Agent skillsResearch and web

AI Research Skills: an ML engineering manual your agent can follow

Researchers lose more time debugging infrastructure than testing hypotheses, and agents trip over dozens of frameworks. Orchestra Research gathered 98 skills across 23 categories: each one covers a single tool such as Megatron-Core, vLLM or TRL, while the autoresearch skill runs the whole lifecycle and routes to the right domain skill.

13,897stars on GitHub · #45

Open on GitHub

Install

  • skills CLI
    npx skills add Orchestra-Research/AI-Research-SKILLs
  • npm
    npx @orchestra-research/ai-research-skills
  • Claude Code plugin
    /plugin marketplace add orchestra-research/AI-research-SKILLs

Download ZIP

Author
Orchestra Research
Licence
MIT
Pack
98 skills in the pack
Updated
11 October 2026
Agents
Claude Code, Hermes Agent, OpenCode, Qoder, Cursor, Gemini CLI, OpenClaw

What it does

  • autoresearch uses two loops: an inner one iterates fast on experiments, an outer one synthesises results and steers the project.
  • Fine-tuning and post-training guides for Axolotl, LLaMA-Factory, PEFT, Unsloth, TRL, GRPO, OpenRLHF, SimPO, verl and more.
  • Distributed training and optimisation: DeepSpeed, FSDP2, Megatron-Core, Flash Attention, bitsandbytes, GPTQ, AWQ, GGUF.
  • Serving and evaluation with vLLM, TensorRT-LLM, llama.cpp, SGLang and lm-eval-harness, plus mech interp via TransformerLens and SAELens.
  • Further sections cover ML paper writing with LaTeX templates and citation checks, RAG, agent frameworks and multimodal models.

When to use it

  • You need a LoRA or QLoRA fine-tune on a single GPU without guessing at flags.
  • You are standing up an inference server and choosing between vLLM, SGLang and TensorRT-LLM.
  • You want the agent to carry a project from literature survey to a paper draft on its own.

Ask your agent

  • Set up a QLoRA fine-tune of Llama 3 8B with Unsloth on one A100 and explain each hyperparameter
  • Serve this model with vLLM and tune it for 50 concurrent requests
  • Start autoresearch: test whether speculative decoding speeds up generation on our benchmark

Questions

How do I install AI Research Skills?

Interactively with npx @orchestra-research/ai-research-skills, which detects your agents, or with npx skills add Orchestra-Research/AI-Research-SKILLs. Claude Code users can also add the marketplace and install categories as plugins.

Which agents do AI Research Skills support?

The installer detects Claude Code, Hermes Agent, OpenCode, Qoder, Cursor, Gemini CLI and others. autoresearch can run continuously through Claude Code /loop or the OpenClaw heartbeat.

Can I install only some of the 98 skills?

Yes. The installer offers everything, a quickstart bundle, whole categories or single skills, and in Claude Code a category installs as a plugin such as fine-tuning@ai-research-skills.

Research and webAll skills