Installation guide

dzhng/skills installation guide

MIT-licensed collection of 22 README-enumerated Agent Skills for Claude Code and other skill-compatible coding harnesses, centered on an autonomous software-factory loop that explores unknowns, writes and implements living specs, reviews each slice, records agent-made choices, validates visuals, and closes shipped specs into rationale records.

Install `dzhng/skills` for Claude Code on macOS Tahoe (Apple silicon)

What this repository is

`dzhng/skills` is a public, MIT-licensed collection of reusable agent skills for software-development work. The repository describes the skills as “small, composable, and hackable” and says they work with Claude Code as well as several other agent harnesses.

For Claude Code, the best-supported classification is:

This is not a native Apple silicon application, and the supplied files do not describe any Intel-versus-Apple-silicon difference. There is therefore no repository-supported Rosetta step, Homebrew step, architecture flag, or macOS Tahoe workaround to apply.

Before you begin

You need a working Claude Code setup and a Terminal environment in which `npx` is available. The repository’s documented installer is invoked through `npx`, but the supplied files do not state a required Node.js version or provide a first-party command for installing Node.js, npm, `npx`, or Claude Code.

If Terminal reports that `npx` is not found, stop there and install it using the current first-party instructions for your chosen Node.js distribution. Do not guess at a package-manager command from this guide: none is documented by this repository.

You will also need internet access while `npx` obtains and runs the skills installer. Close or save unrelated Terminal work before proceeding so it is clear which prompts belong to this installation.

Open Terminal on macOS Tahoe

  1. Open **Spotlight** by pressing **Command–Space**.
  2. Type **Terminal**.
  3. Press **Return**.

You can run the documented installer from the Terminal window. The repository does not require cloning it first and supplies no clone-based installation procedure that needs to be used here.

Choose an installation method

The README documents two installer forms:

  1. Install from the repository with the standard add command.
  2. Add `--list` to choose individual skills.

For a layperson, selective installation is the more transparent starting point because it lets you see and choose the skills rather than requesting the whole collection at once.

Recommended: choose individual skills

Copy and paste this exact documented command into Terminal, then press **Return**:

npx skills add dzhng/skills --list

Follow the installer’s on-screen selection prompts. Choose Claude Code if the installer asks which supported harness or agent should receive the skills. Select only the skills you understand and expect to use. The repository does not supply the installer’s exact prompt sequence, so the wording you see may differ; do not assume an option that is not displayed.

Useful areas described by the repository include:

Some skills call for capabilities beyond Claude Code itself. In particular, the `codex` skill expects a separately installed and authenticated Codex CLI, and the `claude` skill expects Claude Code on the command path and authenticated. Installing this collection does not, based on the supplied evidence, install or authenticate those external command-line tools.

Alternative: request the repository’s skills directly

If you want the default installation rather than choosing from a list, copy and paste the repository’s primary command:

npx skills add dzhng/skills

Follow every prompt shown by the installer. Do not add `sudo`, architecture flags, package-manager options, or destination paths: the supplied repository files do not document any of those additions.

Understand what may be installed

The README lists a broader library of skills, while `.claude-plugin/plugin.json` explicitly names a Claude plugin called `skills` containing twelve paths:

The README’s installer may discover skills independently of that plugin manifest, but the supplied installer implementation was not provided. Consequently, this guide cannot promise whether the default command installs every README-listed skill, only the manifest’s set, or presents another scope determined by the installer. The `--list` form is the safest way to observe the choices actually offered.

The README also says a skill folder can be copied into a harness skills directory such as `.claude/skills/`. That supports manual installation in principle, but it does not provide a complete, exact manual copy command or define whether your intended setup should be global or project-local. To avoid inventing commands or destinations, this guide uses the documented `npx` installer instead.

Confirm the result

Treat the installer’s successful completion message and its displayed list of selected targets as the primary installation confirmation. The supplied repository files do not document a dedicated status, list, health-check, or uninstall command, so this guide does not invent one.

Next, open the Claude Code environment where you intended to use the skills. Ask Claude Code to use one of the skills you selected, or invoke it by name through the interface your Claude Code version provides. Exact host UI and invocation syntax can change, and the repository does not provide a universal verification command.

A good first test is a low-risk, read-oriented task. For example, if you selected `eli5`, use it to explain an existing specification or proposed change. If you selected `review`, use it only when you have a finished substantive change: that skill coordinates a shape review, diff review, and documentation pass. Do not start with a skill that delegates implementation or performs broad autonomous work until you have read its `SKILL.md` and understand its permissions.

macOS and safety notes

The `preview-shots` skill is specifically macOS-oriented and uses Preview to display saved image artifacts. That makes it relevant to macOS Tahoe, but installing the collection does not mean every skill will run automatically.

Read a skill before allowing it to execute tools. The inspected `codex` and `claude` skills can invoke separately authenticated agents and discuss write-enabled or bypassed permission modes for narrowly defined tasks. Their own guidance emphasizes explicit user approval, isolated worktrees for broad permissions, diff review, and verification. Installation alone should not be treated as consent for delegated edits or unrestricted commands.

Likewise, the `renderer` skill assumes a project already has the relevant browser, GPU, renderer, and test setup. The repository does not document a universal dependency-install command for those project-specific tools.

Troubleshooting boundaries

npx skills add dzhng/skills --list

At that point, installation is complete when the installer reports success for Claude Code and the selected skill is available in the intended Claude Code environment.