AI-Agent-Skills

A shared collection of Agent Skills — portable SKILL.md-based capabilities that work across multiple coding agents (Claude Code, OpenAI Codex, OpenCode, Gemini CLI, Cursor, and more).

A skill is just a folder containing a SKILL.md (YAML frontmatter + Markdown instructions) plus optional scripts/ and references/. Every skill in this repo conforms to the Agent Skills specification — only spec-defined frontmatter fields (name, description, license, compatibility, metadata), validated in CI by npm run validate — so the same folder installs into any supporting agent. Licensed under MIT; contributions welcome, see CONTRIBUTING.md.

Requirements

What you need depends on which skills you install. The skill itself (SKILL.md) is just text and needs nothing — the requirements come from any scripts a skill ships and the credentials it talks to.

General

After an auto-install, Python may not be on PATH for the current terminal session — open a new terminal (or re-run the installer) so the freshly installed python is found.

Per-skill

Repository layout

Skills are grouped into category folders for overview (skills/<category>/<name>/). The grouping is repo-side only — every installer discovers skills by their SKILL.md and installs them flat into the agent's skills dir, so an agent still sees one namespace (~/.claude/skills/<name>/). Each category folder is listed in .claude-plugin/plugin.json so the Claude Code marketplace finds them too.

AI-Agent-skills/
├── README.md
├── .gitignore                 # ignores real config.json / secrets
├── package.json               # enables: npx github:whobat/AI-Agent-skills
├── bin/cli.js                 # npx installer (interactive + flags)
├── install.ps1                # script installer (Windows PowerShell)
├── install.sh                 # script installer (macOS/Linux/Git-Bash)
├── .claude-plugin/            # marketplace.json + plugin.json (lists the category folders)
└── skills/
    ├── nav-2009/              # category folder (7 NAV 2009 skills)
    ├── ax-retail/             # Dynamics AX 2012 + Retail POS
    ├── edi/                   # EDI / e-document exchange (Peppol BIS 3.0 validation, …)
    ├── sql-server/ · windows-ops/ · documents/ · time-tracking/ · repo-tooling/
    └── time-tracking/
        └── 7pace-time-tracker/   # one folder per skill; name == this folder
            ├── SKILL.md            # the skill manifest + instructions
            ├── REFERENCE.md        # detailed reference (optional)
            ├── skill.install.json  # optional: runtime requirements, pip packages, credential-setup command
            ├── config.example.json # template — copy to config.json (gitignored)
            └── scripts/
                ├── timetracker.py
                └── test_timetracker.py

Available skills

Install as a Claude Code plugin (marketplace)

Claude Code users can install everything through the native plugin system — updates then arrive automatically with every push to this repo:

/plugin marketplace add whobat/AI-Agent-skills
/plugin install ai-agent-skills@ai-agent-skills

The plugin exposes all skills in skills/. For Codex/OpenCode (or if you want the runtime auto-install and credential setup), use the npx installer below instead.

Install with npx (recommended)

No clone needed — run it straight from GitHub:

# Interactive (pick agent + skill, optional credential setup)
npx github:whobat/AI-Agent-skills

# Non-interactive
npx github:whobat/AI-Agent-skills --agent claude --skill all
npx github:whobat/AI-Agent-skills --agent codex  --skill 7pace-time-tracker --auth

Flags: --agent claude|codex|opencode|all (all = every agent in one run) · --skill all|<name> · --update (refresh installed skills to the latest repo versions — installs nothing new) · --auth (run a skill's credential setup after install) · --symlink · --list (add --agent to see installed vs latest versions) · -y/--yes · -h.

# Update everything you have installed, across all agents, in one go:
npx -y github:whobat/AI-Agent-skills --update --agent all --yes

(The ai-agent-skills-update skill wraps exactly this — say "update my AI-Agent-skills" to your agent. It only touches skills from this repo.)

The installer copies skill folders into the agent's skills dir and never copies real `config.json` (only config.example.json). With --auth (or when prompted), it runs the skill's credential setup so your tokens are entered securely and saved locally — see Configuration & secrets. If a skill's credentials are already configured (its configPath file exists, e.g. ~/.7pace/config.json), the installer detects it and asks whether to update the tokens or keep them (keep is the default) instead of prompting for setup from scratch.

Install with the bundled scripts (no Node)

If you've cloned the repo and prefer not to use Node:

Windows (PowerShell):

./install.ps1 -Agent claude -Skill all
./install.ps1 -Agent codex -Skill 7pace-time-tracker

macOS / Linux / Git-Bash:

./install.sh --agent claude --skill all
./install.sh --agent codex --skill 7pace-time-tracker

-Agent/--agent: claude · codex · opencode · all (every agent in one run). -Skill/--skill: all or a skill folder name. Add -Update/--update to refresh installed skills only (installs nothing new; run git pull first — the comparison is against your clone), -Symlink/--symlink to link instead of copy, or -Yes/--yes to auto-apply runtime installs and skill updates without prompting.

Manual install (per agent)

A skill = the folder skills/<category>/<name>/. Copy (or symlink) that folder into the agent's skills directory as `<name>/` (drop the category — agents want a flat skills dir). The npx/script installers do this flattening for you; the commands below are the manual equivalent.

Tip — one copy, many agents: ~/.agents/skills/ is read by Codex and OpenCode, and Claude Code reads ~/.claude/skills/. Symlinking a skill into both covers most setups: ``bash ln -s "$PWD/skills/time-tracking/7pace-time-tracker" ~/.claude/skills/7pace-time-tracker ln -s "$PWD/skills/time-tracking/7pace-time-tracker" ~/.agents/skills/7pace-time-tracker ``

Examples (manual copy) — note the skills/*/* glob, which flattens the category folders:

# Claude Code — all skills (flattened from their category folders)
cp -r skills/*/* ~/.claude/skills/
# Codex — single skill (drop the category in the destination)
cp -r skills/time-tracking/7pace-time-tracker ~/.agents/skills/7pace-time-tracker

# Claude Code — all skills (PowerShell)
Copy-Item skills/*/* "$HOME/.claude/skills/" -Recurse

"Pi" / other agents

There is no single standard for every agent. Two fallbacks that always work:

1. SKILL.md as instructions — paste the SKILL.md body into the agent's rules/instructions file (e.g. AGENTS.md, a system prompt, or a custom command).
2. Script is agent-agnostic — skills/7pace-time-tracker/scripts/timetracker.py is a plain Python CLI. Any agent (or you) can run it directly; see its --help.

Configuration & secrets

Skills that need credentials ship a config.example.json. Never commit real tokens — .gitignore already excludes config.json (and .env, etc.).

Easiest — guided setup (recommended): the installer's --auth runs it for you, or run it directly:

python skills/7pace-time-tracker/scripts/timetracker.py --auth

It prompts (tokens hidden) and writes ~/.7pace/config.json (chmod 600). For 7pace-time-tracker that's the 7pace Bearer token (worklog CRUD) and an optional Azure DevOps PAT (work-item search) — two separate systems, both kept out of git.

Manual alternative:

cp skills/7pace-time-tracker/config.example.json ~/.7pace/config.json   # then edit with your tokens

The script reads ~/.7pace/config.json by default; override with --config <path>.

Verifying a skill (tests)

Skills that bundle scripts may include tests. CI (GitHub Actions) runs everything below on Windows and Linux for every push/PR, plus npm run validate, which checks every skill against the Agent Skills specification. evals/<skill>.json holds per-skill evaluation scenarios (query + expected behavior) for manual/LLM-driven behavior testing.

Installer tests live in tests/installers/ — one suite per installer. They are hermetic: each test runs the installer against a throwaway repo copy with fixture skills and a fake HOME, so your real skills dirs, configs, and network are never touched. They cover install/all, secret stripping (config.json/__pycache__), version updates (incl. preserving an installed config.json), error exits, and the credential detection (configured → update-or-keep / "already configured").

npm test                                                    # npx installer (Node 18+)
bash tests/installers/test_install_sh.sh                    # install.sh (needs bash 4+)

Invoke-Pester -Path ./tests/installers/Install.Tests.ps1 -Output Detailed   # install.ps1

# Python (7pace-time-tracker)
cd skills/7pace-time-tracker/scripts
python -m unittest test_timetracker -v                 # unit tests, no network
RUN_INTEGRATION=1 python -m unittest test_timetracker -v   # + live API (needs config.json)

# PowerShell / Pester — every script-backed skill ships a *.Tests.ps1 next to its script
Install-Module Pester -MinimumVersion 5.0.0 -Scope CurrentUser
Invoke-Pester -Path (Get-ChildItem skills -Recurse -Filter '*.Tests.ps1').FullName -Output Detailed

Runtime requirements (skill.install.json)

A skill can declare runtimes it needs in skill.install.json. All three installers ensure each requirement is present (and ≥ minVersion when detectable) before finishing, auto-installing any that are missing:

{
  "requirements": [
    {
      "name": "PowerShell 7",
      "detect": "pwsh",
      "minVersion": "7.0",
      "wingetId": "Microsoft.PowerShell",
      "brewCask": "powershell",
      "githubRepo": "PowerShell/PowerShell",
      "url": "https://github.com/PowerShell/PowerShell/releases/latest"
    }
  ]
}

• Windows: winget install <wingetId>; if winget is missing, the installer

downloads the latest win-<arch>.msi from github.com/<githubRepo> and runs it silently (msiexec /quiet).

• macOS: brew install --cask <brewCask>.
• Linux / no package manager: prints the url to install manually.

Requirements that cannot be auto-installed (e.g. NAV 2009's finsql.exe, which only ships with the NAV Classic client) instead declare detectPaths (file locations to check — the installer also accepts these as present) and "warnOnly": true plus a help text: the installer then prints a warning with the checked paths and guidance and continues — the skill still installs, since the tool may live on another machine or a custom path (the script's -FinSqlPath overrides detection at runtime).

If a runtime is installed but not yet on PATH for the current shell, open a new terminal. (The same file may also declare pipPackages, authCommand, authHelp, and configPath — the location of the skill's credential file, e.g. ~/.7pace/config.json.)

Updating installed skills

Each SKILL.md carries a version under metadata: version: in its frontmatter (per the Agent Skills spec; legacy top-level version: in already-installed copies is still recognized). On any install run, the installer compares the installed version of every other skill in the target agent's directory against the repo version, and offers to update the ones that differ (showing old -> new). Updates re-copy the skill files while preserving any `config.json` inside the skill folder. Pass -Yes (PowerShell) / --yes / -y (bash, npx) to apply updates non-interactively; omit it to be prompted.

So after you change a skill, bump its metadata.version and re-run the installer — anyone who already has it installed is offered the update automatically.

Adding a new skill

Copy template/skill-name/ to skills/<name>/ and follow CONTRIBUTING.md. In short:

1. Spec-compliant frontmatter only (name matching the folder, description ≤1024 chars with when-to-use triggers, license: MIT, metadata.version, optional compatibility). npm run validate enforces this.
2. Put any code in scripts/ (with tests next to it), reference docs in REFERENCE.md, and a config.example.json if it needs secrets.
3. If it needs a runtime (Python packages, PowerShell 7, …), declare it in skill.install.json (see Runtime requirements).
4. Add 3+ evaluation scenarios in evals/<name>.json and a row to Available skills above. Bump metadata.version whenever you change the skill so installed copies are offered the update. Commit.

Sources

• Agent Skills — the open standard
• Codex Agent Skills
• OpenCode Agent Skills
• Agent Skills in the Claude Code SDK