Reclaim disk space with an AI agent news update

Developer Rıza Sabuncu launched ai-disk-cleanup, an open-source prompt framework that transforms AI coding agents like Claude Code, Cursor, and Codex into disciplined disk auditors across macOS, Linux, and Windows, as detailed in github.com's report. The project provides structured prompts in seven languages that inspect storage top-down, categorize findings by deletion risk, and require user confirmation before executing removals.

This article examines the tiered auditing model, privilege limits, and benchmark results that cleared 111 GB on a 494 GB macOS disk at 99 percent capacity. It is written for software engineers seeking to eliminate bloat from old runtimes, container images, and dead worktrees without accidentally deleting critical project files.

Reclaim disk space with an AI agent news: what actually changed

The project introduces localized prompt files in English, Turkish, Simplified Chinese, Spanish, Russian, Brazilian Portuguese, and Japanese inside the prompts directory. The English prompt, disk-cleanup.en.md, serves as the authoritative specification, though all shell commands remain in English across translations. Instead of relying on manual du and sort commands that misinterpret storage metrics, the prompt instructs agents to detect the host operating system and traverse directories top-down. This avoids traps where allocated sizes deceive users, such as a Docker.raw file reporting 60 GB that actually consumes 14 GB on disk.

Standard cleanup checklists overlook heavy artifacts outside user caches, leaving massive directories undetected. The audit targets hidden accumulations, including a 23 GB directory of dead git worktrees from an abandoned refactoring branch and two 15 GB iOS simulator runtimes located in the system Library directory that standard home scans ignore. By using specific cleanup commands native to each tool rather than blunt rm -rf invocations, the agent prevents broken dependency trees. The top-down audit navigates directly to root offenders, avoiding misidentified targets and ensuring accurate measurement before any deletion step.

Reclaim disk space with an AI agent news: how it works

Users deploy the tool by copying the prompt text beneath the divider line into an agent session. The auditor scans top-down, checking for storage anomalies like sparse disk images, double-counted symlinks, operating system snapshots holding reclaimed capacity, asynchronous deletions, and non-Rust directories named target. After measuring real disk consumption, the assistant categorizes every candidate into a three-tier risk table and halts immediately. It requires explicit user approval before running any destructive commands, ensuring human verification remains in control before files are wiped.

The framework separates data into three risk levels: Tier 1 covers regenerable caches, build outputs, package downloads, and dead worktrees. Tier 2 contains virtual machine disks, container images, and outdated SDK runtimes where internal data disappears permanently. Tier 3 includes downloads, browser profiles, and unused applications requiring manual discretion. When approved, the agent executes cleanups step by step using dedicated subsystem commands rather than generic deletion scripts. It reports newly freed space after each action and delivers a closing summary specifying tasks requiring manual developer completion.

Reclaim disk space with an AI agent news: why it matters now

A real-world test on a 494 GB macOS volume at 99 percent utilization demonstrated how effective the methodology is in practice. Starting with just 5.2 GB available, the audit reclaimed 111 GB, expanding free storage to 117 GB and lowering disk usage to 71 percent. The run eliminated 4.9 GB of dead Claude and git worktrees in a single project, 5.7 GB from Rust target folders, 18 GB across 83 node_modules and .next caches, and reduced ~/Library/Caches from 17 GB to 2.4 GB. Another 16 GB was cleared from ~/.cache, npm, Go modcache, and Homebrew components.

The process freed another 36 GB by deleting obsolete iOS 17.5 and 18.0 simulator runtimes, 14 GB from a Docker Desktop VM disk, and 11 GB from an OrbStack VM disk. Stale downloads accounted for 14 GB, comprising an 11 GB firmware file and 42 installers. Privilege boundaries were strictly enforced when encountering root-owned files: a 10 GB simulator dyld cache and a 6 GB dead Nessus install were not deleted automatically, but provided to the developer as ready-to-run sudo one-liners to execute manually.

Reclaim disk space with an AI agent news: who is affected

This workflow targets engineers running coding assistants like Claude Code, Cursor, and Codex on macOS, Linux, and Windows machines. Multi-language development environments routinely accumulate unindexed dependencies, container images, and build files across disparate directories. Because tools like npm, Go, Cargo, and Docker do not automatically purge old data, drives quietly run out of space during active coding. The prompt provides a structured, platform-aware audit that surfaces orphaned worktrees, obsolete toolchains, and hidden caches that normal system maintenance ignores across diverse developer environments.

Operational safety depends directly on running the agent inside an interactive confirmation mode rather than an automated setting. The documentation warns that while the prompt commands the agent to pause for review, only the host environment enforces that boundary. Developers must use Claude Code's standard permission prompts or Cursor's ask mode instead of unattended yolo modes. Running an agent without permission prompts risks premature file deletion before the user can evaluate the tier table, making human confirmation an indispensable requirement during execution.

Reclaim disk space with an AI agent news: what to watch

Users must recognize that all removals performed by the audit are irreversible because commands bypass the operating system trash bin. Developers should create reliable backups before initiating cleanup, especially prior to approving Tier 2 or Tier 3 actions involving virtual disks or personal downloads. The prompt strictly protects vital files, explicitly forbidding the deletion of source code, git history, .env environment variables, cryptographic keys, system keychains, photo stores, mail archives, cloud-sync folders, active virtual machine disks, and operating system snapshots. Any unintended deletion represents a software bug that operators should halt and report immediately.

The prompt safeguards against indirect prompt injection by instructing agents to treat filenames, downloaded scripts, and README files strictly as data rather than instructions. Creator Rıza Sabuncu manages the repository using an open translation workflow where community additions require BCP 47 language codes while keeping shell commands in English. Because disk-cleanup.en.md serves as the definitive source of truth, any methodological revisions must be applied there first before updating other languages. Developers who benefit from the project can send coffee donations directly to Sabuncu.