AI Agent Skills
whobat/AI-Agent-skills
Skills portáteis reutilizáveis para agentes de código múltiplos
Ver no GitHubColeção portável de skills reutilizáveis para agentes de código (Claude, OpenAI, Gemini, Cursor, etc). Cada skill é uma pasta com instruções em Markdown e scripts opcionais, seguindo especificação padronizada que funciona em múltiplos agentes. Inclui skills para análise de logs, relatórios, rastreamento de tempo e mais.
Instalação
npx convez-ai whobat/AI-Agent-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Estender capacidades de agentes de código (Claude Code, Cursor, OpenCode)
- Automatizar tarefas repetitivas em PowerShell, Python e Node.js
- Rastreamento de tempo e análise de eventos do sistema
- Geração de relatórios com branding customizado
Veredito

Projeto bem estruturado com especificação clara e CI/CD robusto. Ideal para quem quer estender agentes de código, mas requer conhecimento de múltiplas linguagens (PowerShell, Python, Node.js) e gerenciamento de dependências. Documentação sólida, mas adoção depende do suporte do seu agente específico.
README
A tradução para o português deste README ainda está na fila.
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
| Requirement | When it's needed | Notes |
|---|---|---|
| Node.js 18+ | Only for the npx installer (bin/cli.js). | Not needed if you use install.ps1 / install.sh or copy folders manually. |
| Python 3.8+ | Any skill that ships .py scripts (currently 7pace-time-tracker). | All three installers auto-detect Python and offer to install it when a Python-based skill is selected — via winget (Windows), brew (macOS), or apt-get/dnf (Linux). If no package manager is found, install manually from python.org. The installers also pip install each skill's Python packages (declared in skill.install.json). |
| PowerShell 7+ | Any skill that declares it as a requirement (currently win-eventlog-triage and nav2009-sql-performance). | All three installers auto-install it when such a skill is selected: winget install Microsoft.PowerShell on Windows (falling back to downloading + silently running the latest MSI from GitHub if winget is absent), brew install --cask powershell on macOS. See Runtime requirements. |
| A supported package manager | Only for the auto-install above. | winget / brew / apt-get / dnf. Without one, install the runtime yourself, then re-run. |
After an auto-install, Python may not be on
PATHfor the current terminal session — open a new terminal (or re-run the installer) so the freshly installedpythonis found.
Per-skill
| Skill | Runtime | Python packages | Credentials |
|---|---|---|---|
| branded-report | Python 3.8+ | markdown, python-docx, beautifulsoup4, pillow, pypdfium2, pypdf (installed automatically) | A local theme.json (brand colors/fonts/logo). PDF additionally needs a headless Chrome/Edge/Chromium (warn-only). |
| 7pace-time-tracker | Python 3.8+ | requests (installed automatically) | config.json with a 7pace Bearer token (worklog CRUD) + optional Azure DevOps PAT (work-item search). See Configuration & secrets. |
| win-eventlog-triage | PowerShell 7+ (auto-installed) | — | A tier-admin credential, prompted each run and never stored. WinRM must be enabled on the target servers. |
| rds-profile-triage | PowerShell 7+ (auto-installed) | — | A tier-admin credential, prompted each run and never stored. WinRM (or DCOM/WMI for the dead-WinRM fallback) on the target RDS hosts. Read-only. |
| nav2009-development | — (knowledge-only) | — | — |
| nav2009-sql-performance | PowerShell 7+ (auto-installed) | — | Windows integrated auth by default (or -SqlCredential, prompted). Needs VIEW SERVER STATE + VIEW DATABASE STATE on the SQL Server. Read-only. |
| nav2009-object-management | NAV 2009 Classic client (finsql.exe — from a NAV install, not auto-installable; warned if missing) | — | Manual Object Designer runbook (NAV 2009 has no object CLI — that arrived in NAV 2013). Developer license for .txt/compile; end-user license suffices for .fob import. SQL read access for verification. |
| nav2009-service-tier-admin | PowerShell 7+ (auto-installed) | — | Local admin to read service config / restart; remote needs CIM/WinRM + admin on the target. Inventory is read-only; restart is opt-in. |
| nav2009-db-maintenance | PowerShell 7+ (auto-installed) | — | Windows integrated auth by default (or -SqlCredential, prompted). db_backupoperator/db_owner (backup), ALTER (index), db_owner/sysadmin (CHECKDB). Defaults to a dry run. |
| nav2009-permissions-security | — (knowledge-only) | — | — |
| nav2009-troubleshooting | — (knowledge-only) | — | — |
| ai-agent-skills-update | Node.js 18+ (for the npx path; warn-only) | — | — |
| sqlserver-perf-triage | PowerShell 7+ (auto-installed) | — | Windows integrated auth (or -SqlCredential, prompted). VIEW SERVER STATE + VIEW DATABASE STATE. Read-only. |
| ax2012-sql-performance | PowerShell 7+ (auto-installed) | — | Windows integrated auth (or -SqlCredential, prompted). VIEW SERVER STATE + VIEW DATABASE STATE. Read-only. |
| retail-pos-fleet-health | PowerShell 7+ (auto-installed) | — | Admin credential on the targets, prompted each run. WinRM on the targets. Read-only. |
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
├── 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
| Skill | What it does | Extra setup |
|---|---|---|
| branded-report | Standardized branded documents: turns a Markdown report into HTML + PDF + DOCX from one source and one theme, so every report looks identical. The theme (brand colors/fonts/logo) can be auto-extracted from a PowerPoint/Office template (.pptx/.docx/.xlsx/…) or sampled from an image, website, or PDF. | Python 3.8+, pip install markdown python-docx beautifulsoup4 pillow pypdfium2 pypdf, a local theme.json; PDF needs a headless browser. |
| 7pace-time-tracker | 7pace Timetracker (Azure DevOps): create/edit/delete time entries via REST API, plus free-text work-item search. | Python 3.8+, pip install requests, and a config.json (see Configuration & secrets). |
| win-eventlog-triage | Windows Event Log triage: pulls Critical/Error events from one or many servers in parallel over WinRM, groups them, and returns JSON the agent turns into a critical-first summary. | PowerShell 7+ (auto-installed by the installer), WinRM on the targets, and a tier-admin credential (prompted each run — nothing stored). |
| rds-profile-triage | RDS session-host profile & WinRM-host triage: read-only health of terminal-server profiles over WinRM (CIM/DCOM fallback when the WinRM host won't even launch) — profile-hive leaks, temp-profile sprawl, ProfileList corruption (the "Element not found" chain that kills WinRM), the roaming-profile path read RAW, drain state, and User-Profiles-Service failures grouped by the actual user. Built to dodge the REG_EXPAND_SZ and double-hop misdiagnosis traps. | PowerShell 7+ (auto-installed), WinRM (or DCOM/WMI) on the targets, a tier-admin credential (prompted each run). |
| nav2009-development | Dynamics NAV 2009 / C/AL development: coding patterns, key/SIFT design, review checklist, customization architecture, reports, integrations, and BC-upgrade posture. Knowledge-only (no scripts). | None. |
| nav2009-sql-performance | NAV 2009 SQL performance triage: read-only DMV snapshot of the SQL Server behind a NAV 2009 database (top queries, waits, blocking, deadlocks, missing/unused indexes, SIFT views, fragmentation, stale stats) as JSON, plus a NAV-specific interpretation guide for the agent. | PowerShell 7+ (auto-installed), VIEW SERVER STATE/VIEW DATABASE STATE on the SQL Server. |
| nav2009-object-management | NAV 2009 object deployment runbook: guides export/import/compile of C/AL objects (.fob/.txt) through the Classic client's Object Designer — NAV 2009 has no object CLI (finsql command= arrived in NAV 2013) — and verifies results via a read-only SQL query against the Object table. Knowledge-only. | NAV 2009 Classic client with an appropriate license on the machine doing the manual steps. |
| nav2009-service-tier-admin | NAV 2009 Service Tier / NAS admin: inventories NST instances (status, account, ports, parsed CustomSettings.config, NAS/Job Queue config) as JSON; optional opt-in restart in the correct order. Local or via -ComputerName. | PowerShell 7+ (auto-installed); local admin (or remote CIM/WinRM). |
| nav2009-db-maintenance | NAV 2009 SQL maintenance (action side of the perf skill): backup, DBCC CHECKDB, rebuild/reorganize NAV-owned indexes, and update statistics — defaulting to a dry run that prints the exact T-SQL. Never creates/drops indexes (NAV owns them). | PowerShell 7+ (auto-installed), SQL maintenance permissions. |
| nav2009-permissions-security | NAV 2009 security & permissions: Roles/permission sets, object permissions (incl. indirect), security filters, Windows vs Database logins, NAV↔SQL synchronization and Standard vs Enhanced models, and the license-vs-permission distinction. Knowledge-only. | None. |
| nav2009-troubleshooting | NAV 2009 triage runbook: maps a reported symptom (RTC won't connect, Service Tier won't start, login/permission/license errors, posting/locking failures, NAS/Job Queue stopped, deployment/compile errors, crashes) to a likely cause and routes to the right NAV 2009 skill. Knowledge-only. | None. |
| sqlserver-perf-triage | Generic SQL Server performance triage (any instance/database): read-only DMV snapshot — top queries, waits, blocking, deadlocks, missing/unused indexes, fragmentation, stale stats, config — as JSON, with a generic interpretation guide. | PowerShell 7+ (auto-installed), VIEW SERVER STATE/VIEW DATABASE STATE. |
| ax2012-sql-performance | Dynamics AX 2012 (R3) SQL performance triage: the same read-only snapshot, interpreted through an AX lens — known bloat tables and their cleanup routines, RCSI/TF4136/MAXDOP posture, AOT-owned indexes, and Retail channel DB / CDX sync health. | PowerShell 7+ (auto-installed), VIEW SERVER STATE/VIEW DATABASE STATE. |
| retail-pos-fleet-health | POS / Windows fleet health sweep over WinRM: stopped auto-start services, local SQL instances with per-DB sizes (flags databases nearing the SQL Express 10 GB limit), low disk, recent error counts — ranked warnings as JSON. Read-only. | PowerShell 7+ (auto-installed), WinRM on the targets, admin credential (prompted). |
| ai-agent-skills-update | Update skills installed from THIS repo (and only this repo — other skills are never touched): runs the installer in --update mode across all agents, refreshes every installed repo skill to the latest published version, installs nothing new, preserves local config.json. Triggers on "update my AI-Agent-skills". | Node.js 18+ for the npx path (warned if missing; clone + install.ps1/install.sh -Update works without). |
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.
| Agent | Personal/global skills dir | Project-level |
|---|---|---|
| Claude Code | ~/.claude/skills/<name>/ | .claude/skills/<name>/ |
| OpenAI Codex | ~/.agents/skills/<name>/ | <repo>/.agents/skills/<name>/ |
| OpenCode | ~/.config/opencode/skills/<name>/ (also reads ~/.claude/skills/ and ~/.agents/skills/) | .opencode/skills/<name>/ |
| Gemini CLI / Cursor / Cline / Windsurf | place under their skills dir, or reuse ~/.claude/skills/ or ~/.agents/skills/ | project .agents/skills/ |
| Any other agent | If it supports the SKILL.md standard, drop the folder in its skills dir. If not, the bundled Python script still works standalone — call it from the shell (see below). |
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: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:
- SKILL.md as instructions — paste the
SKILL.mdbody into the agent's rules/instructions file (e.g.AGENTS.md, a system prompt, or a custom command). - Script is agent-agnostic —
skills/7pace-time-tracker/scripts/timetracker.pyis 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>; ifwingetis missing, the installer downloads the latestwin-<arch>.msifromgithub.com/<githubRepo>and runs it silently (msiexec /quiet). - macOS:
brew install --cask <brewCask>. - Linux / no package manager: prints the
urlto 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:
- Spec-compliant frontmatter only (
namematching the folder,description≤1024 chars with when-to-use triggers,license: MIT,metadata.version, optionalcompatibility).npm run validateenforces this. - Put any code in
scripts/(with tests next to it), reference docs inREFERENCE.md, and aconfig.example.jsonif it needs secrets. - If it needs a runtime (Python packages, PowerShell 7, …), declare it in
skill.install.json(see Runtime requirements). - Add 3+ evaluation scenarios in
evals/<name>.jsonand a row to Available skills above. Bumpmetadata.versionwhenever you change the skill so installed copies are offered the update. Commit.