Orchestra-Research/AI-Research-SKILLs installation guide

Install AI Research SKILLs for Claude Code on Apple-silicon macOS Tahoe

What this repository is

**Repository:** [Orchestra-Research/AI-Research-SKILLs](https://github.com/Orchestra-Research/AI-Research-SKILLs) **License:** MIT, although the README warns that libraries referenced by individual skills can have different licenses.

**Classification:** **Claude Code skill/plugin library** using **context injection and orchestration**, with a **Claude extension** installation mode.

The evidence supports this classification because the repository packages research guidance as `SKILL.md` files that Claude Code can load when relevant. It also supplies an **autoresearch** orchestration skill that routes research work to domain-specific skills. As an alternative to the general installer, the repository publishes category plugins through a Claude Code marketplace. It is not presented as a standalone desktop application or a replacement for Claude Code.

The README describes **98 skills in 23 categories**, including ideation, model architecture, fine-tuning, distributed training, evaluation, inference, RAG, mechanistic interpretability, safety, paper writing, and autonomous research orchestration. Some older statistics remain elsewhere in the README, but its June 2026 update says the inventory was reconciled to 98 skills and 23 categories.

Before you begin

This guide is specifically written for **macOS Tahoe on Apple silicon**. The supplied repository documentation does not state a minimum macOS release, provide a separate Apple-silicon build, or document a Rosetta requirement. Its recommended installer is an npm package invoked with `npx`, so the same documented command is used on Apple silicon.

You need:

1. **Claude Code already installed and usable.** The repository installs skills into coding agents; it does not install Claude Code itself.
2. **`npx` already available in Terminal.** The repository documents an `npx` installer but does not provide a first-party command for installing Node.js or npm. If macOS reports that `npx` is not found, stop and install a current supported Node.js/npm distribution using its official documentation, then return here. Do not improvise a package-manager command from this guide.
3. **Internet access.** The installer downloads the selected skills from GitHub.
4. Enough disk space for the selected content. GitHub metadata reports a repository size of about 21 MB, but the installed size and future updates can differ.

Installing these skills does **not** automatically install every research framework described by them. For example, individual skills discuss packages such as vLLM, SAELens, pyvene, and many others. Those are separate tools with their own platform, Python, GPU, API-key, and licensing requirements. On an Apple-silicon Mac, some GPU-focused workflows may require remote or non-macOS infrastructure. Install only the skill library now; add a framework later only when its relevant first-party instructions call for it.

Recommended installation: interactive npm installer

Open **Terminal** on your Mac. You can find it in Applications → Utilities, or open it with Spotlight.

Run the repository’s recommended interactive installer:

npx @orchestra-research/ai-research-skills

The package may need to be downloaded before its interface appears. Follow the prompts shown by the installer. According to the supplied package README, it will:

1. Detect installed coding agents.
2. Let you choose everything, selected categories, individual skills, or a quick-start bundle.
3. Download the skills from GitHub.
4. Install the selected content for the agents you choose.

For a normal personal setup, select **Claude Code** when offered. The global installation is the default. The installer stores one canonical copy under `~/.orchestra/skills/` and creates links under Claude Code’s `~/.claude/skills/` directory. This avoids maintaining a separate full copy for every supported agent.

The root README says the global installer falls back to copying on Windows; that fallback is not relevant to a normal macOS setup. No administrator command is documented, and you should not add `sudo` to the installer command.

Choosing what to install

For a first trial, the interactive **quick-start bundle** or a few relevant categories is easier to review than all 98 skills. Installing everything is appropriate if you expect to work across the complete research lifecycle and accept the larger body of model-visible guidance.

The central **Autoresearch** skill is an orchestration layer. The repository says it can manage a workflow from literature review and ideation through experiments, synthesis, and paper writing, routing to domain skills as needed. This is powerful, but generated research, citations, experimental claims, code, and safety conclusions still require human verification.

Verify the installation

After the installer finishes, use the package’s documented list command:

npx @orchestra-research/ai-research-skills list

Review the displayed installed skills. This is the repository-supported verification step available in the supplied files. The documentation does not provide a separate macOS health-check command.

Then open or restart Claude Code so it can discover the installed skill links. Ask Claude Code for help on a topic matching an installed skill—for example, planning an ML paper or selecting an evaluation workflow. Skills are described as auto-activating when relevant topics are discussed, so you do not normally run a skill as a macOS executable.

Remember what this verification proves: it confirms that the installer tracks the skills. It does not prove that every external framework discussed inside those skills is installed or compatible with your Mac.

Optional: install all skills without interactive selection

If you deliberately want the complete global library, the package README documents this direct command:

npx @orchestra-research/ai-research-skills install --all

This uses the same global model: canonical storage in `~/.orchestra/skills/` with links for detected agents such as Claude Code. Use the interactive installer instead if you want to see and choose agents or categories.

To install the documented post-training category directly, the package README provides:

npx @orchestra-research/ai-research-skills install post-training

The supplied files do not document every accepted command-line category spelling in one complete command table. Use the interactive installer for any category not explicitly demonstrated rather than guessing a command.

Optional: project-local installation

A local installation is useful when one research project needs its own selected skills, when a team wants to review and version those files, or when different projects should use different skill sets.

First, open Terminal in the intended project directory using Finder’s supported “New Terminal at Folder” workflow or your usual navigation method. This guide does not provide an invented directory-changing command because no project path can be known in advance.

To copy all skills into the current project’s agent directories, run:

npx @orchestra-research/ai-research-skills install --all --local

For only the documented post-training category, run:

npx @orchestra-research/ai-research-skills install --category post-training --local

For Claude Code, local skills are copied beneath the project’s `.claude/skills/` directory. The installer also creates `.orchestra-skills.json`, which tracks installed content. Unlike the default global installation, local mode uses direct copies rather than symlinks. The package README says this makes per-project sets reproducible and suitable for version control.

Verify a local installation from that same project directory:

npx @orchestra-research/ai-research-skills list --local

Use local mode intentionally: if the installer detects several coding agents, it may create corresponding skill directories for the agents selected during setup.

Alternative: Claude Code marketplace

The root README also documents a native Claude Code marketplace route. The following are **Claude Code slash commands**, not macOS Terminal commands. Enter them inside an active Claude Code session.

First add the marketplace:

/plugin marketplace add orchestra-research/AI-research-SKILLs

Then install a category. For example:

/plugin install fine-tuning@ai-research-skills

Other documented examples are:

/plugin install post-training@ai-research-skills
/plugin install inference-serving@ai-research-skills
/plugin install distributed-training@ai-research-skills
/plugin install optimization@ai-research-skills

Choose either the npm installer or marketplace workflow as your primary method. The npm installer is the repository’s recommended human installation and supports global or project-local sets; the marketplace installs category plugins directly through Claude Code.

Updating and removing skills

Update a global npm-managed installation with:

npx @orchestra-research/ai-research-skills update

Update a project-local installation from its project directory with:

npx @orchestra-research/ai-research-skills update --local

The supplied package README explicitly documents local removal:

npx @orchestra-research/ai-research-skills uninstall --local

The root README says the interactive installer can uninstall all or selected skills. For global removal, rerun the interactive interface rather than guessing an undocumented flag:

npx @orchestra-research/ai-research-skills

Safe first use

Treat installed skills as guidance supplied to Claude Code, not as proof that generated procedures are correct. Review commands, scripts, API calls, cloud-resource requests, and file changes before approving them. Some skill folders can contain optional helper scripts, templates, or assets. Research workflows may involve paid APIs, remote GPUs, credentials, downloaded models, or publication templates, none of which are configured merely by installing this library.

For macOS Tahoe on Apple silicon, begin with planning, ideation, documentation, or paper-writing tasks that do not require specialized CUDA hardware. When moving to training or inference, confirm the chosen framework’s current official Apple-silicon support and requirements separately. The installation is complete once the list command shows your selected skills and Claude Code can discover them.