AI Research Skills: a 98-skill library that turns a coding agent into a research assistant
Comprehensive open-source library of AI research and engineering skills for any AI model. Package the skills and your claude code/codex/gemini agent will be an AI research agent with full horsepower. Maintained by Orchestra Research.
At a glance
- What is it?
- AI Research Skills packages 98 skills in 23 categories so a coding agent such as Claude Code, Codex or Gemini CLI can carry research work from ideation to paper writing. It is guidance written for agents, so what you get out of it depends on the agent that reads it.
- Who is it for?
- Install AI Research Skills if you already run Claude Code, Codex or Gemini CLI on machine learning work and want framework guidance loaded into the agent before it touches training or inference code. Start with the npx installer and a single category such as fine-tuning, and read what lands in ~/.orchestra/skills/ before letting the autoresearch skill run an experiment unattended.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 92 days ago.
- What is it written in?
- Mainly TeX, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the skills library gives an agent
AI Research Skills starts from a complaint most machine learning researchers will recognise: the README says researchers spend more time debugging infrastructure than testing hypotheses. Its answer is not a new framework. It is a library of skills, packaged instructions that a coding agent loads and follows, covering the tools a research project touches along the way, from Megatron-LM and vLLM to TRL.
The library is built for people who already hand work to an agent. The installer detects Claude Code, Hermes Agent, OpenCode, Qoder, Cursor and Gemini CLI, and the repository topics add Codex. If you are looking for a Python package to import into your own training script, this is the wrong shape: nothing here runs on its own. The skills change what an agent knows when you ask it to fine-tune a model, write an evaluation or draft a paper.
Autoresearch and the two-loop routing
The centre of the library is the autoresearch skill. The README describes it as the orchestration layer that manages the full research lifecycle with a two-loop architecture and routes to domain skills as each stage needs them. The domain skills are grouped into 23 categories: ideation, ML paper writing, model architecture, tokenization, fine-tuning, mechanistic interpretability, data processing, post-training, safety and alignment, distributed training, infrastructure, optimization, evaluation, inference, MLOps, agents, RAG, prompt engineering, observability, multimodal work, emerging techniques and an agent-native research artifact category, plus autoresearch itself.
The repository layout mirrors that list, with numbered top-level folders from 0-autoresearch-skill/ through 22-agent-native-research-artifact/. The primary language GitHub reports is TeX, which fits the ML Paper Writing category shipping LaTeX templates and citation checks. The practical consequence is that the library is mostly text and templates; its value is in how specific each skill is about a given framework, not in executable code.
Installing AI Research Skills with the npx installer
The README's recommended route is one interactive command:
npx @orchestra-research/ai-research-skillsAccording to the README, the installer auto-detects which coding agents you have, installs skills to ~/.orchestra/skills/ and links them into each agent with symlinks, falling back to copies on Windows. It offers everything at once, a quickstart bundle, whole categories or individual skills. The same package lists and refreshes what you installed:
npx @orchestra-research/ai-research-skills list
npx @orchestra-research/ai-research-skills updateBecause the skills live in one directory and each agent points at it through a link, an update reaches every agent at the same time. That is convenient when it works and worth knowing when it does not: a skill that changes behaviour after an update changes it for Claude Code and Gemini CLI together. The README also gives agents a welcome document to read so they can install the library themselves, which saves typing but means the agent decides what to install.
Installing by category through the Claude Code marketplace
Claude Code users have a second route that installs categories instead of the whole library. The README adds the repository as a marketplace and then installs categories by name:
/plugin marketplace add orchestra-research/AI-research-SKILLs
/plugin install fine-tuning@ai-research-skills
/plugin install inference-serving@ai-research-skillsThe fine-tuning category covers Axolotl, LLaMA-Factory, PEFT and Unsloth, and the inference-serving category covers vLLM, TensorRT-LLM, llama.cpp and SGLang, according to the comments next to those commands. This is the approach to prefer when an agent only ever works on one part of a project. Loading 98 skills into an agent that will only fine-tune adapters adds instructions it has to sort through; a single category keeps its context about the task at hand. The cost is that the autoresearch layer, which routes between categories, has less to route to.
Where the library stops being useful
The README describes its skills as research-grade and production-ready, with documentation sourced from official repositories, real GitHub issues and production workflows. Those are the project's own claims, and a skill is only as current as the framework version it was written against. Fast-moving tools such as vLLM or TRL change flags and defaults between releases, and an agent following an outdated skill will produce confident, wrong commands.
The repository files show some loose ends. The root package.json is a stub: its version is 1.0.1 while releases are tagged v1.7.x, its licence field says ISC while the repository licence is MIT, and its test script only prints that no test is specified. The installer lives in a separate npm package, so none of this affects installation, but it does mean there is no automated check in the root of the repository that the skills still match the tools they describe.
The larger limit is the promise of autonomous research from idea to paper. The library can give an agent the steps; it cannot judge whether an experiment answered the question, and nothing in the README describes how results are validated.
Releases, maintenance and the MIT licence
The release history shows steady work through the first half of 2026. Version 1.4.0 on 2026-03-16 introduced the autoresearch skill for autonomous research from idea to paper. Version 1.7.1 on 2026-06-15 was about inventory consistency, a drift guard and security hardening, and version 1.7.2 the next day added Qoder agent support. The last push to the repository was on 2026-06-16.
The library is MIT licensed, which allows copying individual skills into your own agent configuration, editing them and keeping them private. For a team that already maintains its own agent instructions, that may be the most useful way to use it: take the skills for the frameworks you actually run, pin them to the versions you use, and update them on your own schedule instead of through the shared installer.
Editorial conclusion
Install AI Research Skills if you already run Claude Code, Codex or Gemini CLI on machine learning work and want framework guidance loaded into the agent before it touches training or inference code. Start with the npx installer and a single category such as fine-tuning, and read what lands in ~/.orchestra/skills/ before letting the autoresearch skill run an experiment unattended.
Frequently asked questions
What is an example of a skill in AI Research Skills?
The Fine-Tuning category is one example: it packages guidance for Axolotl, LLaMA-Factory, PEFT and Unsloth, and Claude Code users can install it on its own from the marketplace as fine-tuning@ai-research-skills.
What skills are needed for an AI researcher, according to this library?
The library splits research work into 23 categories, from ideation and ML paper writing to fine-tuning, post-training, distributed training, inference, evaluation, safety and alignment, and MLOps, with an autoresearch skill that routes between them.
How do I install AI Research Skills?
Run npx @orchestra-research/ai-research-skills. The interactive installer detects your coding agents, installs the skills to ~/.orchestra/skills/ and links them into each agent, copying them instead on Windows.
Which coding agents can use AI Research Skills?
The installer auto-detects Claude Code, Hermes Agent, OpenCode, Qoder, Cursor and Gemini CLI, and the repository topics also list Codex and Gemini.
Community notes