Chinese Prompt Normalizer

Why full-width characters quietly break Chinese prompts

When you copy Chinese text into a prompt — from a chat app, a web page, a PDF or a document — you rarely get the clean characters you think you do. Chinese input methods and rich-text sources are full of full-width variants: full-width punctuation (，。！？：；（）), full-width Latin letters and digits (ＡＢＣ１２３), and the full-width space U+3000, which looks identical to an ordinary space but is a completely different character. To a human eye the prompt reads fine; to a large language model, and to any code that parses the prompt, those characters are distinct from their ASCII counterparts. The result is subtle: delimiters that no longer match, variable names that silently differ, JSON that fails to parse, and formatting that drifts when the same prompt is pasted across DeepSeek, Qwen, Doubao, Kimi, ERNIE or Zhipu. A normalizer fixes this once, deterministically, before the prompt ever reaches a model.

This tool is built as a pure text transformer in the same spirit as a code formatter: text goes in, cleaned text comes out, and the same input always produces the same output. There is no model, no randomness and no network call. Each rule is a small, well-defined string operation — full-width punctuation maps to its half-width twin, full-width letters and digits shift back into the ASCII range, runs of spaces collapse to one, blank lines collapse to a single blank line, line edges are trimmed, and curly quotes, en/em dashes and ellipses are unified to straight quotes, a hyphen and three dots. Because the rules are explicit and ordered, you can reason about exactly what happened to your text, and the before / after / changed counters give you an immediate, honest read on how much was touched.

A deterministic transformer, with every rule under your control

Crucially, every rule is an independent toggle, because the right answer depends on your intent. If you are writing Chinese prose for a Chinese model, you may well want to keep ，。！？ as full-width punctuation and only tidy the whitespace and stray full-width letters — so the punctuation rule comes off and the rest stay on. If you are embedding a JSON block, a function call or a code snippet inside the prompt, half-width punctuation and straight quotes matter a great deal, so you turn those rules on. Nothing is forced: the defaults are a sensible clean-up, but you remain in control of each transformation, and you can watch the output update live as you flip each switch.

Treat normalization as the last, boring step before you ship a prompt — the equivalent of running a linter. Paste your draft, glance at the "characters changed" tile to confirm it actually did something, copy the result, and move on. Because everything runs locally in your browser, you can paste sensitive internal prompts, draft system messages or proprietary instructions without anything ever leaving your device, being sent to a model, or being stored. It pairs naturally with a prompt builder: assemble the structure first, then run it through the normalizer so the version you paste into DeepSeek, Qwen, Doubao or Kimi is clean, consistent and free of the invisible full-width gremlins that quietly degrade Chinese prompts. Two seconds of tidying here saves a surprising amount of head-scratching later, when a prompt that "looks identical" behaves differently on two models.