My reading inbox in Obsidian (what I call "Clippings") had ballooned to just under 500 clips. Each one had sat there since I imported them from Pocket and added more with the Obsidian Web Clipper. A majority stayed unread, waiting for my attention, which I never gave. Scrolling through the bookmarks haul feels overwhelming. I had achieved some success with an earlier Gemini + Obsidian experiment. Now, I pointed Claude Desktop's Filesystem MCP server at the whole folder and asked it to sort the pile. Instead of simple bookmark organization, it became a prompt playground as Claude found patterns individual bookmarks don't reveal.

Claude scans every clip in one pass You don't need to open articles one by one

Claude Desktop's Filesystem MCP server can read an entire folder. I pointed it at my Clippings and asked it to read the front matter and body of every file in a single prompt. In one shot, Claude had gone through all of the saved links. This method of accessing a bunch of articles in one sweep beats manually reviewing the massive backlog of unread articles.

Read every file in my Clippings folder, including the frontmatter, and give me a summary of what's in there.

I worry that scanning through messy articles with embedded images, raw HTML artifacts, or large code dumps could eat up my Claude tokens or throw up execution errors. I haven't seen any evidence of the latter in my limited testing so far, though the former is self-evident.

The good news is that the first scan was fast, deep, and pretty accurate. As an alternative experiment, I can ask Claude to run two passes. The first will be a frontmatter-only scan that creates an index file, followed by a deep relevance analysis on that single index file.

Sort by context, not age Your oldest clips might be your most valuable

I asked Claude to bucket every clip into four groups: Act now, Reference, Archive, and Delete. It checked each clip against the notes in my active project folders rather than just the save date. A lot of book summary webpages and learning-related articles landed in "Reference" because they are loosely connected to some of my projects, and only a few are backlinked to other notes. As Claude told me — they are solid material tied to a real project, but not urgent.

Sort these clips into four buckets: act now, reference, archive, delete. Check each one against my active project notes, not just the save date.

Claude marked dozens of clips for removal that I'd once thought I'd need "someday." Let's give Claude the benefit of the doubt, as many of my read-it-later bookmarks are usually from serendipitous finds. Claude went by the absence of a connection to any active project, or broken/duplicates to mark the articles for deletion. Anyways, surfacing duplicates — if there are many — is a useful way to declutter redundant bookmarks.

While scanning through my someday clips, I discovered topics I'd already learned or apps I'd since dropped. Deleting them removed more noise. I found timeless evergreen guides. I also found Claude misinterpreting outdated software documentation as "highly relevant" because the topic-matching score was high. That's why a manual pass through Claude's answers is always vital.

Claude surfaces clusters you didn't see Your saves reveal what you actually care about

Beyond sorting, I asked Claude what patterns it noticed across the whole folder. It grouped clips into themes I'd never consciously tracked or tagged in the Obsidian note. A good chunk of my "productivity" saves turned out to be about the more granular "action over analysis" theme. This type of wholesale analysis can better show your evolving interests than your tags can.

Look across all these clips and tell me what themes or patterns keep coming up, even ones I haven't tagged.

I'm not diligent about tagging all my notes or even filling out the frontmatter. So, Claude's assessment, as it cuts across notes and folders, is especially useful for this type of bird's-eye view. This works even if I don't use folders to organize my Obsidian notes.

Claude's clusters explained why I kept saving that content in the first place, and that distinction pointed me toward what to read next rather than just what I'd already collected.

It flags the surprising keepers Some good clips were about to get deleted

A few clips Claude had almost archived turned out to be the most useful when I skimmed them again. It marked three as "unclear value" instead of auto-deleting them, which gave me a chance to double-check. A prompt can force Claude to admit uncertainty, making it more trustworthy than the version that's always confident.

If you're not confident a clip should be archived or deleted, mark it "unclear value" instead of guessing.

I wanted a clean, decisive list. Getting an "unclear value" pile back felt like the tool hadn't finished the job with the first prompt we discussed. But this works really well as a second filter before a purge.

That pile was small, just about two dozen, and it caught real keepers I'd lost in a fully automated purge. A little uncertainty saved more value than a forced yes-or-no answer would have.

Make a final call Treat the sort as a draft, not a final culling list

Claude's sort was a first pass, not a final answer. I quickly reviewed all fifty "keep" clips myself before moving them out of the inbox, and I skimmed the delete pile once before clearing it. The takeaway, as always, is to use AI sorting to cut down the work, not to replace your judgment.

Code snippets, quick recipes, single-paragraph quotes, or saves with just website links can often look like "low value " to an AI sorter. Claude will frequently push these into the Delete pile unless explicitly told to evaluate information density over word count. I added this to the first prompt we saw.

Judge value by information density, not word count. A short clip with a complete answer counts as high value.

Also, the five minutes I spent double-checking caught one clip Claude had flagged as a duplicate that wasn't actually a duplicate. One idea is to physically move ambiguous clips into a dedicated Review_Uncertain subfolder in Obsidian.

Run this sort on your own inbox this weekend

If your reading inbox has outgrown casual scrolling, don't wait for it to become a mountain. Point Claude Desktop's Filesystem MCP at your folder, ask it to sort by relevance and flag patterns, then spend fifteen minutes reviewing what it hands back. You'll get a cleaner vault, and probably a surprise or two about what you actually read for.