Skip to content
You commented “CAVEMAN”: here it is

Token Audit Checklist: Find Where You're Wasting 50 Percent

Run this 10-item audit and identify exactly where you're burning 50 to 75 percent of your Claude token budget on waste.

Free checklist · copy-paste ready · built by the BrainVaultAI team

Most people burn tokens in four places: verbose system prompts, repeated context in every message, over-explained instructions, and asking Claude to confirm understanding before acting. None of that helps output quality. All of it inflates cost. The caveman principle: say the minimum that produces the right output. Not minimal effort, minimum signal. Claude is not a human who needs warmup. It needs precision. Run this audit on your three most-used prompts. Most people find 40-60% reduction is achievable without any quality loss.

Token Audit Checklist
Run this on any prompt you use regularly: SYSTEM PROMPT AUDIT
[ ] Does my system prompt repeat anything that's already in CLAUDE.md? Remove it.
[ ] Do I explain WHO Claude is for more than 2 sentences? Trim to 1.
[ ] Am I giving examples when the instruction is clear without them? Remove examples.
[ ] Do I have "Please" or polite filler? Delete all of it.
[ ] Am I telling Claude to "think carefully" or "take your time"? Remove, it's noise. MESSAGE AUDIT
[ ] Do I re-explain the context on every message? Move it to CLAUDE.md or a file.
[ ] Do I ask Claude to confirm it understood before answering? Remove, just answer.
[ ] Do I ask for things I then ignore? (e.g., "include 3 options" then use 1), stop asking.
[ ] Am I including background Claude doesn't need for this specific task? Cut it.
[ ] Is my prompt more than 150 words for a task that takes Claude 3 seconds? Trim it. QUICK MATH
Count words in your prompt. Estimate what 50% of that is. Try the shorter version.
If output quality holds, you're done. If it drops, add back only what was missing.
Prompt Compression Prompt
I'm going to give you a prompt. Your job is to compress it, remove everything that doesn't change the output quality. Do not change the intent. Do not remove constraints that matter. Just cut the fat. For each thing you remove, tell me why it was unnecessary.
For each thing you keep, tell me why it's load-bearing. Target: 40-60% word reduction without output quality loss. Here's the prompt:
[paste your prompt]

Want the editable version?

We'll send the editable version of this asset, plus a short series of our highest-leverage AI plays for operators. No spam. Unsubscribe in one click.

Free. No spam. Unsubscribe in one click. By submitting you agree to receive emails from BrainVaultAI.

More free AI guides in the full library · follow @brainVaultAI