A MakeUseOf writer recently discovered that Windows’ built-in cleanup tool had left 143 gigabytes of unnecessary files languishing on a nearly full C: drive—until Anthropic’s Claude Code intervened. That’s more than some entire Windows installations consume, and it underscores a persistent frustration for power users: native storage management, for all its polish, still leaves a vast frontier of digital detritus untouched.

Claude Code, an AI coding assistant that runs in the terminal, was not designed for disk janitorial work. But when the writer fed it a simple directive—“Find large, safe-to-delete files I haven’t used in months”—it tapped PowerShell, du, and its own knowledge of common bloat patterns to comb through every corner of the file system. The results were startling, revealing entire virtual machines, obsolete Android emulator images, and swollen AppData folders that Windows Storage Sense either did not see or chose to ignore.

Storage Sense’s blind spots

Storage Sense, introduced in Windows 10 and refined in Windows 11, is designed as a set-it-and-forget-it housekeeper. When disk space runs low, it can automatically delete temporary files, empty the recycle bin, and remove unused local copies of OneDrive documents. Its rule set is intentionally conservative: it targets well-known junk locations like %TEMP%, Prefetch, and old Windows Update backups, but it stops well short of poking into user folders where the truly massive waste piles up.

Microsoft’s design philosophy here is understandable. Windows must protect inexperienced users from accidentally wiping personal photos or critical documents. But that caution creates a growing gap for enthusiasts, developers, and long-time Windows users whose systems accumulate terabytes of invisible waste. Virtual machine hard disks, Docker layers, NuGet caches, old Android SDK images, and forgotten node_modules folders can consume hundreds of gigs without ever tripping Storage Sense’s sensors.

How Claude Code spotted the 143 GB

Claude Code is Anthropic’s agentic coding tool, built on the same Claude family of models that powers enterprise chat interfaces. It lives in a terminal session and can write and execute code, interpret output, and iterate on a problem. The MakeUseOf author installed it on Windows via WSL2, giving it access to the entire NTFS file system through Linux-based utilities and Windows interop commands.

The initial prompt was deceptively straightforward: “Analyze my C: drive and identify all directories larger than 1 GB that have not been modified in the last six months. Suggest which ones are safe to delete.” Claude Code responded by running find and du piped through a series of filters, then cross-referencing the results with known software footprints it has in its training data. It recognized Hyper-V .vhdx files, VirtualBox .vdi images, and Android Virtual Device .img files as virtual machine artifacts. It spotted versions of old Windows SDKs and game engine caches. All were presented in a structured, explained list, with each item rated for risk.

The largest single find was a dormant Hyper-V Windows 10 development VM. The .vhdx file alone measured 62 GB, but the VM’s checkpoint chain and automatic snapshots ballooned the total to 91 GB. The author had last used the VM over a year ago and had multiple newer instances already running. Another 20 GB originated from an abandoned Android emulator image for a project that never materialized. The AVD folder inside C:\Users\[User]\.android\avd is notoriously opaque—Storage Sense never touches it because it resides in the user profile, a directory it treats as sacrosanct.

The AppData bloat that followed was equally enlightening. Within AppData\Local, the tool found 15 GB of stale log files from a now-uninstalled game launcher, cached installers from long-completed Visual Studio updates, and orphaned user profiles for applications that had been uninstalled without proper cleanup. Windows’ own Component Store, accessible through DISM, held 8 GB of superseded updates and driver packages that the normal Disk Cleanup wizard often overlooks unless you manually tick the “Windows Update Cleanup” box. Closing out the 143 GB were scattered remnants: a forgotten Downloads folder stuffed with decade-old ISO images and duplicated temp files from a video editing binge.

The human-AI collaboration model

Crucially, Claude Code never acted as an autonomous deletion bot. Every suggested removal required explicit user confirmation. The author could inspect each file path, verify its relevance, and in several cases, move items to an external drive rather than wipe them. This interactive loop is what distinguishes the approach from traditional one-click “PC cleaner” utilities that often overpromise and occasionally delete the wrong things.

The AI also acted as an educator. When flagging the Android emulator folder, it explained that system-images can be safely re-downloaded if needed later. It reminded the user that Hyper-V checkpoints are not deleted when a VM is removed—a nuance that many even seasoned Windows admins forget. By revealing these hidden relationships, Claude Code transformed a simple cleanup into a learning experience.

Why Storage Sense can’t keep up

Microsoft has steadily improved Storage Sense, but its architecture will always be reactive. It relies on a static whitelist of folders and file types. The company avoids scanning user profile directories because of privacy and safety guardrails. That’s good for a billion consumer devices, but it leaves the knowledgeable minority to fend for themselves.

Third-party tools like WizTree, WinDirStat, and SpaceSniffer have long filled this gap, providing visual treemaps that make it easy to spot large files. But they still require a human to interpret the results and decide what to delete. A novice can stare at a giant blue block labeled AppData and have no idea whether it’s safe to prune. An AI assistant adds a layer of domain expertise, instantly recognizing the digital signatures of known applications and their safe-to-remove caches.

The promise of AI-guided maintenance

This experiment points toward a future where Windows maintenance becomes conversational. Microsoft is already embedding its Copilot assistant across the OS. It’s not hard to imagine a “Clean my PC” button that fires up a local AI model, trained on anonymized disk‑usage patterns, to produce a tailored cleanup proposal. The AI could flag items with confidence scores, explain the impact (e.g., “This will temporarily re-download 2 GB the next time you run Docker”), and schedule deletions for off‑hours.

Such capabilities would be a natural extension of existing Windows features. The Diagnostic Data Viewer already collects telemetry on disk usage. An on-device model could consume that data without ever leaving the machine, preserving privacy. The challenge is ensuring the model never suggests deleting critical system files—a risk that even intelligent models can get wrong. But with proper sandboxing and a mandatory review step, the safety bar could be met.

Risks and precautions

For all the promise, AI-powered cleanup is not without peril. A hallucinating model might confidently assert that System32 is a leftover from an old Windows 7 upgrade. Without robust safeguards, a user could blindly follow the bot into disaster. The Claude Code experiment succeeded because it ran without administrative privileges, required user approval for each action, and was supervised by a technically savvy writer. Replicating that at scale requires building trust and rigorous testing.

Privacy is another concern. Cloud-based assistants would inevitably send file paths and metadata to remote servers, which Microsoft has historically avoided with local-only tools. Keeping the AI local—perhaps leveraging the NPUs in modern Snapdragon X PCs—could resolve that tension while still offloading the cognitive load of manual spring cleaning.

What users can do today

Until native AI cleanup arrives, Windows enthusiasts can mimic the MakeUseOf writer’s approach with a few precautions:

  • Use disk analyzers first: Run WizTree or SpaceSniffer to visualize what’s eating space. Export the list of large directories.
  • Leverage command-line tools with AI assistance: Tools like Claude Code, GitHub Copilot in the terminal, or even ChatGPT with command generation can help you craft precise Get-ChildItem PowerShell scripts that filter by age and size.
  • Always manually review: Never delete anything without understanding what it is. When in doubt, research the folder name online.
  • Back up irreplaceable data: Before any mass cleanup, ensure critical documents, photos, and VM images are safely duplicated.
  • Be especially careful with AppData and ProgramData: These directories contain configuration and state for applications; delete only the ones you’re certain belong to uninstalled software.

Microsoft’s opportunity

The 143 GB find isn’t an outlier—it’s the natural endpoint of a decade of cumulative Windows usage. As storage capacities grow, so does the pile of abandoned artifacts. Microsoft has historically treated disk cleanup as a solved problem, but the proliferation of developer tools, containerization platforms, and virtual machines has changed the game. An intelligent assistant that learns from the collective experience of Windows users could turn monthly “disk full” panics into a thing of the past.

A compelling first step would be to open a Storage Sense API that third-party AI tools can query. By providing a secure, read-only view of the file system with metadata tags for “safe to delete” or “deletable if not used for X days,” Microsoft could foster an ecosystem of smart janitorial agents while still enforcing its safety standards.

Bottom line

The MakeUseOf writer’s experiment with Claude Code demonstrates that even a well-maintained Windows machine can hide a staggering amount of wasted space. It exposes the limitations of built-in tools and points to a smarter, AI-powered path forward. For now, the burden remains on users to hunt down dead weight manually or with half-measures, but the conversation has clearly shifted. The question is no longer whether AI can help us clean our PCs better; it’s when that capability will arrive in every Windows update.