PowerShell + Windows CLI Agent Skill

The missing AI agent skill for Windows. Most coding agents are trained on Linux/bash patterns and produce broken PowerShell, unsafe CMD commands, or treat Windows like a second-class citizen. This skill fixes that.

What problem this solves

AI agents (Claude, Kimi, Cursor, Copilot, etc.) often fail on Windows because they:

• Generate bash idioms that do not exist in PowerShell or CMD.
• Confuse PowerShell's object pipeline with Unix text pipelines.
• Miss Windows-specific concepts: registry, services, event logs, WMI/CIM, UAC, execution policy.
• Produce destructive commands without -WhatIf or elevation warnings.
• Use deprecated aliases like ls, rm, cat, sort in production scripts.

This skill gives agents a structured, safety-first guide to writing correct PowerShell 7 / Windows PowerShell 5.1 and CMD/Batch commands.

Keywords

powershell agent, windows cli agent, ai agent windows, powershell skill, cmd skill, batch ai, windows automation agent, agent skill powershell, powershell vs cmd, windows command line ai, powershell static checker, powershell code generator, windows system administration ai, registry ai, wmi cim agent, active directory ai, event log ai.

Features

• 📘 Comprehensive SKILL.md — frontmatter-triggered guide for agents with explicit scope, definitions, and execution-context guidance.
• 🛡️ Safety-first patterns — -WhatIf, elevation warnings, execution policy guidance, and banned permissive policies.
• ⚡ Command templates — generate common PowerShell/CMD snippets from intent with input validation and escaping.
• 🔍 Static checker — catch dangerous cmdlets, deprecated aliases/WMI, risky execution policies, and missing error handling.
• 📚 Deep references — registry, services, networking, WMI/CIM, Active Directory, pitfalls, PowerShell vs CMD, and bash-to-PowerShell Rosetta stone.
• 🧪 Evals — 20+ regression tests covering safety, format, CMD, and negative cases.
• 🤖 Agent-native — designed to be consumed by AI agents via SKILL.md and AGENTS.md.
• 🔄 Autoresearch-ready — includes the Autoresearch skill/command files so Claude Code/Codex can iterate on the skill itself.

Installation

As an agent skill

Copy or symlink this directory into your agent's skill path:

# Example for Kimi / kimi-code
ln -s "$(pwd)" ~/.agents/skills/powershell-windows-cli

As a GitHub project

git clone https://github.com/UncertaintyDeterminesYou4ndMe/powershell-windows-cli-agent-skill.git

Quick start

Validate PowerShell code

python3 scripts/validate_ps.py -c "Get-Process | Stop-Process"
# Exit code 2: Stop-Process is destructive without -WhatIf

python3 scripts/validate_ps.py -c "Get-Process | Stop-Process -WhatIf"
# Exit code 0

Generate command templates

python3 scripts/generate_template.py list-services --status Stopped --start-type Automatic

python3 scripts/generate_template.py query-events --log System --level Error --hours 24

python3 scripts/generate_template.py test-port --host example.com --port 443 --json

Project structure

powershell-windows-cli-agent-skill/
├── SKILL.md                              # Agent-facing main guide
├── AGENTS.md                             # How agents should use this repo
├── README.md                             # Human-facing project overview
├── LICENSE                               # MIT
├── autoresearch/                         # Autoresearch probe report and outputs
│   └── probe-report.md
├── .claude/                              # Autoresearch skill/commands for Claude Code
│   ├── skills/autoresearch/
│   └── commands/autoresearch/
├── scripts/
│   ├── validate_ps.py                    # Static PowerShell checker
│   └── generate_template.py              # Command template generator
├── references/
│   ├── powershell-vs-cmd.md              # When to use which + translation guide
│   ├── bash-to-powershell.md             # Linux/bash to PowerShell/CMD Rosetta stone
│   ├── registry.md                       # Registry operations
│   ├── services-processes.md             # Services, processes, scheduled tasks
│   ├── wmi-cim.md                        # Modern CIM vs legacy WMI
│   ├── networking.md                     # Network diagnostics and firewall
│   ├── active-directory.md               # AD users, groups, computers
│   └── common-pitfalls.md                # Frequent mistakes and fixes
└── evals/
    └── evals.json                        # 20+ regression tests

Why this matters for AI agents

Most training data for coding assistants is Linux-centric. When users ask:

"How do I list stopped services on Windows?"

Agents often answer with systemctl or fragile bash-like PowerShell. This skill teaches agents to answer with:

Get-Service | Where-Object { $_.Status -eq 'Stopped' }

And to explain that Windows services are objects, not text, and that restarting them may require elevation.

Contributing

Issues and PRs are welcome. Focus areas:

• More command templates
• Better static analysis via PowerShell AST
• Additional references (IIS, Exchange, WinGet, WSL, Intune)
• Eval cases for regression testing

License

MIT — see LICENSE.