Claude Code Finds 30GB of Windows 11 Cleanup Candidates: Why Approval and Backups Matter

That distinction matters. Finding 30GB worth of candidates is not the same as proving that 30GB was safely reclaimed, and Raj’s account does not include a complete deletion log or measured post-cleanup free-space total. Still, the experiment offers a useful model for supervised Windows maintenance: measure first, protect important data, and keep the delete button firmly under human control.

What Claude Code found on the Legion laptop​

Raj describes a 2022 Lenovo Legion 5 Pro running Windows 11 version 26H2, with roughly 213GB occupied on its system drive. Before involving Claude Code, he used WizTree to inspect storage consumption, exported the scan as a CSV, and saved a list of installed applications.

His initial inventory showed more than 110GB under C:\Users, approximately 20GB in the Recycle Bin, an old C:\Temp directory, and a roughly 3GB factory driver folder.

According to XDA, Claude Code then produced a categorized cleanup proposal in about three minutes and 23 seconds. The reported totals were:

These were recommendations for that particular laptop—not a universal Windows debloating checklist.

Among the larger individual discoveries were a roughly 4.5GB PanFone recovery folder associated with an old iPhone recovery operation and a roughly 4.5GB NVIDIA shader cache. Raj also reported a Lenovo factory driver dump, Lenovo Companion, a McAfee upgrade task, Lavasoft Web Companion, and a ReasonLabs antivirus installation with seven services.

The application list reportedly included 23 preinstalled apps, with Clipchamp, Disney+, and Solitaire among the examples.

The account primarily identifies the model as Claude Sonnet 5.5, although two passages call it Opus. That inconsistency prevents a clean model-specific performance claim. More importantly, this was one person’s maintenance session, not a comparative benchmark of AI models or cleanup utilities.

The supported result is a fast candidate audit—not a demonstrated 30GB recovery or a measured performance improvement.

The exclusions were more important than the prompt’s ambition​

Raj did not simply ask Claude to make Windows smaller and hope for the best. His instructions required an audit before changes, a table of proposed removals with sizes and safety assessments, explicit approval before removal, and a record in debloat-log.md.

He also excluded:

• OneDrive.
• System Volume Information and recovery resources.
• His Dev Drive.
• C:\Llama.cpp and C:\Games.
• Model folders and .ckpt checkpoint files.
• The Recycle Bin.

That final exclusion is particularly important when interpreting the headline: the separately reported 20GB in the Recycle Bin was outside the requested cleanup scope. It should not be quietly folded into the 30GB claim.

The broader lesson is that “unused” depends on the owner. An old phone-recovery directory might be disposable on one machine and the only remaining copy of useful material on another. A model checkpoint can look like an enormous, rarely opened file while representing hours of work.

An AI-generated safety rating is therefore a recommendation to investigate, not a permission slip.

Manual approval is not an undo system​

Raj says he switched Claude Code from auto to manual permissions before the cleanup. Anthropic’s user FAQ distinguishes those modes: auto can approve actions after background safety checks, manual asks before risky edits or commands, and plan mode provides read-only planning. The FAQ documents Shift+Tab for cycling permission modes. Manual and read-only are not interchangeable.

For an initial storage investigation, that distinction suggests a more conservative approach: use read-only planning where available, keep the audit explicitly non-destructive, and inspect the active permissions rather than assuming the mode name guarantees a prompt for every file.

There is also a separate recovery limitation. Anthropic’s checkpointing documentation explains that files changed only through shell commands are not tracked for restoration. Its current documentation allows rewind coverage for subsequent shell changes to files already edited through Claude’s file-editing tools, but that is not a general backup of the drive. Do not assume /rewind can recover arbitrary cleanup deletions.

For this kind of work, the practical implication is straightforward: retain independent backups of anything irreplaceable before authorizing destructive changes. A conversation history is not a recovery plan.

Start with Windows’ own cleanup baseline​

The experiment also raises a reasonable question: how much assistance does an AI agent add beyond Windows’ existing tools?

Microsoft documents a built-in review workflow in Windows 11:

1. Open Start > Settings > System > Storage > Cleanup recommendations.
2. Inspect the suggested categories, including temporary files, large or unused files, cloud-synced files, and unused apps.
3. Select only the items you intend to remove.
4. Choose Clean up for the relevant section. Windows displays an estimated space saving before cleanup.

That gives users a baseline before granting another tool access to their files. It does not establish that Windows would have found every item on Raj’s laptop, but neither does XDA’s experiment establish that conventional cleanup tools would have missed them.

For unwanted applications, Microsoft’s documented route is Settings > Apps > Installed apps, followed by the application’s menu and Uninstall. Startup management is a different task: Task Manager > Startup apps lets users disable unnecessary automatic launches without removing the application. Separating those decisions avoids treating every startup entry as an uninstall candidate.

Automatic cleanup needs its own review, too. Microsoft says Storage Sense can remove temporary files and age out Recycle Bin contents; Downloads and cloud content are not managed under its default settings unless configured. Anyone preserving the Recycle Bin as Raj did should check those settings before enabling automation.

A useful assistant, not a universal debloater​

The strongest interpretation of this experiment is not that Claude knows which files every Windows user needs. It is that a supervised agent may help turn a messy storage inventory into a manageable review queue.

Raj supplied boundaries, inspected recommendations, and retained approval authority. Those choices—not merely the model’s speed—made the exercise worth examining.

For readers attempting something similar, success should mean more than a large proposed total: record free space before and after, retain the approved-item log, and confirm that required applications and workflows still function. Without those checks, a tidy recommendation table remains just that—a recommendation table.

The promising role for AI here is storage analyst with supervision, not unattended janitor. On a working PC, the best cleanup is the one that leaves both more space and everything you still need.