Stop Torching Your Claude Credits: The 5-Minute Token Fix
What's draining your Claude usage
Most people blow through their Claude usage far faster than they need to, and the single biggest leak is uploading raw PDFs and screenshots. Claude has to "read" every visual chunk of those, which burns a pile of tokens on words it could have read almost for free as plain text. Here's the full setup to plug that leak and squeeze far more out of the same plan, plus a way to see exactly where your tokens are going. A token is a chunk of text Claude reads or writes, and your plan is metered in tokens.
Fix 1 — Never upload a raw PDF again
A PDF or screenshot makes Claude process the whole visual, so it spends tokens on layout and images and not just the words you actually care about. Pull the text out first: copy it into a Google Doc, export that as a Markdown file (a stripped-down plain-text version), and upload the Markdown instead. Same words, a fraction of the cost. If you do this often, Microsoft's free tool markitdown (github.com/microsoft/markitdown) converts PDFs, Word docs, slides, and spreadsheets into clean Markdown for you in one step.
Fix 2 — Watch your tokens burn in real time
You can't fix what you can't see. Install the free Claude Counter extension (github.com/she-llac/claude-counter), which puts a live token count and your usage bars right on claude.ai, along with a cache timer so you know how long a conversation stays cheap to continue. If you'd rather install straight from the Chrome Web Store, "Claude Usage Tracker" does the same job and even breaks down which files and projects are eating your tokens.
Fix 3 — Paste one instruction that makes every answer cheaper
There's a free repo, claude-token-efficient (github.com/drona23/claude-token-efficient), built around a single drop-in instruction that keeps Claude's answers tight instead of letting it ramble. Grab the full CLAUDE.md from the repo, or just paste this into your customizations (Settings then Personal preferences on claude.ai, or your project's custom instructions):
Be concise and direct. Skip the preamble and pleasantries, and don't restate my question.
Give me the answer first, then only the reasoning I actually need.
Don't pad with summaries or extra caveats unless I ask, and match the length to the task.Every response after that does the same job for fewer tokens.
Fix 4 — Use prompt shortcuts to sharpen the output
These are short directives you drop at the front of a prompt to steer the answer. They work through clear instruction rather than secret keywords, and a few of them genuinely pull their weight:
- "Steps only." — skips the explanation and gives you just the actions.
- "Be terse." — compressed output with no filler.
- "ELI5." — explains it like you're brand new to the topic.
- "Push back if I'm wrong." — gets you a real second opinion instead of automatic agreement.
- "Output only the code." — no commentary wrapped around it.
For a big curated set, there's a free list of 100 Claude prompt shortcuts (gist by Samarth0211) you can browse and steal from.
Stack them
Switch your PDFs to Markdown, keep the counter visible, paste the instruction once, and lean on a few shortcuts, and you'll claw back a real chunk of the usage you've been quietly burning, most of it from a setup you only do one time.
Frequently asked questions
- Why do PDFs cost so many tokens? — Claude processes the whole visual, layout and images included, not just the words, so a screenshot or PDF burns far more than the same text as Markdown.
- Is markitdown free? — Yes. It's Microsoft's free tool for converting PDFs, Word docs, slides, and spreadsheets into clean Markdown.
- What does the one-paste instruction do? — It keeps Claude's answers tight instead of letting it ramble, so every response does the same job for fewer tokens.
- Do the prompt shortcuts use secret keywords? — No. They work through clear instruction, not hidden codes.
Get the tools
- markitdown — https://github.com/microsoft/markitdown
- Claude Counter — https://github.com/she-llac/claude-counter
- claude-token-efficient — https://github.com/drona23/claude-token-efficient