Stop Burning Claude Credits: Three Changes, One Measured

Three changes. One of them is measured, and it is the one worth doing first.

What's actually proven here

The number in circulation for stacking all three is 70%. Read section 5 before you repeat it.

1 · Stop uploading raw PDFs

Every page of a PDF is converted to an image before Claude reads it. You are not paying for the words on the page. You are paying for a picture of the page, plus the words, and the picture costs the same whether the page holds forty words or four hundred.

Anthropic documents both halves of this: text runs about 1,500–3,000 tokens per page depending on density, and because each page is converted into an image, image costs apply on top.

The measurement

Two real PDFs, each read twice, against a control run that reads a one-byte file so the fixed overhead of the harness cancels out. Same document, same content, both ways. Claude Opus 5.

Per page, the PDF cost 1,260 and 1,304 tokens. Nearly identical — across two documents whose text density differs by a factor of six.

That flat per-page number is the whole finding. A page costs what a page costs. The sparse document, the one with almost nothing on it, is where the waste is worst: 80% of what you spent bought you a photograph of white space.

Anthropic's own docs land in the same place from the other direction — a 3-page PDF runs about 1,000 tokens as plain text extraction and about 7,000 with the page images included. Two independent routes, same answer.

The fix

Copy the text out. Upload the text.

1. Open the PDF, select all, copy.
2. Paste into a Google Doc.
3. File → Download → Markdown (.md).
4. Upload the .md instead.

No Google Doc handy? Any of these does the same job:

# macOS / Linux, if you have poppler installed
pdftotext -enc UTF-8 yourfile.pdf yourfile.md

Paste-into-a-plain-text-file works too. The format barely matters — what matters is that it is text and not a picture of text.

When to keep the PDF

Do not do this blindly. Uploading the raw PDF is correct when the thing you need is visual: charts, scanned pages, signatures, tables whose layout carries meaning, handwriting, anything where the picture is the point. Strip it to text and Claude loses the part you were asking about.

Text-only documents — reports, transcripts, guides, contracts, articles — are the ones where you are paying twice for the same words.

How the numbers were produced

Reproduce it yourself if you want to check:

• Both files read through Claude Code's Read tool, --output-format json, MCP servers off, Read the only tool allowed.
• Document cost = that run's cache-write tokens minus a control run that read a one-byte file. That control is what makes it a measurement instead of an anecdote — without it you are measuring the system prompt, not the document.
• Two runs per condition. The markdown runs came back within 4 tokens of each other; the PDF runs within 71.
• Confirmed the mechanism directly rather than assuming it: the 9-page PDF arrived as 9 image blocks, one per page.

Two honest caveats. This was measured through Claude Code, and corroborated against Anthropic's published API behavior — the claude.ai upload box was not separately instrumented. And the markdown side carries line numbers the Read tool adds, which inflates it slightly; the real saving is, if anything, a little better than 60%.

2 · Claude Counter — watch the burn

Free, live, and it does what it says. Verified on the Chrome Web Store: Claude Counter by Hassaan Haider, 4,000 users, 5.0 stars, last updated 29 May 2026.

Install: chromewebstore.google.com/detail/claude-counter/hjoepjapcijlmkjebpdfpbdmlfddpeij (opens in a new tab) Source: github.com/she-llac/claude-counter (opens in a new tab)

It adds three read-only overlays to claude.ai:

• Token count for the current conversation, against the 200,000-token context limit.
• Cache timer — a countdown showing how long the conversation stays in Claude's prompt cache. Continuing inside that window is dramatically cheaper than starting cold.
• Session and weekly usage bars, read from Claude's own /usage endpoint, each with a reset countdown.

Everything runs locally in the browser. Nothing is sent anywhere.

Two things to know before you trust the number

It counts with the wrong tokenizer. The extension bundles gpt-tokenizer (o200k_base), which is OpenAI's. Claude tokenizes differently. Expect the displayed count to run low — typically 15–20% on ordinary prose, and further off on code or non-English text. It also excludes the system prompt. Treat it as a gauge, not a meter: useful for "this thread is getting expensive," useless for "this thread cost exactly N."

It saves nothing on its own. A fuel gauge is not fuel economy. Its value is that it makes the other two changes visible — you can watch the counter move when you swap a PDF for markdown, which is how you confirm the fix is working on your documents.

The session and weekly bars are the exception worth caring about: those come from Claude's own endpoint, so they are accurate, and they are the ones that tell you how close you are to a limit.

3 · The standing instruction

Most of the waste that isn't PDFs is output — preamble, restating your question, "Great question!", a summary of what it just said, an offer to keep going. You pay for every word of it, on every message, forever.

Fixing it takes one paste, one time.

Where it goes: claude.ai → your initials, bottom-left → Settings → Instructions for Claude (this is the box formerly called custom instructions or personal preferences). It loads into every new conversation automatically. In Claude Code, the same text goes in CLAUDE.md.

Paste this:

Read before writing. Do not re-read unchanged files.
Thorough in reasoning, concise in output.
No opener, no closing summary, no offers to continue.
No emojis. No em-dashes.
Answer first, then only the reasoning I need to check you.
Do not guess APIs, versions, flags, or package names — verify before asserting.
Say "I don't know" instead of filling the gap.

Seven lines. Adjust the tone rules to taste; the structural ones — no opener, no closing summary, answer first — are where the savings are.

The catch nobody mentions

The instruction itself costs tokens on every single turn. It is prepended to everything. Keep it short and it pays for itself many times over on long answers; let it grow into a 500-line manifesto and it costs more than it saves. This is the specific failure mode to watch for — the file that starts as seven lines and is forty by March.

It also only helps in proportion to how much Claude was writing. On long generated answers the saving is large. On short factual questions there was little fluff to cut, so there is little to save.

On the repo

The version circulating for this is github.com/drona23/claude-token-efficient — MIT, ~6k stars, and its core file is genuinely about 350 bytes. Its own benchmark reports roughly 63% fewer words out, and 17.4% cost savings on external coding challenges.

Be careful with those two numbers — they are not the same claim. 63% fewer words is not 63% cheaper, because output is only part of what you pay for and input is untouched. The repo's own honest figure is the 17.4%, and it says plainly that the file adds input tokens on every turn, so net savings show up on high-output workflows and not everywhere.

Those are the repo's measurements, not re-run here. The instruction block above is written to stand on its own — you do not need the repo to use it.

4 · Shortcut codes

Already written up in full, separately. Fifty of them, each one actually tested with a paired A/B run — same prompt with and without the code — including the eight whose common descriptions turned out to be wrong.

→ CODES — 50 Claude Prompt Codes

Not repeated here. The short version is that they are prompt patterns, not official Anthropic commands, they shift between model versions, and the honest verdict on all fifty is in that guide rather than in a list of names.

For token burn specifically, the two that matter are TIGHT (end of prompt) and RAW (front) — they do by the message what the standing instruction in section 3 does permanently.

5 · About that 70%

It doesn't hold, and here is exactly where it breaks.

The 70% is three separate savings added together by someone who did not measure any of them. Take it apart:

The counter contributes zero. It is a display. Adding a gauge to your dashboard does not reduce fuel consumption. Any stacked total that includes it is inflated by one full term.

The other two apply to different things and don't compound. The PDF fix cuts input on documents you upload. The standing instruction cuts output on answers Claude writes. If you never upload documents, leg one saves you nothing. If you only ask short factual questions, leg two saves you nothing. They are not multipliers on one bill — they are two separate line items, and which one matters depends entirely on how you work.

What is defensible:

• On a text-only document you upload, converting to markdown first: 60–80% of that document. Measured, twice, two documents, with a control.
• On verbose generated answers, the standing instruction: real, and the only published figure is the repo's own 17.4% on coding tasks.
• Combined, across a realistic month: a large chunk, and nobody has measured which chunk.

If you want a number you can defend, use your own. Install the counter, run one document both ways, and read the difference off the screen. That takes four minutes and it beats repeating someone else's arithmetic.

The four-minute version

1. 1Convert text-only PDFs to markdown before uploading. Keep the PDF only when the visuals are the point. This is the one with a measurement behind it.
2. 2Install Claude Counter. Trust the session and weekly bars; treat the per-message token count as approximate and low.
3. 3Paste the seven-line instruction into Settings → Instructions for Claude. Keep it short forever.
4. 4Codes are in the CODES guide. Start with RAW and TIGHT.

Token measurements run 2026-08-24 on Claude Opus 5 via Claude Code, two runs per condition against a control. Chrome Web Store listing and repo status verified the same day. PDF processing behavior cross-checked against Anthropic's published documentation.