Search skills/

AI Research Skills (Orchestra) Claude skills

98 skills for AI research and ML engineering, plus an autoresearch orchestrator that runs a two-loop experiment and synthesis cycle.

Orchestra-Research/AI-Research-SKILLs 98 skills Planning, TDD & debuggingResearchScience & academia
Our verdict Recommended with caveats

A broad, MIT-licensed ML library with a real autonomous-research orchestrator, but install selectively and add your own spend and permission limits before running it unattended.

4/5 Depth expertise shipped ! Security Broad access 18d Maintenance Updated 18 days ago 14K Adoption stars OK License MIT, commercial OK

What AI Research Skills (Orchestra) does

This library has two layers. The orchestration layer is a skill called autoresearch (about 25 KB, with references and templates). It sets up a workspace with research-state.yaml, a research log and findings.md, then alternates between an inner loop of fast experiments and an outer loop that reviews results, updates findings and picks the next direction. It routes to domain skills for execution and is designed to keep going through Claude Code's /loop or an OpenClaw heartbeat. Its instructions say to run autonomously without asking for confirmation and to keep moving after failures.

The second layer is domain knowledge across about 20 categories of ML work: model architectures (nanoGPT, RWKV, Mamba, TorchTitan), tokenization, fine-tuning (Axolotl, LLaMA-Factory, PEFT, Unsloth), mechanistic interpretability (TransformerLens, nnsight, SAELens), post-training (TRL, GRPO, OpenRLHF), distributed training, inference, evaluation, safety and guardrails, RAG, multimodal, MLOps and observability. The 98 SKILL.md files average about 13 KB and cite official docs and issues, with reference folders, a few templates and only two Python files in the snapshot. An npm installer (npx @orchestra-research/ai-research-skills) detects your agents, installs the skills under ~/.orchestra/skills and symlinks them into each agent. The README calls itself comprehensive; treat the domain skills as documentation digests.

Good for

  • ML engineers who want agents to follow framework-specific recipes for training, fine-tuning and inference
  • Long autonomous experiment loops with persistent findings.md state
  • Mechanistic interpretability and post-training workflows

Not for

  • Installing all 98 skills into one agent without selecting categories
  • Unsupervised runs without budget limits, since the orchestrator is told not to ask or stop
  • Non-ML projects

What's inside depth 4/5

98 skills averaging 13 KB with reference folders, an orchestrator with state templates and a continuity loop, but very few scripts and no evals or tests.

Skills
98
Reference docs
268
Scripts
2
Tests / evals
3
Templates
61
Avg SKILL.md
12.8 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • Node.js for the npx installer
  • GPU and the ML frameworks each skill covers (PyTorch, Axolotl, vLLM and others)
  • Optional: W&B, MLflow or other services for MLOps skills
  • Claude Code /loop or OpenClaw heartbeat for continuous autoresearch

Security review

Broad access

This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.

  • Loads instructions from a server at run timeThe agent install path is to have the agent read https://www.orchestra-research.com/ai-research-skills/welcome.md and follow it, so the install and usage instructions come from the vendor's server rather than the repo.README.md:95
  • Turns off permission prompts or safety checksThe autoresearch skill states it runs fully autonomously, tells the agent not to ask for permission or confirmation, never stop, and to set up a recurring /loop (or OpenClaw cron job) every 20 minutes to keep going.0-autoresearch-skill/SKILL.md:16

A library of about 98 AI-research and ML-engineering skills, mostly documentation, with two example training scripts. The README offers an agent install route in which your agent follows instructions from a vendor-hosted welcome.md, an npx installer that links skills into every detected agent, and an autoresearch skill that tells the agent to run indefinitely without asking for confirmation. Part of the repo was not available to us.

Worth knowing

  • Runs remote code at an unpinned versionInstallation and updates run npx @orchestra-research/ai-research-skills (unpinned), which installs skills into ~/.orchestra/skills and symlinks them into each detected agent.README.md:89
  • Changes files outside your projectThe installer writes skills to ~/.orchestra/skills/ and links/copies them into the config folders of every detected coding agent; autoresearch also creates a recurring scheduled loop.README.md:103
  • Installs with curl | shReference docs include curl -LsSf https://astral.sh/uv/install.sh | sh and wget -O- https://lambdalabs.com/install-lambda-stack.sh | sh - as installation steps for third-party tools.03-fine-tuning/axolotl/references/other.md:310
  • Needs an API key for its own serviceLambda, NDIF and similar skills use LAMBDA_API_KEY / NDIF_API_KEY, sent only to those services.09-infrastructure/lambda-labs/SKILL.md:144
  • Relies on a CLI or package we could not reviewDepends on the @orchestra-research/ai-research-skills npm installer and many third-party ML frameworks and CLIs (pip install, docker run, git+https installs) whose code is not in this repo.README.md:116

What we checked

The repository (529 files) on Oct 29, 2026. Not a full review: 345 of 529 files were not downloaded (large repository; we reviewed the skill folders); it relies on code outside the repository: npm package @orchestra-research/ai-research-skills (the installer); https://www.orchestra-research.com/ai-research-skills/welcome.md (remote bootstrap instructions); Third-party ML packages and install scripts referenced across the 98 skills.

NVIDIA SkillSpector static scan: 102 pattern hits (Privilege Escalation 27, MCP Rug Pull 17, Prompt Injection 10, Data Exfiltration 10, YARA Match 7, Excessive Agency 5). The CRITICAL and several HIGH hits are guard-model skills (llamaguard, prompt-guard, nemo-guardrails) quoting jailbreak and harmful-prompt examples as test data; 'exec' is an example GRPO reward function that executes model-generated code with a comment saying it must be sandboxed in production; docker --cap-add/--ipc=host and pip git+https lines are framework install guides; the 'hidden instructions' hit is an HTML comment in a presentation template telling the agent to fill it in.

A review, not a certification. Skills change after we check them — read the files before installing.

License

MIT — commercial use allowed. The README notes individual skills reference libraries with different licenses.

Can it run unattended?

Yes. The autoresearch skill keeps state in files and is built to continue through /loop or a heartbeat, but it tells the agent never to stop or ask, so you must supply budgets and a sandbox.

Alternatives to AI Research Skills (Orchestra)

Scientific Agent Skills

K-Dense-AI/scientific-agent-skills

K-Dense's library of 177 science skills covering bioinformatics, chemistry, clinical research, data analysis, lab tooling and scientific writing.

With caveats Broad access Depth 4/5

ARIS (Auto-Research-In-Sleep)

wanshuiyin/Auto-claude-code-research-in-sleep

A large markdown skill set for autonomous ML research: idea discovery, experiment runs and cross-model review loops where a second model critiques the first.

With caveats Broad access Depth 4/5

Hugging Face Skills

huggingface/skills

Official Hugging Face skills led by hf-cli, a Hub CLI reference generated from your local install, plus AWS/SageMaker deployment and operations skills.

Recommended Check first Depth 3/5

Academic Research Skills (ARS)

Imbad0202/academic-research-skills

Five very detailed Claude Code skills for the academic pipeline: deep research, paper writing, peer review, revision and systematic-review screening.

With caveats Check first Depth 5/5

FAQ

What is the AI Research Skills library?

It is an open-source library from Orchestra Research with 98 skills covering the AI research lifecycle: an autoresearch orchestrator, ideation, paper writing, and engineering skills for architectures, fine-tuning, post-training, distributed training, inference, evaluation and safety.

How do I install AI Research Skills?

Run npx @orchestra-research/ai-research-skills, which detects your coding agents and installs the skills to ~/.orchestra/skills with symlinks. You can also clone the repo and copy only the categories you need.

Does the autoresearch skill run without supervision?

Yes, that is the intent: it keeps state in research-state.yaml, a log and findings.md, and continues via Claude Code /loop or an OpenClaw heartbeat. Its instructions say not to ask for confirmation, so set time, spend and sandbox limits yourself.

AI Research Skills vs K-Dense scientific skills: which one?

Orchestra's library is focused on ML and AI engineering (training, interpretability, inference). K-Dense's collection targets natural sciences such as biology, chemistry and clinical research. Pick by domain.

Sources: GitHub repository (stars, license, file tree, last push Oct 11, 2026). Review based on the README and SKILL.md files as published.