The Pain: How “System Data” on a 256GB Mac Bloats Past 100GB

Every veteran Mac user shares the same nightmare. That gray “System Data” bar in About This Mac eats 50GB, 80GB, sometimes over 100GB, and nobody can explain what’s inside. The 256GB base MacBook Air gets hit hardest. macOS claims half the drive, a few dev tools claim another chunk, and you’re stuck deleting files just to install the next Xcode update.

Cache hides in three places: Lark, Chrome, and terminal toolchains. Lark’s LarkShell/aha directory caches every document you’ve ever opened. One user accidentally piled up 42GB (per the project README). Chrome’s Service Worker plus debug data can balloon past 50GB (per the source post). Python devs routinely watch ~/.cache/uv quietly swallow 56GB (per the source post). Add Telegram media cache, npm cache, Homebrew downloads, and a one-to-two-year-old dev machine routinely carries 100GB of junk.

Here’s the catch: half that cache is garbage, half is state. Lark’s document cache is safe to delete, but login session files in the same directory force a QR-code re-scan. Chrome’s Service Worker can go, but the IndexedDB next to it holds your web app data. Traditional cleaners list “which folders are big” but never answer “which ones are safe to delete.”

The official fix hurts even more. Apple’s SSD is soldered to the logic board. Jumping from 256GB to 512GB costs $200 through Apple’s upgrade program, and every tier up adds another hundred-plus. That money doesn’t have to leave your wallet.

The Opportunity: Outsource “Should I Delete This?” to AI

Disk Raccoon (硬盘小浣熊), an open-source project, takes a different angle. It builds a structured cache map that flags what’s safe to delete and what’s off-limits, then lets an AI coding assistant work through it. This isn’t another cleanup app. It’s a Skill you install into Claude Code, Cola, or similar AI coding tools. Think of it as a plugin that teaches your AI assistant how to handle this specific task. Tell your agent “check my cache” and it gets to work.

Project lives on GitHub: https://github.com/orange2ai/disk-raccoon, MIT licensed, 11 stars at time of writing (per source post, not independently verified), last updated late August 2026 (per source post, not independently verified). The official tagline says it straight: “macOS cache cleanup Agent: wash before eat. Cache map + three-tier safety levels, only deletes what’s safe. First real run cleared 135GB overnight.”

“Wash before eat” comes from raccoon behavior. They’re the only animals that rinse food before eating. Not refusing to eat, just not eating blindly. That maps perfectly to macOS cleanup’s core tension: big doesn’t mean deletable, deletable doesn’t mean big. Traditional tools solve the first half. Raccoon solves the second.

Side note: knowing how to install AI Skills is itself a billable skill. Charge $30 per cleanup gig, the tool’s free, and everything you save the client is pure profit.

The Path: Three-Tier Safety + Cache Map

Disk Raccoon’s workflow breaks into three steps, each with clear guardrails.

Step 1: Scan with Mole First, Locate the Big Folders

Mole is another open-source tool built specifically for finding large folders. Think of it as Raccoon’s “eyes.” Without it, Raccoon doesn’t know where to look. Mole lists every directory eating 10GB, 20GB, 50GB on your drive. This step is the “medical report” telling the AI where to focus. Raccoon works without Mole, but the AI scans blind and slower.

Step 2: AI Walks the Cache Map, Item by Item

Raccoon ships with a built-in references/cache-map.md where every cache path carries a safety rating. Here’s what real entries look like:

1
2
3
4
5
6
7
8
9
10
11
12
13
14

## ~/.cache/uv
- Level: ✅ Safe to delete
- Type: Python package cache
- Note: uv install rebuilds automatically, no project impact

## ~/Library/Application Support/LarkShell/aha/profile_explorer
- Level: ⚠️ Delete with caution
- Type: Lark document cache
- Note: Delete only profile_explorer; profile_main holds login state

## ~/Library/Application Support/Google/Chrome/Default/Service Worker
- Level: ✅ Safe to delete
- Type: PWA offline cache
- Note: Browser rebuilds automatically

What about paths not on the map? Raccoon’s policy: “figure it out first, don’t touch it if you can’t.” The AI researches each directory (which app owns it, what it stores, what breaks if deleted), then presents a checklist for user approval.

Step 3: Read-Only by Default, sudo Stays with the User

Raccoon’s safety design follows hard rules:

• Read-only by default: First pass is observe-only, report first.
• Reversible when possible: Trash over rm -rf whenever feasible.
• Quit before delete: Close running apps first to avoid rebuild-during-delete loops.
• sudo stays with you: System-level deletes only generate commands; you run them.

Why does read-only-by-default beat one-click cleanup? The recovery cost of one wrong delete (lost chat history, broken project dependencies) dwarfs the time saved clearing a few dozen GB. CleanMyMac’s “scan + clean” design dumps the judgment call on you. Will you actually click “Clean All”? Most people won’t. Raccoon hands the judgment to AI, and you just nod or shake your head on the final list. Decision cost drops to near zero.

The Case Study: How Those 135GB Got Freed

The source post author’s first Raccoon run cleared three big offenders:

That 135GB translates to the cost of Apple’s 512GB upgrade: $200. At roughly 100GB per cleanup, each run is worth about $150. This case is representative because it covers the three most common “disk killers”: dev toolchains (uv), browsers (Chrome), office apps (Lark). A Python dev + Lark + Chrome power user is basically the standard dev setup. If that sounds like you, your drive probably hides a similar 100GB+ cache. One run will likely reproduce this scale.

Why Traditional Cleanup Tools Can’t Solve This

Mac cleanup apps aren’t scarce. CleanMyMac, DaisyDisk, OmniDiskSweeper are all veterans. But they share one blind spot: they tell you “what’s big” but never “what’s safe to delete.”

Take Lark. The LarkShell/aha directory is nearly invisible in Finder, buried under multiple Library layers. CleanMyMac flags it yellow, but clicking in reveals a maze of cryptic subfolders nobody dares touch. Chrome is worse. Service Worker, IndexedDB, Cache Storage, Cookies each follow different rules. Delete the wrong one and your PWA’s offline data vanishes.

Raccoon’s fix is AI-driven “semantic judgment.” Same 50GB Chrome directory: the AI distinguishes Service Worker (deletable), IndexedDB (case-by-case), and login state (untouchable). CleanMyMac flags 50GB of Lark cache yellow with zero explanation; Raccoon tells you which 30GB is safe and which 20GB is login state. That’s why CleanMyMac charges $40/year and still can’t solve this. Rule engines can’t understand path semantics. This is the first time LLMs give cleanup tools actual reading comprehension.

Call to Action: Install the Raccoon Tonight

If you’re on a Mac, constantly fighting storage warnings, and refuse to pay $200 for an upgrade, follow these three steps:

Step 1: Install the Skill

1

git clone https://github.com/orange2ai/disk-raccoon ~/.claude/skills/disk-raccoon

Any agent supporting the SKILL.md protocol works: Claude Code, Cola, and others.

Step 2: Pair It with Mole for the Health Check

Run Mole first to surface the big folders, then let Raccoon judge what’s safe. One tool scans, the other decides. Clean division of labor.

Step 3: Talk to Your Agent

Skip the complex commands. Just say “check my cache” or “drive is full, what can I clean?” The agent walks the cache map and builds your checklist item by item. Nod to clean, shake your head to skip.

The cache map is alive. The project author asks users in the README to file issues or PRs with new cache paths they discover. Drop the path, size, and owning app in your issue. One cleanup session might save the next user $200.