a minimum Node.js version. The skills also assume you already have Python and ML environments set up.

• Node.js (required for npx)
• An AI coding agent such as Claude Code, Codex or Gemini CLI
• Python and ML environment experience (the skills assume it)

Install the skills with the interactive installer

The README recommends the interactive installer. Run it once and it walks you through choosing what to install.

The installer auto-detects coding agents on your machine, installs skills to ~/.orchestra/skills/ and creates symlinks to each agent. On Windows it falls back to copying files. You can install everything, the quickstart bundle, skills by category, or individual skills.

List and update installed skills

Once installed, you can check what is present and pull newer versions with two direct commands. The installer also supports uninstalling all skills or selected ones.

• list: shows installed skills
• update: updates installed skills to the latest versions

Start your agent with the autoresearch skill

The README says the autoresearch skill is the central orchestration layer. It uses a two-loop architecture (inner optimization and outer synthesis) and routes to the domain skills as needed. For agents, the README suggests giving the agent this instruction: read the welcome document at the URL below and follow it to install and use the skills.

Note that the README describes the goal of autonomous research from idea to paper but does not report how well this works.

• Welcome doc for agents: https://www.orchestra-research.com/ai-research-skills/welcome.md

Install through the Claude Code marketplace instead

If you use Claude Code, the README also lists a marketplace route. You add the marketplace once, then install a single category, for example fine-tuning, post-training, inference-serving, distributed-training or optimization.

• fine-tuning covers Axolotl, LLaMA-Factory, PEFT and Unsloth
• post-training covers TRL, GRPO, OpenRLHF, SimPO, verl, slime, miles and torchforge
• inference-serving covers vLLM, TensorRT-LLM, llama.cpp and SGLang

Find the right category for your task

The 23 categories cover the research lifecycle, including ideation, ML paper writing, model architecture, fine-tuning, post-training, distributed training, inference, evaluation, safety, RAG, multimodal and observability. Browse the category table to see which framework skills exist before you ask your agent for help.

Each skill has a linked folder in the repository with its own guidance. The README gives line counts and reference counts for each skill, which helps you judge how much material it contains.

• Use the table to find a framework by task, such as vLLM for serving or LoRA-style tools like PEFT for tuning
• Pick a narrow category if you only need one area, to keep the installed set small