These skill files turn an agent into a research assistant that can design experiments, train models, interpret internals, audit safety, and deploy inference. They are built for AI research scientists, LLM fine-tuning engineers, distributed training engineers, and MLOps teams who need reproducible, agent-guided workflows across the full model lifecycle.
Follow the documented setup, then try a first task.
Installs: Complete bundle of all 98 AI Research Skills, including the autoresearch orchestration layer; the interactive installer also offers quickstart, category, and individual selections. · Claude Code, Hermes Agent, OpenCode, Qoder, Cursor, Gemini CLI
Run the documented interactive installer in a terminal. When it presents the available scopes, select the “everything” choice to install all 98 skills rather than the quickstart, a category, or an individual skill. The installer detects installed coding agents and installs skills under ~/.orchestra/skills/, using symlinks or a Windows copy fallback.
npx @orchestra-research/ai-research-skills
Installs: Complete bundle of all 98 AI Research Skills, including the autoresearch orchestration layer. · Claude Code, Hermes Agent, OpenCode, Qoder, Cursor, Gemini CLI
Enter the following text in an AI-agent session; it is not a terminal command.
Read https://www.orchestra-research.com/ai-research-skills/welcome.md and follow the instructions to install and use AI Research Skills.
Installs: The fine-tuning marketplace category: Axolotl, LLaMA-Factory, PEFT, and Unsloth. · Claude Code
Enter this marketplace-add command in a Claude Code session; it is not a terminal command.
/plugin marketplace add orchestra-research/AI-research-SKILLs
After adding the marketplace, enter this category-install command in the same Claude Code session.
/plugin install fine-tuning@ai-research-skills # Axolotl, LLaMA-Factory, PEFT, Unsloth
Installs: The post-training marketplace category: TRL, GRPO, OpenRLHF, SimPO, verl, slime, miles, and torchforge. · Claude Code
Enter this marketplace-add command in a Claude Code session; it is not a terminal command.
/plugin marketplace add orchestra-research/AI-research-SKILLs
After adding the marketplace, enter this category-install command in the same Claude Code session.
/plugin install post-training@ai-research-skills # TRL, GRPO, OpenRLHF, SimPO, verl, slime, miles, torchforge
Installs: The inference-serving marketplace category: vLLM, TensorRT-LLM, llama.cpp, and SGLang. · Claude Code
Enter this marketplace-add command in a Claude Code session; it is not a terminal command.
/plugin marketplace add orchestra-research/AI-research-SKILLs
After adding the marketplace, enter this category-install command in the same Claude Code session.
/plugin install inference-serving@ai-research-skills # vLLM, TensorRT-LLM, llama.cpp, SGLang
Installs: The distributed-training marketplace category: DeepSpeed, FSDP, Accelerate, Megatron-Core, Lightning, and Ray Train. · Claude Code
Enter this marketplace-add command in a Claude Code session; it is not a terminal command.
/plugin marketplace add orchestra-research/AI-research-SKILLs
After adding the marketplace, enter this category-install command in the same Claude Code session.
/plugin install distributed-training@ai-research-skills
Installs: The optimization marketplace category: Flash Attention, bitsandbytes, GPTQ, AWQ, HQQ, and GGUF. · Claude Code
Enter this marketplace-add command in a Claude Code session; it is not a terminal command.
/plugin marketplace add orchestra-research/AI-research-SKILLs
After adding the marketplace, enter this category-install command in the same Claude Code session.
/plugin install optimization@ai-research-skills
Uses axolotl
Using the axolotl skill, review my Axolotl YAML snippet for FSDP version 2 and return a corrected minimal FSDP section. Check that context_parallel_size divides my total GPU count, and briefly identify any setting you changed. Do not run training.
Uses sentencepiece
Using the sentencepiece skill, draft a small starter consisting of a BPE training command with an 8,000-token vocabulary and a minimal Python encode/decode example. Treat data.txt as a sample placeholder and tell me to replace it with the path to my raw-text corpus. Do not execute anything.
Uses llama-cpp
Using the llama-cpp skill, recommend a suitable default GGUF quantization for my local hardware and draft a simple, non-interactive llama-cli chat command using the local model path and prompt I provide. Explain any GPU-layer offloading choice, but do not download a model or run the command.
Claude
Codex
Cursor
Gemini CLI
OpenClaw
OpenCode
Hermes
Qwen Code
DeepSeek