I’ve been using Claude for six months now, and I just realized I’ve been throwing money at the wall like an idiot. Every time I’d upload a PDF—a twenty-page report, a contract, a spreadsheet with charts—Claude would devour tokens like it was going out of style. Seventy thousand tokens gone before I even typed my first question. I’d sit there watching my credits drain and think, there’s gotta be a better way. Turns out there is, and Microsoft just gave it away for free.

Why This Keeps Coming Up

Here’s the thing nobody tells you when you start with Claude: PDFs are bloated. A Word document with formatting, images, and layout instructions is basically a mess of hidden code that Claude has to parse through. Every font change, every margin adjustment, every embedded image metadata—it all counts as tokens. So you’re paying to teach Claude how to ignore ninety percent of what’s in the file before it even gets to the actual content you care about.

I was uploading reports and getting slammed. A single PDF could be anywhere from fifty to one hundred thousand tokens depending on how it was formatted. Then I’d ask Claude a follow-up question, and boom—another ten thousand tokens. Three or four documents in a week, and I was hemorrhaging credits. I started wondering if there was a preprocessing step I was missing, something to clean files before they hit Claude’s tokenizer.

What I Actually Found Out

Microsoft released MarkItDown earlier this year, and it’s exactly what I was looking for. It’s free, open-source (MIT license), and it does one thing beautifully: it strips all the formatting noise from documents and converts everything to clean Markdown. PDFs, Word docs, Excel files, PowerPoints, even YouTube transcripts—it handles all of it. You run the file through MarkItDown, and what comes out is pure text in Markdown format. No hidden formatting. No bloat.

The results are ridiculous. That twenty-page PDF that was eating seventy thousand tokens? After running it through MarkItDown, it drops to somewhere between twenty and thirty thousand tokens. That’s a thirty to seventy percent reduction depending on how formatting-heavy the original file was. The content is identical. Claude sees everything it needs. But it’s not wasting tokens on layout instructions and image metadata.

I tested this with a real client contract—dense legal document, lots of formatting. Before: ninety-three thousand tokens. After MarkItDown: thirty-two thousand tokens. I asked Claude the same questions on both versions, got the same answers, but saved sixty thousand tokens on the input alone. At current pricing, that’s money in my pocket.

What You Can Do With This Today

The workflow is simple. You download MarkItDown (it’s on GitHub, literally free), you run your PDF or document through it, and you get a clean Markdown file in seconds. Then you paste that Markdown into Claude instead of the original file. Same information, way fewer tokens. If you’re processing multiple documents or doing regular analysis work, you can automate this—pipe files through MarkItDown as part of your pipeline before they hit Claude.

I’ve started doing this for everything now. Client reports, research papers, contracts, even spreadsheets I need Claude to analyze. It’s become muscle memory. File comes in, runs through MarkItDown, goes to Claude. No more watching my credits disappear into formatting overhead. And because I’m burning fewer tokens per document, I can afford to be more thorough with my prompts, ask more follow-up questions, and actually use Claude the way it’s meant to be used.

The GenX move here is recognizing that someone smarter already solved this problem and just gave it to you. No subscription, no upsell, no corporate wrapper. Microsoft built a tool, released it with zero fanfare, and it just works. If you’re serious about keeping your AI costs reasonable while scaling your work, this is table stakes. Download it, test it on your next PDF, watch your token usage drop. You’ll wonder why you weren’t doing this from the beginning.

More soon.

Watch the Full Video

I covered all of this in a short video too — sometimes it’s easier to watch than read.

Watch on TikTok →