Mdscript — AI skill for Claude Code

MDS - Markdown Script

[](https://github.com/dean0x/mdscript/actions/workflows/ci.yml) [](https://crates.io/crates/mds-cli) [](https://www.npmjs.com/package/@mdscript/mds) [](LICENSE)

LLM prompts grow into copy-pasted walls of text that drift across agents, models, and environments. MDS gives you variables, functions, imports, and conditionals so you can write prompts once and compose them everywhere, compiled to clean Markdown.

Built for AI engineers who manage prompt libraries across agents, models, and environments.

Quick Start

**Install the CLI** (Rust):

cargo install mds-cli

**Or install via npm** (Node or browser):

npm install @mdscript/mds

**Create and compile a starter template**:

mds init                       # creates hello.mds
mds build hello.mds            # compiles to hello.md
mds build hello.mds -o -       # stdout

**For multi-file prompt libraries**, templates compose via `@import` — for example:

---
model: claude-sonnet
tools: [search, calculator]
---

@import "./safety.mds" as guard
@import "./personas.mds" as persona

{{persona.code_reviewer("TypeScript")}}

{{guard.safety_rules()}}

## Available Tools

@for tool in tools:
- **{{tool}}**
@end

@if model == "claude-sonnet":
Use extended thinking for complex tasks.
@end

See [`examples/prompt-library/`](examples/prompt-library/) for a complete reusable prompt library using `@export`/`@import` (personas, formatting, guardrails). If imports from cross-directory libraries are rejected with a containment error, see [Project root](#project-root).

Unlike general-purpose template engines, MDS is Markdown-native: no delimiters to escape, no runtime to configure. The compiler catches undefined varia