If you're using AI assistants daily, you're probably burning tokens you don't need. Not because your prompts are bad — because they're carrying dead weight.
I run a token audit once a week. It takes 10 minutes and consistently cuts my usage by 30-40%. Here's the exact checklist.
The 10-Minute Token Audit
Step 1: Find Your Heaviest Prompts (2 min)
Open your last 10 conversations. Sort by length. Your top 3 longest prompts are where the waste lives.
Most people discover one or two prompts that are 3x longer than they need to be. These are your targets.
Step 2: Strip Redundant Context (3 min)
For each heavy prompt, ask:
- Am I pasting the entire file when the model only needs 20 lines?
- Am I including "just in case" context that the model never references?
- Am I repeating instructions the system prompt already covers?
A common pattern: developers paste an entire 500-line file when the bug is on line 47. Instead:
// Bug is on line 47. Here's the relevant section (lines 40-55):
[paste only those lines]
// The function is called from:
[paste the one call site]
That's 30 lines instead of 500. Same result, 94% fewer tokens.
Step 3: Kill Boilerplate Instructions (2 min)
Remove instructions your model already follows by default:
- ❌ "Please write clean, well-documented code" (it already does)
- ❌ "Make sure to handle errors" (too vague to help anyway)
- ❌ "Use best practices" (meaningless)
Replace with specific constraints:
- ✅ "Return errors as
{ ok: false, error: string }, never throw" - ✅ "Max 40 lines per function"
- ✅ "No external dependencies"
Specific constraints are shorter AND more effective than generic instructions.
Step 4: Check for Conversation Bloat (2 min)
Long conversations accumulate context. After 10+ messages, you're paying for the entire history in every request.
The fix: start a new conversation when you switch tasks. Don't carry a refactoring discussion into a debugging session. Each new chat starts with a clean (and cheap) context.
A simple rule: one task, one conversation.
Step 5: Audit Your System Prompts (1 min)
If you use custom system prompts, measure them. I've seen system prompts that are 2,000 tokens of instructions the model already follows. Every message pays for that overhead.
Trim your system prompt to constraints that actually change behavior. Everything else is wasted tokens on every single request.
The Numbers
Before my first audit:
- Average prompt: ~1,200 tokens
- Average conversation: ~15,000 tokens total
After:
- Average prompt: ~700 tokens
- Average conversation: ~8,500 tokens total
That's a 43% reduction. At scale, that's real money.
Make It a Habit
Set a weekly reminder. Every Friday, spend 10 minutes on this checklist. You'll save more in tokens than the time costs you — and your prompts will get better as a side effect.
Lean prompts aren't just cheaper. They're faster, more focused, and produce better output. Turns out, the model works better when you stop burying the signal in noise.
Top comments (0)