Save 90% of Your Tokens in Claude (the PDF method)
Every PDF you paste into Claude costs you. One free Microsoft command turns any document into Markdown, plus the manual method if you never touch a terminal.
May 24, 2026 · 5 min read
Every PDF you paste into Claude costs you.
When you drop a raw document into Claude, you're not only paying for the useful info. You're paying for all the noise around it.
The Claude AI Lab is my Skool community where I share my Claude systems and the more advanced modules. Access is $67/month.
Join the Lab →Claude has to parse the layout, the images, the headers, the margins. You pay for noise.
Vision tokens per image. Multiply that by five or ten screenshots in one session and your context is full.
The result: you fill your context with junk tokens, you burn through your plan faster, and you slow Claude down on every reply.
Convert every document to .md before handing it to Claude. You keep 100% of the information and pay a fraction of the price.
There are two ways to do it. One command if you are willing to open a terminal, and a fully manual route if you are not. Both end with the same file.
The automatic method
Microsoft maintains a free, open source tool that does exactly this job: it swallows a file and hands back clean Markdown. It is called MarkItDown.
python3 --version in a terminal. You need 3.10 or newer. If you do not have it, install it from python.org.pip install 'markitdown[all]'. The quotes matter, they stop the terminal from interpreting the brackets.markitdown --help. If the help text prints, you are ready.After that, converting a document is a single line:
markitdown my-report.pdf -o my-report.md
The .md file appears next to the original. You drag it into Claude exactly as you would have dragged the PDF.
What used to take two minutes of copy and paste per document now takes two seconds. Across ten documents, you go from twenty minutes to thirty seconds.
And it does not stop at PDFs. The same tool converts Word, Excel, PowerPoint, HTML, CSV, JSON, EPub, ZIP archives whose contents it walks through, images whose text it reads, and even audio it transcribes.
pip install 'markitdown[pdf]'. Lighter and faster to set up.One honest caveat, and it comes from Microsoft: the output is built to be read by a machine, not to render a document perfectly for the eye. Unusual columns and complex layouts sometimes come out flattened. For your use case that is exactly right. Claude does not need your layout, it needs your text.
The manual method
If you would rather not open a terminal, you get the same result in four steps and about thirty seconds.
docs.new in your browser and paste. You can keep headings, bold and lists, the conversion preserves them.File, then Download, then Markdown (.md). The structure comes out clean: headings, lists, links, bold. You end up with an extremely light text file.A Markdown file is plain text with structure. No encoded images, no layout, no PDF metadata. Claude gets the pure signal without the visual noise.
This method has one advantage the automatic one does not: you filter as you copy. On an eighty page report where three pages matter, you only convert those three.
The four habits
Converting your documents fixes the biggest line item. These four habits fix the rest.
.md directly.The cleaner, more structured and more minimal your input, the faster, sharper and more stable Claude gets. Markdown is simply the plainest expression of that principle.
Where to start tonight
Take the document you handed Claude most recently. Convert it with the command, then ask the same question again on the .md.
You will notice two things: the answer is at least as good, and your conversation lasts a lot longer before it starts to tire.
Once you have seen it once, you will never hand over a raw PDF again.
Want to go further?
In the Lab, I share the configs, instructions and architectures I keep refining to turn Claude into a real copilot.
A dedicated session or program, tailored to your tools and use cases.
And day-to-day, I post one reel a day on Instagram: @quentin_iamarketing