Chinese Few-Shot Example Formatter

Examples teach the model better than descriptions

Few-shot prompting is the simplest reliable upgrade you can make to a Chinese-model prompt. Instead of describing in words exactly how you want an answer formatted, you show the model two or three worked examples — pairs of input and the ideal output — and let it generalise the pattern. For "fixed-pattern" work like classification, extraction, rewriting or format conversion, this is consistently more dependable than prose instructions, because the model sees the target shape directly rather than inferring it. This formatter keeps that structure for you: write one task line, add a few input/output pairs, and it assembles a clean block — task heading first, each example numbered and prefixed with "输入：" / "输出：", and a trailing "现在请处理：" line ready for your real data. The result is exactly the kind of few-shot prompt a careful prompt engineer would hand-assemble, only built in seconds and ready to paste into any model.

The order matters. Leading with a one-sentence task line tells the model what the job is and by what criteria, so when it reaches the examples it reads them as demonstrations of that goal rather than as random text. After the task, the examples do the heavy lifting. Each pair fixes a little more of the format and the judgement you want: how to phrase the output, which fields to include, how to handle the awkward middle cases. Because the tool numbers and formats every pair identically, the model sees a stable repeating structure and tends to continue it faithfully — which is precisely why few-shot is so good at locking in a consistent output shape across many later inputs.

A few good pairs beat a wall of instructions

The art of few-shot is choosing the right examples, not piling on more. Two or three high-quality pairs that cover the typical case and the edge cases you actually care about — empty values, ambiguity, malformed input — will outperform ten examples written carelessly. Every example also costs context on every single turn, so a bloated block both raises your token bill and risks diluting the instruction the model should be following. The discipline is the same as writing good unit tests: pick the cases that pin down the behaviour you care about, demonstrate them cleanly, and stop. If the model handles those, it will usually handle the rest.

Because the output is structured plain text, the same few-shot block is portable across every major Chinese model and works just as well on ChatGPT, Claude or Gemini; the trailing hand-off line gives the model an unambiguous place to continue. And because the whole tool runs locally in your browser, you can iterate freely — add a pair, tweak an output, copy again, test — without anything you type ever leaving your device, being sent to a model, or being stored. Treat the first block as a draft: run it, watch where the model drifts, then add or sharpen the example that covers that exact case. Two or three rounds of that usually turn a mediocre, inconsistent response into a clean, repeatable one, and you keep a tidy, reusable few-shot prompt at the end.