# TypeChat — Schema-First Natural Language UIs

> Build typed natural language interfaces by turning intent into validated JSON via TypeScript types. Replace prompt engineering with schema engineering.

## Install

Save as a script file and run:

# TypeChat — Schema-First Natural Language UIs

> Build typed natural language interfaces by turning intent into validated JSON via TypeScript types. Replace prompt engineering with schema engineering.

## Quick Use

1. Install:
   ```bash
   npm install typechat
   ```
2. Run:
   ```bash
   node -e "import('typechat').then(()=>console.log('typechat ok')).catch(e=>{console.error(e); process.exit(1);})"
   ```
3. Verify:
   - Run one example project and confirm invalid model output is repaired into valid typed JSON.


---

## Intro

Build typed natural language interfaces by turning intent into validated JSON via TypeScript types. Replace prompt engineering with schema engineering.

- **Best for:** Product teams who need typed actions (APIs, workflows) from an LLM with validation and repair loops
- **Works with:** TypeScript/JavaScript, any LLM backend, JSON schema-style validation flows
- **Setup time:** 12 minutes


### Quantitative Notes

- Setup time ~12 minutes (npm install + run one example in the repo)
- GitHub stars + forks (verified): see Source & Thanks
- Typical loop: generate JSON → validate → repair (1–2 retries)


---

## Practical Notes

If you’re building an agent that triggers actions, define a narrow union type for allowed intents and keep each intent payload minimal. Add deterministic validation (schema + business rules) and log all repairs—repairs are where many hidden bugs appear. Over time, grow the schema set rather than growing prompts.

**Safety note:** Always validate typed JSON with business rules (not just schema) before executing real actions.

### FAQ

**Q: Do I still need prompt engineering?**
A: Much less. You focus on defining types/schemas; prompts become smaller and more stable.

**Q: How does it handle invalid JSON?**
A: It validates and can re-prompt the model to repair non-conforming output.

**Q: Is it only for TypeScript apps?**
A: TypeChat provides TypeScript/JS support, and the repo also references other language variants.

---

## Source & Thanks

> GitHub: https://github.com/microsoft/TypeChat
> Owner avatar: https://avatars.githubusercontent.com/u/6154722?v=4
> License (SPDX): MIT
> GitHub stars (verified via `api.github.com/repos/microsoft/TypeChat`): 8,652
> GitHub forks (verified via `api.github.com/repos/microsoft/TypeChat`): 413


---

<!-- ZH -->

# TypeChat——用类型定义自然语言界面

> 通过 TypeScript 类型把用户意图约束为可校验的 JSON 输出，构建更安全、可维护的自然语言界面；用 schema engineering 替代脆弱的 prompt engineering。

## 快速使用

1. 安装：
   ```bash
   npm install typechat
   ```
2. 运行：
   ```bash
   node -e "import('typechat').then(()=>console.log('typechat ok')).catch(e=>{console.error(e); process.exit(1);})"
   ```
3. 验证：
   - Run one example project and confirm invalid model output is repaired into valid typed JSON.


---

## 简介

通过 TypeScript 类型把用户意图约束为可校验的 JSON 输出，构建更安全、可维护的自然语言界面；用 schema engineering 替代脆弱的 prompt engineering。

- **适合谁（Best for）:** 需要把 LLM 输出变成“可执行的类型化动作”（API/工作流）并做校验与修复的产品团队
- **兼容工具（Works with）:** TypeScript/JavaScript、任意 LLM 后端、JSON schema 风格的校验与修复流程
- **安装时间（Setup time）:** 12 分钟


### 量化信息

- 跑通约 12 分钟（npm 安装 + 跑一个示例）
- GitHub stars + forks（已核验）：见「来源与感谢」
- 典型流程：生成 JSON → 校验 → 修复（1–2 次重试）


---

## 实战要点

做“会触发动作”的 agent 时，建议先把允许的意图收敛成一个窄的 union type，并让每个 intent 的 payload 尽量小。加入确定性的校验（schema + 业务规则）并记录所有修复过程——修复环节往往暴露隐藏 bug。长期演进优先扩展 schema 集，而不是无止境堆 prompt。

**安全提示：** 在执行真实动作前，务必用业务规则做二次校验（不仅仅是 schema 校验）。

### FAQ

**Q: 还需要做 prompt 工程吗？**
A: 需求会小很多：重点转向类型/Schema 设计，prompt 更短也更稳。

**Q: 模型输出不是合法 JSON 怎么办？**
A: 先校验，再通过进一步交互让模型修复为符合类型的输出。

**Q: 只能用于 TypeScript 吗？**
A: 核心支持 TypeScript/JS；仓库也提供其它语言的实现/说明链接。

---

## 来源与感谢

> GitHub：https://github.com/microsoft/TypeChat
> Owner avatar：https://avatars.githubusercontent.com/u/6154722?v=4
> 许可证（SPDX）：MIT
> GitHub stars（已通过 `api.github.com/repos/microsoft/TypeChat` 核验）：8,652
> GitHub forks（已通过 `api.github.com/repos/microsoft/TypeChat` 核验）：413


---
Source: https://tokrepo.com/en/workflows/typechat-schema-first-natural-language-uis
Author: Agent Toolkit