Filler in a prompt is the language you add so the request sounds like a favor. "Please can you help me, I would like you to write something nice, thank you." None of those words names a task, a source, a limit, or a format. They are tokens anyway. cl100k_base will count "please" as a token, "thank you" as two, and the whole cushion as a dozen or two. Whether that is waste depends on what else you sent.
On a 30-token prompt, a dozen tokens of cushion is a third of the input. On a 3,000-token prompt that is mostly a contract, the same dozen tokens are invisible next to a duplicated clause. The density score flags the cushion because it is a writing problem, not because it is always a cost problem. The reasons list will say "6 filler phrases" on the worked example on the home page, and the token count only drops by one after the rewrite, because the headings cost what the cushion saved. Both facts can be true. Delete the filler so the task is visible. Do not expect the invoice to flinch unless the filler was a paragraph you repeat on every call.
The phrases the score treats as filler are fixed: please, kindly, can you, could you, would you, I would like, I would like you to, I want you to, help me, if possible, thank you, thanks, just, really, very, basically, and actually. Each hit costs four points, and the penalty stops at 28 so a long apology cannot hide a prompt that is otherwise a brief. Longer phrases are counted instead of the shorter ones inside them. "I would like you to" is one hit, not also "I would like." The list is blunt on purpose. "Just" is sometimes meaningful ("just the subject line, not the body"). The score cannot tell. You can. If the reasons line mentions "just" and you meant the restriction, leave the word and ignore that one hit.
Here is a before and after that is almost pure filler, so the token change is the point. Before: "Hi, I would like you to please summarize this for me if possible. Thanks so much, it would really help." After: "Summarize this." Paste both into the counter. The first is a small pile of tokens and no object, because the word "this" does not carry the document. The second is a verb with no object either. Deleting filler did not create a task. It revealed that the task was never there. The next edit is to paste the document under a Context heading and say what kind of summary: five bullets, no new facts, under 120 words. That second edit changes the token count upward and changes the answer. It is the one that matters.
Politeness aimed at the model's output is a different sentence and should stay if you mean it. "Write the reply in a calm tone" is an instruction about the product, not a cushion around the instruction. "Please write the reply" is cushion. The difference is whether deleting the word changes the spec. "Really professional" in the home-page example becomes "professional" under the optimizer, and the spec is the same. "Professional" itself may be vague, but it is not filler on this list. If you need it sharper, replace it with something you can check: no slang, no exclamation marks, no promise of a date.
There is a kind of filler that the list does not catch and that costs more. The preamble you leave in a system prompt: three sentences on how the company was founded, a paragraph of values, a note that the assistant should be helpful, harmless, and honest which repeats the provider's own defaults. None of those words is "please." All of them are paid on every call. Read the system text aloud and mark every sentence that does not change a decision. Those sentences are the filler that shows up on a monthly bill. The cost note walks through a 600-token cut on a 20,000-call day. That cut is this paragraph, not the word "kindly."
Filler also crowds the model's attention in a softer way. A short prompt that starts with two lines of apology and ends with "make it nice" spends its clearest positions on nothing. Models tend to weight the start and the end. If both are cushion, the constraint in the middle is the part that gets dropped. Moving the constraint into its own line does more than deleting "thanks," and doing both is cheap. The optimizer puts constraint-shaped sentences under a heading. Look at what it moved. If your only constraint was trapped inside "please try not to invent a date if possible," the heading version is easier to see and easier to keep.
Do not strip filler inside quoted material. The customer's "please help me" is evidence. The optimizer on this site removes filler phrases from the draft, which can also hit the customer's sentences if you leave them in the same blob. That is a reason to fence the customer's text before you optimize, or to label it "Context:" so it is moved rather than edited. A support reply that silently deletes the customer's "please" from the record you are answering is a worse prompt, not a leaner one. The delimiter note shows a fence around exactly this kind of text.
Teams sometimes add filler back because a terse prompt feels rude to the human editors, not because the model needs it. Keep a comment in the repo for the humans. "This used to say please. We deleted it on purpose." The model does not read the comment. The next editor does, and they will not "fix" the tone by restoring a dozen tokens and a vaguer verb. Tone, if you need it, is one line in the output section: "Calm, plain, no emoji." That line is a spec. "Thanks so much" is not.
Measure the before and after on a prompt you actually send. If the count drops by 8 tokens and the answers on ten real inputs stay acceptable, keep the cut and move on. If the count drops by 8 and you spent an hour, you optimized the wrong layer. Go find the paragraph that is 400 tokens and says the same thing as the line under it. Filler words are the easy lesson. Repeated substance is the expensive one. Learn the easy lesson in one pass so you can see the expensive one clearly.
Questions
Should every prompt be as short as possible?
No. A prompt should be as short as the task allows. A constraint that prevents a known failure is worth more tokens than it costs. Filler is the part you can delete without changing the spec. If you cannot describe what changed, it was filler.
Why does the score punish "very" and not "professional"?
"Very" almost never changes a spec. "Professional" sometimes does, and sometimes hides the lack of one. The list is a filter for cushion, not a style guide. Read the prompt after the flagged words are gone and see if a task remains.
Will a model answer worse if I delete "please"?
Not in any way you should budget for. Models are not waiting to be asked nicely. They are continuing a pattern. A clear verb and a clear limit produce a more stable pattern than a polite fog around an unclear verb.