← Terse Blog
Prompt Habits

The Politeness Tax

By Luz Wang·Founder, Terse·Updated Jul 17, 2026

Politeness, hedging, and conversational filler typically make up 15–20% of a prompt sent to an AI model — and the model answers no better for any of it. Every "could you please possibly" is a set of tokens you bought that carry zero information. Cutting them saves real money and often produces sharper answers.

What Manners Cost

Humans are trained to be polite, and the training runs deep. "Please," "thank you so much," "if it's not too much trouble," "I was just wondering if you could" — to another person, that's warmth. To a tokenizer, it is dead weight billed at the same rate as your actual request.

Measured across typical developer prompts, three categories of filler account for 15–20% of input tokens:

Politeness: "please," "would you mind," "thank you so much!"
Hedging: "I think maybe," "perhaps we could," "it might be nice if"
Throat-clearing: "I was just wondering whether," "so basically what I'm trying to do is"

Any single instance is trivial — a few tokens. But prompts are sent hundreds of times a day, every day, and in agent sessions each message is also re-billed as history on every later turn. Across a year of daily use, filler adds up to millions of tokens spent on words the model effectively discards. For what those tokens are worth in dollars on your specific usage, the free tier math breakdown runs the numbers.

The Counterintuitive Part: Terser Prompts Get Better Answers

The instinct is that politeness at least does no harm. In practice, wordiness often does. Hedged prompts are vague prompts: "maybe we could possibly look at improving this somehow" forces the model to guess what "improving" means, what "this" is, and how much license it has. Rewritten as "cut this function's runtime," the request has one interpretation.

Before: "Hi! I was wondering if you could maybe help me
         figure out why this might be running a bit slowly,
         if that's not too much trouble?"
After:  "Why is this function slow? Profile and fix."
Saved:  ~24 tokens — and the second prompt is clearer.

Removing hedging is not just compression; it forces you to decide what you actually want before you ask. That discipline — the heart of the terse mindset — improves answers from models and colleagues alike. Direct commands consistently perform as well as or better than padded requests in practice; the model does not have feelings to bruise, and it does not reward deference with effort.

Cutting the Tax Without Thinking About It

Nobody hand-edits every prompt for long. The sustainable fix is to automate it:

Start with the verb. Delete the opener and lead with "Write," "Fix," "Explain," "List." If the first word of your prompt is "I," there is usually a shorter version.
Let an optimizer strip filler on the fly. Terse's Normal mode removes politeness, hedging, and filler automatically before your prompt is sent — meaning intact, token count down. It never touches code, quoted strings, or negations, so "do NOT delete the database" stays exactly as written.
Watch the before/after diff. Terse shows exactly which words were removed from each prompt. Within a week you learn which of your habits cost the most — most people are surprised by how often they hedge.

When to Keep the Soft Words

Tone words are the payload in some prompts. If you're drafting an email, a user-facing message, or creative writing where voice matters, the qualifiers are content, not filler — use a gentler optimization mode and keep them. The politeness tax only applies to words aimed at the model rather than at a human reader. Knowing the difference is most of what token optimization is.

Cut the Filler Automatically

Terse strips politeness, hedging, and filler from every prompt before it's sent — 15-40% fewer tokens with your meaning intact. Free tier includes 1,500 optimizations a week.

Download Terse

Frequently Asked Questions

Does saying please to an AI improve its answers?

No measurable improvement. Models respond to the informational content of a prompt, not its politeness. Direct, specific commands perform as well as or better than padded polite requests, because they leave less to interpretation.

How much of a typical prompt is filler?

Measured across everyday developer prompts, politeness, hedging, and throat-clearing account for roughly 15-20% of input tokens. Terse's Normal mode typically removes that share without changing the prompt's meaning.

Is it safe to strip words from prompts automatically?

Rule-based stripping is safe when it protects semantically critical words. Terse never removes negations (not, no, never), conditionals (if, unless), code, quoted strings, or ALL-CAPS emphasis — only words that serve as conversational filler.