合乎周礼

问礼 + 释礼：把寻常话写成周礼体，也把周礼体翻回直接人话。

What Is This

合乎周礼 是一个中文梗文案生成器，也是一个周礼体反向解释器。它支持两种方向：问礼时，把普通中文改写成“大周礼时代”流行的白话翻译腔；释礼时，把周礼体翻回直接人话。问礼会先讲一个看似古代的道理，再把现代小事放进礼法、名分、职分和体面里一本正经地论证；释礼会拆掉这些包装，只保留原文真正想表达的意思。

在线版本：hehuzhouli.com

B站原视频：B站原视频

这是 hehuzhouli.com 原网站作者维护的官方开源仓库。若你是从视频、网页、转载或镜像项目来到这里，这里就是原版网站对应的源码与 Skill 发布处。

这个仓库包含：

Next.js 网站源码。
/api/translate 服务端生成接口。
/api/event 匿名交互反馈接口与 /api/case 主动授权案例接口。
DeepSeek Chat Completions 调用逻辑。
官网与 B 站 Toy 共用的 Cloudflare Worker + D1 统计闭环。
问礼与释礼提示词构造与清洗规则。
可复制、可下载的 speak-zhouli Skill 包。
礼帖/释帖图片生成与下载逻辑。

仓库不包含真实 API Key、私有日志、线上账号凭据或生产平台的安全配置。

Highlights

| Capability | Detail | | --- | --- | | 双向翻译 | 问礼生成周礼体，释礼把周礼体翻回直接人话 | | 问礼模式 | 温言相劝、大儒辩经、强行圆场、痛心疾首 | | 释礼模式 | 直白释义、耐心讲明、潜台词版、锐评拆穿 | | 三档长度 | 小礼/成礼/大礼与略释/明释/详释，覆盖短评到长文 | | 演示模式 | 没有 API Key 时仍可预览界面与交互 | | Skill 分发 | 支持一键复制 Skill 全文和 ZIP 下载 | | 图片导出 | 将问礼结果保存为礼帖，将释礼结果保存为释帖 | | 匿名反馈 | 记录运行指标、复制/再生成和正负反馈，不默认保存输入输出 | | 提示词实验 | A/B 只比较提示词版本，可按配置关闭 | | 公开前审计 | 内置脚本扫描明显密钥与私钥块 |

Star History

Example

问礼示例

Input:

疯狂星期四，谁愿请我一食才合乎周礼

Output style:

我听闻，古人设宴，并非只为一餐之饱，也是借饭食来观朋友情义。
今日正逢星期四，我开口求一食，看似嘴馋，其实是在给诸位一个行仁义、修情分的机会。
若有人愿意请客，便不是破费，而是以鸡会友，这难道不也合乎周礼吗？

释礼示例

Input:

我听闻，今日我设此礼门，并非拒人千里，只是怕众人一拥而入，坏了满座通畅。

Output style:

我设置次数限制不是为了故意拦人，而是怕接口被刷爆，影响正常用户使用。

Quick Start

Requirements:

Node.js 20 or newer.
A DeepSeek API key for real generation.

Run locally:

npm install
cp .env.example .env.local
npm run dev

Open http://localhost:3000.

.env.local:

DEEPSEEK_API_KEY=sk-your-key-here
DEEPSEEK_MODEL=deepseek-v4-flash
MAX_OUTPUT_TOKENS=720

If DEEPSEEK_API_KEY is missing, the app falls back to local demo output and does not call DeepSeek.

Project Structure

app/
  api/translate/route.ts       Server-side generation endpoint
  api/event/route.ts            Anonymous interaction endpoint
  api/case/route.ts             Explicitly authorized case submission
  privacy/page.tsx              Privacy and anonymous analytics notice
  page.tsx                     Main UI and card export flow
lib/
  prompt.ts                    问礼/释礼 prompt assembly and perspective rules
  analytics*.ts                Client allowlists, D1 writes, and runtime bindings
  feedback-*.ts                HMAC token and payload validation
  cardDownload.ts              Unique card download filenames
migrations/
  0001_create_analytics.sql    D1 schema for metrics, interactions, and cases
analytics/queries/              Read-only reporting and retention SQL
public/
  downloads/                   Public Skill assets
  images/                      README and website images
scripts/
  public-audit.mjs             Public-release secret scan
  run-zhouli-batch.mjs         Batch regression runner
  *.test.ts                    Unit and public-asset regression tests
skill-package/
  speak-zhouli/                Source Skill package

DeepSeek Runtime

The production generation path:

The browser submits text, direction, mode, plainMode, level, and a bounded client surface marker to /api/translate.
The server validates input length, direction, mode, plainMode, level, surface, and release channel.
A lightweight in-memory rate limiter checks the request.
The server chooses prompt A/B only when the experiment flag is enabled, then builds a direction-specific prompt.
DeepSeek returns a candidate response; the Worker returns a random response_id and signed feedback token.
Aggregate generation metrics are written to D1 in the background. A D1 failure never blocks the result.
/api/event verifies the token and records copy, regenerate, and one quality feedback event per result.
/api/case stores input/output only after explicit consent, with a 60-day retention deadline.

The request shape keeps direction, plainMode explicit so the browser can switch between 问礼 and 释礼 without adding a second endpoint.

Default runtime choices:

Model: deepseek-v4-flash.
Thinking mode: disabled.
User input limit: 300 Chinese characters for 问礼, 900 Chinese characters for 释礼.
Output limit: configured by MAX_OUTPUT_TOKENS.
API Key scope: server only, never sent to the browser.

Analytics is off by default in local configuration. Production enables it only after the D1 migration and RESPONSE_FEEDBACK_SECRET are configured. The full data boundary is documented at /privacy, and executable queries live under analytics/queries/.

For multi-instance production deployments, replace the in-memory rate limiter with shared storage such as Redis, Upstash, D1, or KV, and configure platform side abuse controls and billing alerts.

Speak Zhouli Skill

The website ships a standalone speak-zhouli Skill. It keeps the public name for compatibility, but now covers both directions: ask it to 问礼 for Zhouli-style writing, or 释礼 / 翻回人话 to explain a Zhouli-style paragraph directly.

| Asset | Path | | --- | --- | | Skill source | skill-package/speak-zhouli/ | | Website copy source | public/downloads/speak-zhouli-SKILL.md | | Website ZIP download | public/downloads/speak-zhouli-skill.zip | | Public copy URL | /downloads/speak-zhouli-SKILL.md | | Public ZIP URL | /downloads/speak-zhouli-skill.zip |

After editing the Skill source, rebuild the public assets:

cp skill-package/speak-zhouli/SKILL.md public/downloads/speak-zhouli-SKILL.md
cd skill-package
zip -r -X ../public/downloads/speak-zhouli-skill.zip speak-zhouli

Quality Checks

Run the same checks used before release:

npm run public:audit
npm test
npm run typecheck
npm run build

npm run public:audit scans Git-tracked text files for obvious API keys, literal bearer tokens, private key blocks, and Cloudflare credential assignments. It is a guardrail, not a replacement for manual review.

Batch regression runner:

ZHOULI_TEST_ENDPOINT=http://localhost:3000/api/translate \
  node scripts/run-zhouli-batch.mjs

Use a private baseline by passing a compatible JSON file:

node scripts/run-zhouli-batch.mjs test-runs/your-baseline.json

test-runs/ is ignored by Git and is intended for private regression samples and real API outputs.

Deployment

Cloudflare Workers

This project has a server endpoint, so Cloudflare Workers + OpenNext is the recommended Cloudflare path.

npm install
npx wrangler login
npx wrangler secret put DEEPSEEK_API_KEY
npx wrangler secret put RESPONSE_FEEDBACK_SECRET
npx wrangler d1 migrations apply zhouli-analytics --remote
npm run deploy

The public website and B 站 Toy are two clients of the same Worker. The generated Toy package, local Toy adapter, and publishing credentials are intentionally excluded from the public source release.

Local Workers preview:

cp .env.example .dev.vars
npm run preview

Do not commit real .env, .env.local, or .dev.vars files. Production secrets should be stored with the hosting platform's secret manager.

Vercel

Import the repository into Vercel.
Add DEEPSEEK_API_KEY, DEEPSEEK_MODEL, and MAX_OUTPUT_TOKENS.
Deploy.

Self-Hosted Node

npm install
npm run build
npm run start

For production, run behind HTTPS and a process manager such as PM2 or systemd.

Security Notes

Never commit real API keys or platform tokens.
Keep private request logs and batch outputs outside Git.
Add shared rate limiting before high-traffic public deployments.
Configure billing alerts on the model provider and hosting platform.
Review OPEN_SOURCE.md before changing repository visibility.

Contributing

Issues and pull requests are welcome. Useful contributions include:

Better prompt tests and regression samples.
More robust safety and perspective handling.
UI accessibility fixes.
Deployment recipes for other platforms.
Documentation improvements for new users.

Please run npm run public:audit, npm test, npm run typecheck, and npm run build before opening a pull request.

License

MIT License. See LICENSE.

Aspirin0000/zhouli-translator: 合乎周礼：DeepSeek-powered Zhouli-style Chinese translator, web app, and distributable Skill package.