Create Claude Code Taskbar Buttons on Windows 10 & 11 with PowerShell

His approach is useful because it is simple. It also has a few rough edges, so before you build your own row of AI buttons, it helps to know which parts are documented by Anthropic and which parts are his own setup.

The idea: one task, one button​

Ghosh describes a familiar progression. First he typed out long prompts every time. Then he kept a prompt library in Obsidian, which meant opening two apps and copying and pasting between them. Then came Claude Projects and Skills, which store instructions inside Claude. All of these still require opening Claude next to whatever app you're working in.

He says he tried Raycast-style launchers and AutoHotkey scripts before settling on PowerShell scripts pinned to the taskbar. Each button works like this:

• Input: The script reads whatever is on the clipboard, either text or a screenshot.
• Processing: It sends that content, plus a saved prompt, to Claude Code in headless mode.
• Output: Depending on the script, the result goes back to the clipboard, appears in a pop-up window, gets pasted at the text cursor, or is saved as a file.

The engine is Claude Code's print mode. With the -p flag, Claude Code runs its normal agent loop with no chat interface. BuildThisNow's guide notes that it supports --output-format text|json|stream-json, reads stdin so you can pipe data in and redirect out, and pre-approves tools with --allowedTools and --permission-mode so unattended runs never block on a prompt. In practice, claude -p behaves like any Unix filter: text goes in and an answer comes out. That makes it easy to drive from a PowerShell script.

Section summary: Each taskbar button runs one PowerShell script. The script feeds clipboard content and a saved prompt to claude -p and sends the answer wherever you want it.

The four buttons Ghosh actually uses​

Here are his four workflows, which make good starting points:

1. Proofreader. He uses voice typing, which leaves transcription and punctuation errors. He copies the dictated text, clicks the button, and gets a corrected version back on his clipboard.
2. Reply. Many of his messages are standard answers that send people to particular resources. He calls these "if-else statements" in prose. His prompt lists the resources and when to recommend each one, and Claude writes a reply that fits the question.
3. Explainer. He copies an error code or a screenshot. Claude runs a quick web search and shows a floating window with an explanation and sources.
4. Fact Checker. He copies a statement, a tweet, or a screenshot. Claude checks it against what it knows plus web results and shows which parts hold up, with links, in a floating window.

He also mentions Claude's "Record a Skill" feature for teaching Claude a more complex task that a shortcut can then trigger.

Building it: Claude writes the script for you​

Ghosh doesn't write the PowerShell himself. He asks Claude Code or the Claude desktop app, both of which can save files to your PC, to build three things:

• A PowerShell script that reads the clipboard (text or image), calls claude -p, and returns the result to the clipboard or shows it in a window you close yourself.
• A separate prompt file next to the script, so you can change the instructions without editing any code.
• Windows notifications when a job starts, when it finishes, and when something goes wrong.

His prompt also tells Claude to use --system-prompt-file so the button's instructions replace Claude Code's default coding prompt, --safe-mode so his personal Claude Code settings don't interfere, and to turn off any tools the button doesn't need. At the end, Claude should give step-by-step directions for creating the desktop shortcut.

Fact-check: do those flags exist?​

Yes, and this is worth confirming before you paste the prompt. Anthropic's current CLI reference lists both flags:

• --safe-mode starts Claude Code with your customizations turned off: CLAUDE.md, skills, plugins, hooks, MCP servers, custom commands, and auto memory. According to Anthropic, authentication, model selection, built-in tools and permissions still work normally, and that is how it differs from --bare. Note that managed (organization-wide) policy settings still apply, so a work laptop may behave differently.
• --system-prompt-file is one of Anthropic's system prompt flags. It replaces the default prompt with text loaded from a file. Anthropic warns that replacing the prompt also removes the default tool guidance and safety instructions, so "you take responsibility for whatever your task still needs." A proofreading button doesn't need coding conventions. A button that browses the web should have its limits spelled out in your prompt file.

The same reference adds controls that suit a one-click button:

There's a catch with --tools "". Explainer and Fact Checker depend on web search, so locking every tool down will break them. Match the tool list to the job. Spending limits matter for scripts too. Amux's headless guide warns that a retry loop on a bad prompt can spend $100 in an hour. Always set per-invocation or per-day budget guards. That warning is aimed at API billing, but a button you click without thinking is the same kind of risk.

Section summary: Ghosh's flags are real and documented. Disable tools for text-only buttons, keep web access for research buttons, and set a budget cap if you pay per API call.

Prerequisites the article skips​

"Without opening Claude" means without opening the Claude app. You still need the Claude Code engine. Anthropic's setup documentation lists:

• OS: Windows 10 version 1809 or later, or Windows Server 2019 or later
• Hardware: at least 4 GB of RAM, x64 or ARM64
• Network: an internet connection
• Account: Pro, Max, Team, Enterprise or Console. The free Claude.ai plan does not include Claude Code. Third-party providers such as Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry are also supported.

A quick setup check on Windows:

1. Install Claude Code from PowerShell or CMD. You don't need admin rights. One documented option is winget install Anthropic.ClaudeCode. Anthropic notes that WinGet installs don't update themselves, so run winget upgrade Anthropic.ClaudeCode from time to time.
2. Open a new terminal and run claude --version. If you see a version number, the install worked. If Windows says claude isn't recognized, the install folder isn't on your PATH yet.
3. Run claude once and sign in through the browser.
4. If something seems wrong, claude doctor prints install and settings diagnostics without starting a session.

Git for Windows is optional. Without it, Claude Code runs shell commands through PowerShell.

Pinning and prettifying the button​

Ghosh found that Claude could write the script but couldn't create the shortcut or pin it to the taskbar. You do that part yourself:

1. Create a desktop shortcut using the target Claude gives you. As a general Windows convention, not something from his article, these targets usually call powershell.exe with -File and the script path, and often add -WindowStyle Hidden to keep the console out of sight. If Windows blocks the script, check your execution policy.
2. Right-click the shortcut and choose Show more options > Pin to taskbar. That's the path in Windows 11's context menu.
3. Every button starts with the same generic PowerShell icon. Download an icon in .ICO format, then open the shortcut's Properties > Change Icon and select the file so you can tell your buttons apart.

The screenshot caveat​

Text is the easy case. Anthropic's CLI reference describes print-mode input as text or streaming JSON, and none of the documentation reviewed here says you can pipe a raw Windows clipboard image straight into claude -p. Ghosh says his script handles screenshots, but that comes from his own implementation. It is not built-in behavior you can count on. One approach, which is my suggestion and not his documented method, is to have the script save the clipboard image as a temporary file and point the prompt at that file. That requires Claude to have a file-reading tool enabled. Test it on your own install before you rely on it.

Privacy, and whether this beats Copilot​

Ghosh argues that Microsoft's "AI as a first-class citizen" promise for Windows 11 is mostly limited to Copilot+ PCs and still feels poorly integrated, while CLI tools like Claude Code or Codex plus a little PowerShell fit how people actually work. That's opinion, and he has an obvious preference. Copilot is built into the OS and needs no scripting. Claude Code requires a paid account and some upkeep.

Remember that your clipboard often holds passwords, client emails and internal screenshots, and every click sends its contents to a cloud model. Ghosh does mention that local servers such as Ollama or LM Studio can keep data on your machine, but I haven't tested those here. If you have a "Fact Checker" button, check what you've copied before you click it.

Bottom line: This is one of the more practical ways to bring AI into Windows. It's cheap if you already pay for Claude, the flags are documented, and you can change any button by editing a text file. Start with a button that only handles text, keep its tools locked down, set a spending cap, and expect to spend some time getting screenshot handling to work.