Principles
- Delete or shorten fixed phrases only. No AI is involved, and nothing is summarised or paraphrased. Asking a model to “compress” a prompt reliably drops the one constraint that mattered, and a cheaper answer to the wrong question is not a saving.
- Never touch code. Fenced code blocks are lifted out before any rule runs and put back unchanged afterwards. Indentation and blank lines inside them are content.
- Never remove words that can carry an instruction.
just,simply,briefly,onlyandexactlylook like filler and are deliberately not on any list. “Just the total” is a different request from “the total”. - Always visible. Every change is listed on the answer’s report card.
Order of operations
The order matters. Removing a bare “please” first would turn “could you please summarise this” into “could you summarise this”, leaving the “could you” behind and saving almost nothing. So long phrases run before the single words inside them.- Protect fenced code blocks.
- Formatting cleanups.
- Wordy phrases shortened.
- Stock politeness removed, in six groups.
- Seams tidied where something was removed mid-sentence.
- Restore code blocks and trim the start and end.
1. Formatting cleanups
2. Wordy phrases shortened
This is the only rule that changes wording rather than only deleting. It runs before the deletions. Matches ignore case; the replacement is written in lower case, so “In order to ship” at the start of a sentence becomes “to ship”. Change label: “shortened N wordy phrase(s)”. There are 50 entries.Where these come from
The list was built from two public, MIT-licensed plain-language databases:retext-simplify (324 entries, derived from US government plain-language guidance) and too-wordy (236 entries, from the write-good family of prose linters). Neither is installed; the lists were read once, filtered, and reviewed by hand.
- Mechanical filter. Only multi-word phrases with exactly one replacement that is strictly shorter were kept, and anything containing a negation, quantifier or modal (“not”, “all”, “only”, “must”, “may”) was dropped, because those change the claim rather than its length. 58 entries survived.
- Human review of the 58. Entries that change register rather than length were removed (“adversely impact” → “hurt”), as were entries that change meaning (“benefit from” → “enjoy”, “is responsible for” → “handles”), one ungrammatical entry, and military and legal boilerplate.
- A few very common phrases missing from both sources were added by hand (“the majority of”, “on a regular basis”, “in the near future”).
retext-simplify also marks 21 entries as safe to delete outright, including “there is”, “it is”, “very”, “currently” and “appropriate”. None of those are applied. Deleting “there is” from “there is a bug in the export” leaves “a bug in the export”, which is not the same statement.
3. Stock politeness removed
Each group is matched case-insensitively. Except for greetings, the phrases are removed wherever they appear in the message, not only at the start. Each group that fires adds one line to the report card.Request preambles, long forms (label: “removed request preamble”)
Removed when followed by more text:- I was wondering if you could / can / might
- I would be grateful if you could / would
- I’d appreciate it if you could / would; I’d really appreciate it if you could / would
- would it be possible to; would it be possible for you to
- is there any way you could / can; is there any chance (that) you could / can
- do you think you could / can
- if it’s not too much trouble
- when you get a chance; when you have a chance; when you get a moment; when you have a moment
- sorry to bother you
- quick question; real quick
Request preambles, short forms (label: “removed request preamble”)
Removed when followed by more text:- please could you; please can you
- could you; could you please; could you kindly
- can you; can you please; can you kindly
- would you; would you please; would you kindly; would you mind
- will you; will you please
- I would like you to; I’d like you to
- I need you to; I want you to
- I’m hoping you can
- may I ask you to
Sign-offs (label: “removed sign-off”)
- thanks so much; thanks a lot; thanks a million; thanks in advance; thanks again; thanks kindly (also “thank” without the s)
- thank you so much; thank you very much; thank you in advance; thank you kindly
- much appreciated; many thanks
- I appreciate your help; I appreciate it; I appreciate this
- appreciate your help; appreciate it
- cheers
Greetings (label: “removed greeting”)
Only at the very start of the message:- hi, hello or hey, optionally followed by “there”, optionally followed by “hope you’re well”, “hope you are well”, “hope you’re doing well” or “hope you are doing well”
Redundant framing (label: “removed redundant framing”)
- as an AI; as an AI language model
- you are an AI; you are an AI language model
- as a language model
- I know you’re an AI, but
Closing hedges (label: “removed closing hedge”)
- if that makes sense; if this makes sense
- hope that makes sense; hope this makes sense
- if you don’t mind; if you do not mind
- if you would be so kind
Bare courtesy words (label: “removed politeness”)
Last, once every phrase containing them has gone:- please; kindly (when followed by more text)
4. Tidying the seams
Removing a phrase mid-sentence leaves its spacing behind. Afterwards:- spaces before
? ! . , ; :are removed, - spaces after an opening
(or[are removed, - runs of two or more spaces become one,
- punctuation and spaces stranded at the very start are removed (so “You are an AI language model. Explain OAuth” does not become ”. Explain OAuth”).
Worked examples
How savings are counted
tokens trimmed = ceil(original characters ÷ 4) − ceil(sent characters ÷ 4), floored at zero.
This is the same four-characters-per-token approximation used elsewhere for estimates. It is shown as an estimate only; billing always uses the provider’s reported token counts.
CorePillar’s own measurement on a six-message sample of deliberately wordy business English removed 45 of about 121 tokens, a 37% reduction. Two of the six messages were left untouched. That is a best case on text chosen to contain these patterns. A terse, well-written prompt saves nothing.