UNPKG

@eimen/token-saver

Version:

Strip polite fluff from AI prompts to cut token use and cost.

90 lines (61 loc) 2.1 kB
# Token Saver Cut polite fluff out of your prompts so you spend less on tokens. > 04/17/2025: Sam Altman admits that saying "Please" and "Thank You" to ChatGPT is wasting millions of dollars in computing power. Disclaimer: Only works with English as of now. ## Install ```bash npm i @eimen/token-saver # or yarn add @eimen/token-saver / pnpm add @eimen/token-saver ``` ## Quick start ### ESM ```js import TokenSaver from "@eimen/token-saver"; const saver = new TokenSaver(); const out = saver.process("Please summarise this text, thank you!"); console.log(out.cleaned); // "Summarise this text." console.log(out.estimatedTokensSaved); // 3 ``` ### CommonJS ```js const TokenSaver = require("@eimen/token-saver"); ``` ## Options | Option | Default | What it does | | ----------------------- | :-----: | -------------------------------------------------- | | `removePoliteness` | `true` | kill _please_, _thank you_ and similar niceties | | `removeFillers` | `true` | kill _um_, _uh_, _actually_, _well_ … | | `removeRedundantIntros` | `true` | kill “I was wondering if you could” style prefixes | | `tokenCharRatio` | 4 | char‑per‑token estimate for savings calculation | Pass a partial options object to the constructor: ```js new TokenSaver({ removeFillers: false }); ``` ## API surface ```ts clean(prompt: string): string process(prompt: string): { original: string; cleaned: string; charsSaved: number; estimatedTokensSaved: number; } ``` Types are shipped in `index.d.ts`. ## CLI one‑liner ```bash npx @eimen/token-saver "Please help me. Thanks!" # prints "Help me." ``` _(CLI is just `node -e` sugar; no global install needed.)_ ## Why bother - ChatGPT Turbo is ~4 characters per token. Two polite words can cost 12 tokens. - Multiply by millions of prompts. Real money. - Faster requests, shorter logs. ## Contributing ```bash git clone https://github.com/eimenhmdt/token-saver cd token-saver npm i npm test ``` ## License MIT