skillsets

Agent skills from other providers, in one folder for each provider.

skillsets.mp4

Why?

• Load skills only when you need them. If you install every skill globally, each agent session loads all of them. That fills the context and makes name conflicts. Here you choose the providers for each session.
• Use one copy with many agents. The same folders work in Claude Code, Codex, OpenCode, Pi, Cursor CLI and T3 Code.
• Review changes before you use them. Skills run with the full permissions of your agent. The skills are in git, and a daily workflow opens a PR when a provider changes, so you can read the diff before you merge it.
• Take only the skills you want. You can take all skills from a provider, or only some of them.

Providers

Folder	Source	Skills
mattpocock/	mattpocock/skills	All
pstack/	cursor/plugins/pstack	All
9arm/	thananon/9arm-skills	debug-mantra, post-mortem, scrutinize
superpowers/	obra/superpowers	All
humanlayer/	humanlayer/skills	All
addyosmani/	addyosmani/agent-skills	All
tstack/	t3dotgg/tstack/skills	All

Usage

The fastest way: choose skillsets and an agent, and the CLI prints the command (or the T3 Code settings) for you.

npm install   # one time
npx .

npm start does the same.

Without cloning, run npx narze/skillsets. The CLI puts the skills in ~/.skillsets (not a temporary npx folder), so the printed paths stay valid. It uses git clone --depth 1 of the repository that npx installed from (so a fork clones itself), or the repository in package.json if npm does not say, and copies the files from the npx package if git or the network is not there. In a terminal, it shows the clone URL so you can check or change it (empty copies the package instead), and asks before it replaces an existing ~/.skillsets. Without a terminal it asks nothing and replaces the folder. To clone another repository (for example your fork), run npx narze/skillsets --repo owner/repo or set SKILLSETS_REPO. It takes owner/repo or any git URL, asks no URL question, and stops with an error if it cannot clone (it does not copy the package then). Set SKILLSETS_HOME to use another folder. npx . in a clone uses the clone and downloads nothing.

Each tool loads only the providers you choose, for one session. Use an environment variable when the tool has one, for example in an app such as T3 Code that cannot pass CLI options.

Tool	Environment variable	CLI option
Claude Code	CLAUDE_CODE_PLUGIN_DIRS	--add-dir
OpenCode	OPENCODE_CONFIG_CONTENT	None
Codex CLI	None	-c 'plugins={...}'
Pi	None	--skill
Cursor CLI	None	--plugin-dir

Open the section for your tool.

T3 Code

Set these in T3 Code settings, for each provider:

Provider	Field	Value
Claude	Variables	CLAUDE_CODE_PLUGIN_DIRS = /path/to/skillsets/pstack:/path/to/skillsets/9arm
OpenCode	Variables	OPENCODE_CONFIG_CONTENT = {"skills":{"paths":["/path/to/skillsets/pstack/.claude/skills"]}}
Codex	Launch arguments	-c 'plugins={"pstack@skillsets"={enabled=true}}' (run codex-install.sh first)
Pi	Launch arguments	--skill /path/to/skillsets/pstack/.claude/skills --skill /path/to/skillsets/9arm/.claude/skills
Cursor	None	Not supported in T3 Code. See the "Cursor CLI" section.

• Claude: do not use --add-dir in "Launch arguments". T3 Code keeps only one value for each flag.
• OpenCode: leave "Server URL" blank, so T3 Code starts the server with the variable.
• Cursor is not supported in T3 Code. T3 Code runs Cursor through the Cursor SDK in its own process. It has no "Launch arguments", and it reads only CURSOR_API_KEY from "Variables". Cursor loads skills only from your user folders (for example ~/.cursor/skills, ~/.agents/skills, ~/.claude/skills), the project folders, and installed plugins. To use these skills with Cursor, use the Cursor CLI (see the "Cursor CLI" section).

Claude Code

Environment variable

Set CLAUDE_CODE_PLUGIN_DIRS to the provider folders, separated by ::

CLAUDE_CODE_PLUGIN_DIRS="/path/to/skillsets/pstack:/path/to/skillsets/9arm"

Each provider folder is a Claude Code plugin (.claude-plugin/plugin.json). So Claude Code adds the provider name to each skill name, for example /pstack:poteto-help, and names in different providers do not conflict.

CLI option

Add the provider folders to a Claude Code session with --add-dir:

claude --add-dir "/path/to/skillsets/pstack" "/path/to/skillsets/mattpocock" "/path/to/skillsets/9arm"

The skills load only in that session. To load one provider, add only its folder.

• --add-dir accepts more than one path. If you give a prompt, put it before --add-dir, or Claude reads the prompt as a folder: claude "your prompt" --add-dir "/path/to/skillsets/pstack"
• Most pstack skills set disable-model-invocation: true. Claude does not use them by itself. Type the slash command, for example /poteto-help.
• mattpocock and pstack both have tdd and teach, and superpowers and addyosmani both have test-driven-development. With --add-dir, these names conflict, and you can get the skill from the other provider. CLAUDE_CODE_PLUGIN_DIRS adds the provider name, so it has no conflicts.

OpenCode

Set OPENCODE_CONFIG_CONTENT to a config with the .claude/skills folder of each provider:

OPENCODE_CONFIG_CONTENT='{"skills":{"paths":["/path/to/skillsets/pstack/.claude/skills","/path/to/skillsets/9arm/.claude/skills"]}}'

• Give the .claude/skills folder, not the provider folder.
• OpenCode adds this config to your usual config, so your usual skills also load.
• OpenCode v2 uses a background service. The variable must be set for the process that starts the server. T3 Code starts the server itself when its "Server URL" is blank. If you connect to a server that is already running, start that server with the variable.

Codex CLI

Codex has no flag or environment variable that adds a skills folder for one session. --add-dir gives write access only. This repo is a Codex plugin marketplace instead, with one plugin for each provider folder.

Install the plugins one time:

/path/to/skillsets/codex-install.sh

The script adds this repo as the skillsets marketplace, installs each provider as a plugin, and turns the plugins off in ~/.codex/config.toml. So they do not load in a normal session.

Turn on the providers you want for one session with -c:

# One provider
codex -c 'plugins={"pstack@skillsets"={enabled=true}}'

# Some providers
codex -c 'plugins={"pstack@skillsets"={enabled=true},"9arm@skillsets"={enabled=true}}'

# All providers
codex -c 'plugins={"pstack@skillsets"={enabled=true},"mattpocock@skillsets"={enabled=true},"9arm@skillsets"={enabled=true}}'

• Use the inline table form above. The dotted form -c 'plugins."pstack@skillsets".enabled=true' does not turn the plugin on.
• Codex merges this table with config.toml, so your other plugins do not change.
• Codex adds the plugin name to each skill name, for example $pstack:poteto-help. So names in different providers do not conflict.
• Codex does not read disable-model-invocation, so it can use all pstack skills by itself.
• Codex uses a copy of each plugin. After ./fetch.sh, run codex-install.sh again to update the copies.
• The script uses $CODEX_HOME if it is set.

Pi

Give each provider's .claude/skills folder to --skill. You can use --skill more than one time:

pi --skill "/path/to/skillsets/pstack/.claude/skills" --skill "/path/to/skillsets/9arm/.claude/skills"

• Give the .claude/skills folder, not the provider folder. Pi does not look in the hidden .claude folder.
• Pi reads disable-model-invocation, so it does not use most pstack skills by itself. Type /skill:<name>, for example /skill:poteto-help.
• Pi also loads your usual skills. To load only the skills you give with --skill, add --no-skills.

Cursor CLI

Give each provider folder to --plugin-dir. You can use --plugin-dir more than one time:

cursor-agent --plugin-dir "/path/to/skillsets/pstack" --plugin-dir "/path/to/skillsets/9arm"

• Cursor reads the .claude-plugin/plugin.json in each provider folder.
• The Cursor CLI has no environment variable for this. CURSOR_CONFIG_DIR does not change where Cursor looks for user skills.
• This does not work in T3 Code. See the "T3 Code" section.

Layout

<provider>/
├── skills-lock.json          # source and hash of each skill
├── .claude-plugin/plugin.json # Claude Code plugin, made by fetch.sh
├── .codex-plugin/plugin.json # Codex plugin, made by fetch.sh
└── .claude/skills/<skill>/SKILL.md
.claude-plugin/marketplace.json  # Codex marketplace, made by fetch.sh

Claude Code loads skills only from <dir>/.claude/skills/ in a folder added with --add-dir.

Update the skills

./fetch.sh

The script deletes each provider folder and downloads it again, so it also removes skills that the provider removed. Use git status to see the changes.

The Fetch skills GitHub Actions workflow runs fetch.sh every day. If the skills changed, it opens a PR from the fetch-skills branch, or updates the PR if it is already open.

Add a provider

Add one line to fetch.sh, before write_marketplace:

fetch_skills <owner>/<repo>[/<path>] <folder> [skill...] [-skill...]

Without skill names, the script downloads all skills of the provider. Put - before a skill name to leave that skill out, for example fetch_skills cursor/plugins/pstack pstack -statechart. The script stops with an error if the provider has no skill with that name.

Use a new folder name with only lowercase letters, numbers and -. fetch.sh deletes and downloads each provider folder again, so it stops with an error if the name is not valid, or if the folder exists and fetch.sh did not make it (for example lib or test).

Then run ./fetch.sh, add the folder to the "Providers" table, and commit. To use the new provider in Codex, run codex-install.sh again.

Ask an agent to add it

Open an agent (for example Claude Code or Codex) in this repo and send this prompt. Replace the source, folder and skill names:

Add a skills provider to this repo. Follow the rules in CLAUDE.md.
- Source: <owner>/<repo>[/<path>]   (the value for `npx skills add`)
- Folder: <folder>
- Skills: all   (or a list, for example: debug-mantra, post-mortem)
Run ./fetch.sh, check that only the new folder changed, update the provider table in README.md, and commit.

Make your own skillsets

1. Click Use this template at the top of this repo on GitHub, then Create a new repository.
2. Clone your new repo.
3. In fetch.sh, delete the fetch_skills lines that you do not want, and add your providers (see "Add a provider"). Change the owner name narze in write_marketplace to your name.
4. Run ./fetch.sh. It deletes and downloads each provider folder again. Delete the folders of providers that you removed from fetch.sh.
5. Update the "Providers" table in this README, and commit.
6. To let the daily workflow open PRs, go to Settings → Actions → General and select Allow GitHub Actions to create and approve pull requests.

You can also ask an agent to do steps 3 to 5:

Make this repo my own skillsets. Follow the rules in CLAUDE.md.
- Keep these providers: <folders to keep, or none>
- Add these providers: <source, folder and skills for each>
- Marketplace owner name: <your name>
Remove the other providers from fetch.sh and delete their folders, run ./fetch.sh, update the provider table in README.md, and commit.